diff --git "a/parse/train/SJ1Xmf-Rb/SJ1Xmf-Rb_middle.json" "b/parse/train/SJ1Xmf-Rb/SJ1Xmf-Rb_middle.json" new file mode 100644--- /dev/null +++ "b/parse/train/SJ1Xmf-Rb/SJ1Xmf-Rb_middle.json" @@ -0,0 +1,41263 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 410, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 411, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 411, + 96 + ], + "score": 1.0, + "content": "FEARNET: BRAIN-INSPIRED MODEL FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 300, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 300, + 117 + ], + "score": 1.0, + "content": "INCREMENTAL LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 290, + 146 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 291, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 291, + 149 + ], + "score": 1.0, + "content": "Ronald Kemker and Christopher Kanan∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 113, + 147, + 257, + 190 + ], + "lines": [ + { + "bbox": [ + 111, + 145, + 258, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 258, + 158 + ], + "score": 1.0, + "content": "Carlson Center for Imaging Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 155, + 250, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 250, + 170 + ], + "score": 1.0, + "content": "Rochester Institute of Technology", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 167, + 227, + 179 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 227, + 179 + ], + "score": 1.0, + "content": "Rochester, NY 14623, USA", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 179, + 250, + 191 + ], + "spans": [ + { + "bbox": [ + 113, + 179, + 250, + 191 + ], + "score": 1.0, + "content": "{rmk6217,kanan}@rit.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 245, + 468, + 409 + ], + "lines": [ + { + "bbox": [ + 142, + 246, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 246, + 469, + 257 + ], + "score": 1.0, + "content": "Incremental class learning involves sequentially learning classes in bursts of ex-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "score": 1.0, + "content": "amples from the same class. This violates the assumptions that underlie methods", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 278 + ], + "score": 1.0, + "content": "for training standard deep neural networks, and will cause them to suffer from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 291 + ], + "score": 1.0, + "content": "catastrophic forgetting. Arguably, the best method for incremental class learning", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "is iCaRL, but it requires storing training examples for each class, making it chal-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "score": 1.0, + "content": "lenging to scale. Here, we propose FearNet for incremental class learning. Fear-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "Net is a generative model that does not store previous examples, making it mem-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "ory efficient. FearNet uses a brain-inspired dual-memory system in which new", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "memories are consolidated from a network for recent memories inspired by the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 344, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 357 + ], + "score": 1.0, + "content": "mammalian hippocampal complex to a network for long-term storage inspired by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "score": 1.0, + "content": "medial prefrontal cortex. Memory consolidation is inspired by mechanisms that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 364, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 470, + 379 + ], + "score": 1.0, + "content": "occur during sleep. FearNet also uses a module inspired by the basolateral amyg-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "score": 1.0, + "content": "dala for determining which memory system to use for recall. FearNet achieves", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 388, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 469, + 399 + ], + "score": 1.0, + "content": "state-of-the-art performance at incremental class learning on image (CIFAR-100,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 399, + 383, + 410 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 383, + 410 + ], + "score": 1.0, + "content": "CUB-200) and audio classification (AudioSet) benchmarks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 433, + 206, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 208, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 208, + 448 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 459, + 504, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "In incremental classification, an agent must sequentially learn to classify training examples, without", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "necessarily having the ability to re-study previously seen examples. While deep neural networks", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "(DNNs) have revolutionized machine perception (Krizhevsky et al., 2012), off-the-shelf DNNs can-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "not incrementally learn classes due to catastrophic forgetting. Catastrophic forgetting is a phe-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 501, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 516 + ], + "score": 1.0, + "content": "nomenon in which a DNN completely fails to learn new data without forgetting much of its pre-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 513, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 506, + 526 + ], + "score": 1.0, + "content": "viously learned knowledge (McCloskey & Cohen, 1989). While methods have been developed to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 537 + ], + "score": 1.0, + "content": "try and mitigate catastrophic forgetting, as shown in Kemker et al. (2018), these methods are not", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 535, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 506, + 549 + ], + "score": 1.0, + "content": "sufficient and perform poorly on larger datasets. In this paper, we propose FearNet, a brain-inspired", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 546, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 481, + 559 + ], + "score": 1.0, + "content": "system for incrementally learning categories that significantly outperforms previous methods.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "The standard way for dealing with catastrophic forgetting in DNNs is to avoid it altogether by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "mixing new training examples with old ones and completely re-training the model offline. For large", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "datasets, this may require weeks of time, and it is not a scalable solution. An ideal incremental", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "learning system would be able to assimilate new information without the need to store the entire", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "training dataset. A major application for incremental learning includes real-time operation on-board", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "embedded platforms that have limited computing power, storage, and memory, e.g., smart toys,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "score": 1.0, + "content": "smartphone applications, and robots. For example, a toy robot may need to learn to recognize objects", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "within its local environment and of interest to its owner. Using cloud computing to overcome these", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "resource limitations may pose privacy risks and may not be scalable to a large number of embedded", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "devices. A better solution is on-device incremental learning, which requires the model to use less", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 673, + 244, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 244, + 686 + ], + "score": 1.0, + "content": "storage and computational power.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 690, + 503, + 712 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "In this paper, we propose an incremental learning framework called FearNet (see Fig. 1). FearNet", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 714 + ], + "score": 1.0, + "content": "has three brain-inspired sub-systems: 1) a recent memory system for quick recall, 2) a memory", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 204, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 205, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 205, + 734 + ], + "score": 1.0, + "content": "∗Corresponding author.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 410, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 411, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 411, + 96 + ], + "score": 1.0, + "content": "FEARNET: BRAIN-INSPIRED MODEL FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 300, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 300, + 117 + ], + "score": 1.0, + "content": "INCREMENTAL LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 290, + 146 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 291, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 291, + 149 + ], + "score": 1.0, + "content": "Ronald Kemker and Christopher Kanan∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 111, + 133, + 291, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 147, + 257, + 190 + ], + "lines": [ + { + "bbox": [ + 111, + 145, + 258, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 258, + 158 + ], + "score": 1.0, + "content": "Carlson Center for Imaging Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 155, + 250, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 250, + 170 + ], + "score": 1.0, + "content": "Rochester Institute of Technology", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 167, + 227, + 179 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 227, + 179 + ], + "score": 1.0, + "content": "Rochester, NY 14623, USA", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 179, + 250, + 191 + ], + "spans": [ + { + "bbox": [ + 113, + 179, + 250, + 191 + ], + "score": 1.0, + "content": "{rmk6217,kanan}@rit.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 111, + 145, + 258, + 191 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 245, + 468, + 409 + ], + "lines": [ + { + "bbox": [ + 142, + 246, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 246, + 469, + 257 + ], + "score": 1.0, + "content": "Incremental class learning involves sequentially learning classes in bursts of ex-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "score": 1.0, + "content": "amples from the same class. This violates the assumptions that underlie methods", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 278 + ], + "score": 1.0, + "content": "for training standard deep neural networks, and will cause them to suffer from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 291 + ], + "score": 1.0, + "content": "catastrophic forgetting. Arguably, the best method for incremental class learning", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "is iCaRL, but it requires storing training examples for each class, making it chal-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "score": 1.0, + "content": "lenging to scale. Here, we propose FearNet for incremental class learning. Fear-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "Net is a generative model that does not store previous examples, making it mem-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "ory efficient. FearNet uses a brain-inspired dual-memory system in which new", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "memories are consolidated from a network for recent memories inspired by the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 344, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 357 + ], + "score": 1.0, + "content": "mammalian hippocampal complex to a network for long-term storage inspired by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "score": 1.0, + "content": "medial prefrontal cortex. Memory consolidation is inspired by mechanisms that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 364, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 470, + 379 + ], + "score": 1.0, + "content": "occur during sleep. FearNet also uses a module inspired by the basolateral amyg-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "score": 1.0, + "content": "dala for determining which memory system to use for recall. FearNet achieves", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 388, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 469, + 399 + ], + "score": 1.0, + "content": "state-of-the-art performance at incremental class learning on image (CIFAR-100,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 399, + 383, + 410 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 383, + 410 + ], + "score": 1.0, + "content": "CUB-200) and audio classification (AudioSet) benchmarks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15, + "bbox_fs": [ + 141, + 246, + 470, + 410 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 433, + 206, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 208, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 208, + 448 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 459, + 504, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "In incremental classification, an agent must sequentially learn to classify training examples, without", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "necessarily having the ability to re-study previously seen examples. While deep neural networks", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "(DNNs) have revolutionized machine perception (Krizhevsky et al., 2012), off-the-shelf DNNs can-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "not incrementally learn classes due to catastrophic forgetting. Catastrophic forgetting is a phe-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 501, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 516 + ], + "score": 1.0, + "content": "nomenon in which a DNN completely fails to learn new data without forgetting much of its pre-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 513, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 506, + 526 + ], + "score": 1.0, + "content": "viously learned knowledge (McCloskey & Cohen, 1989). While methods have been developed to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 537 + ], + "score": 1.0, + "content": "try and mitigate catastrophic forgetting, as shown in Kemker et al. (2018), these methods are not", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 535, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 506, + 549 + ], + "score": 1.0, + "content": "sufficient and perform poorly on larger datasets. In this paper, we propose FearNet, a brain-inspired", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 546, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 481, + 559 + ], + "score": 1.0, + "content": "system for incrementally learning categories that significantly outperforms previous methods.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 459, + 506, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "The standard way for dealing with catastrophic forgetting in DNNs is to avoid it altogether by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "mixing new training examples with old ones and completely re-training the model offline. For large", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "datasets, this may require weeks of time, and it is not a scalable solution. An ideal incremental", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "learning system would be able to assimilate new information without the need to store the entire", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "training dataset. A major application for incremental learning includes real-time operation on-board", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "embedded platforms that have limited computing power, storage, and memory, e.g., smart toys,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "score": 1.0, + "content": "smartphone applications, and robots. For example, a toy robot may need to learn to recognize objects", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "within its local environment and of interest to its owner. Using cloud computing to overcome these", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "resource limitations may pose privacy risks and may not be scalable to a large number of embedded", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "devices. A better solution is on-device incremental learning, which requires the model to use less", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 673, + 244, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 244, + 686 + ], + "score": 1.0, + "content": "storage and computational power.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 563, + 506, + 686 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 690, + 503, + 712 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "In this paper, we propose an incremental learning framework called FearNet (see Fig. 1). FearNet", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 714 + ], + "score": 1.0, + "content": "has three brain-inspired sub-systems: 1) a recent memory system for quick recall, 2) a memory", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "system for long-term storage, and 3) a sub-system that determines which memory system to use for", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "a particular example. FearNet mitigates catastrophic forgetting by consolidating recent memories", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "into long-term storage using pseudorehearsal (Robins, 1995). Pseudorehearsal allows the network", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 130 + ], + "score": 1.0, + "content": "to revisit previous memories during incremental training without the need to store previous training", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 281, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 281, + 140 + ], + "score": 1.0, + "content": "examples, which is more memory efficient.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 689, + 505, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "system for long-term storage, and 3) a sub-system that determines which memory system to use for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "a particular example. FearNet mitigates catastrophic forgetting by consolidating recent memories", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "into long-term storage using pseudorehearsal (Robins, 1995). Pseudorehearsal allows the network", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 130 + ], + "score": 1.0, + "content": "to revisit previous memories during incremental training without the need to store previous training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 281, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 281, + 140 + ], + "score": 1.0, + "content": "examples, which is more memory efficient.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 324, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 325, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 325, + 155 + ], + "score": 1.0, + "content": "Problem Formulation: Here, incremental class learn-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 325, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 166, + 165 + ], + "score": 1.0, + "content": "ing consists of", + "type": "text" + }, + { + "bbox": [ + 167, + 155, + 175, + 164 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 154, + 271, + 165 + ], + "score": 1.0, + "content": "study-sessions. At time", + "type": "text" + }, + { + "bbox": [ + 272, + 155, + 276, + 164 + ], + "score": 0.67, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 154, + 325, + 165 + ], + "score": 1.0, + "content": ", the learner", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 325, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 201, + 177 + ], + "score": 1.0, + "content": "receives a batch of data", + "type": "text" + }, + { + "bbox": [ + 201, + 166, + 213, + 176 + ], + "score": 0.87, + "content": "B _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 165, + 280, + 177 + ], + "score": 1.0, + "content": ", which contains", + "type": "text" + }, + { + "bbox": [ + 280, + 166, + 292, + 176 + ], + "score": 0.86, + "content": "N _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 165, + 325, + 177 + ], + "score": 1.0, + "content": "labeled", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 170, + 330, + 196 + ], + "spans": [ + { + "bbox": [ + 102, + 170, + 202, + 196 + ], + "score": 1.0, + "content": "training samples, i.e.,", + "type": "text" + }, + { + "bbox": [ + 203, + 176, + 291, + 192 + ], + "score": 0.92, + "content": "B _ { t } ~ = ~ \\{ ( \\mathbf { x } _ { j } , y _ { j } ) \\} _ { j = 1 } ^ { N _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 170, + 330, + 196 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 190, + 326, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 141, + 204 + ], + "score": 0.92, + "content": "\\mathbf { x } _ { j } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 190, + 326, + 204 + ], + "score": 1.0, + "content": "is the input feature vector to be classified and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 326, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 117, + 215 + ], + "score": 0.86, + "content": "y _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 202, + 326, + 216 + ], + "score": 1.0, + "content": "is its corresponding label. The number of training", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 325, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 141, + 225 + ], + "score": 1.0, + "content": "samples", + "type": "text" + }, + { + "bbox": [ + 142, + 214, + 155, + 225 + ], + "score": 0.87, + "content": "N _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 214, + 325, + 225 + ], + "score": 1.0, + "content": "may vary between sessions, and the data", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 224, + 325, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 325, + 236 + ], + "score": 1.0, + "content": "inside a study-session is not assumed to be indepen-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 325, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 325, + 248 + ], + "score": 1.0, + "content": "dent and identically distributed (iid). During a study", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 247, + 325, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 325, + 258 + ], + "score": 1.0, + "content": "session, the learner only has access to its current batch,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 257, + 325, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 325, + 269 + ], + "score": 1.0, + "content": "but it may use its own memory to store information", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 268, + 325, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 325, + 279 + ], + "score": 1.0, + "content": "from prior study sessions. We refer to the first session", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 279, + 325, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 325, + 291 + ], + "score": 1.0, + "content": "as the model’s “base-knowledge,” which contains ex-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 290, + 325, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 163, + 302 + ], + "score": 1.0, + "content": "emplars from", + "type": "text" + }, + { + "bbox": [ + 163, + 290, + 195, + 301 + ], + "score": 0.9, + "content": "M \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 290, + 325, + 302 + ], + "score": 1.0, + "content": "classes. The batches learned in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 301, + 325, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 325, + 313 + ], + "score": 1.0, + "content": "all subsequent sessions contain only one class, i.e., all", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 312, + 275, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 117, + 325 + ], + "score": 0.83, + "content": "y _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 312, + 275, + 324 + ], + "score": 1.0, + "content": "will be identical within those sessions.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 109, + 329, + 305, + 340 + ], + "lines": [ + { + "bbox": [ + 107, + 329, + 307, + 341 + ], + "spans": [ + { + "bbox": [ + 107, + 329, + 307, + 341 + ], + "score": 1.0, + "content": "Novel Contributions: Our contributions include:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 334, + 159, + 502, + 334 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 159, + 502, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 334, + 159, + 502, + 334 + ], + "spans": [ + { + "bbox": [ + 334, + 159, + 502, + 334 + ], + "score": 0.967, + "type": "image", + "image_path": "9ec0369767a96c98cf47fe8ba08e1c364ce5daf78138cb6ec61cc3ee815bb4b4.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 334, + 159, + 502, + 171.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 334, + 171.5, + 502, + 184.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 334, + 184.0, + 502, + 196.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 334, + 196.5, + 502, + 209.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 334, + 209.0, + 502, + 221.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 334, + 221.5, + 502, + 234.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 334, + 234.0, + 502, + 246.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 334, + 246.5, + 502, + 259.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 334, + 259.0, + 502, + 271.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 334, + 271.5, + 502, + 284.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 334, + 284.0, + 502, + 296.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 334, + 296.5, + 502, + 309.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 334, + 309.0, + 502, + 321.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 334, + 321.5, + 502, + 334.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 332, + 345, + 504, + 400 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 332, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 332, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 1: FearNet consists of three brain-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 331, + 354, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 331, + 354, + 505, + 369 + ], + "score": 1.0, + "content": "inspired modules based on 1) mPFC (long-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 331, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 331, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "term storage), 2) HC (recent storage), and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 331, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 331, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "3) BLA for determining whether to use", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 332, + 388, + 428, + 400 + ], + "spans": [ + { + "bbox": [ + 332, + 388, + 428, + 400 + ], + "score": 1.0, + "content": "mPFC or HC for recall.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + } + ], + "index": 35.75 + }, + { + "type": "text", + "bbox": [ + 131, + 351, + 323, + 406 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 325, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 325, + 363 + ], + "score": 1.0, + "content": "1. FearNet’s architecture includes three neu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 140, + 360, + 325, + 375 + ], + "spans": [ + { + "bbox": [ + 140, + 360, + 325, + 375 + ], + "score": 1.0, + "content": "ral networks: one inspired by the hip-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 373, + 325, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 325, + 384 + ], + "score": 1.0, + "content": "pocampal complex (HC) for recent memo-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 384, + 325, + 396 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 325, + 396 + ], + "score": 1.0, + "content": "ries, one inspired by the medial prefrontal", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 395, + 325, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 395, + 325, + 407 + ], + "score": 1.0, + "content": "cortex (mPFC) for long-term storage, and one", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 136, + 406, + 504, + 428 + ], + "lines": [ + { + "bbox": [ + 141, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "inspired by the basolateral amygdala (BLA) that determines whether to use HC or mPFC", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 141, + 416, + 184, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 416, + 184, + 429 + ], + "score": 1.0, + "content": "for recall.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 130, + 439, + 504, + 483 + ], + "lines": [ + { + "bbox": [ + 129, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 129, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "2. Motivated by memory replay during sleep, FearNet employs a generative autoencoder for", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 141, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 141, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "pseudorehearsal, which mitigates catastrophic forgetting by generating previously learned", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 142, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 142, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "examples that are replayed alongside novel information during consolidation. This process", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 142, + 472, + 332, + 484 + ], + "spans": [ + { + "bbox": [ + 142, + 472, + 332, + 484 + ], + "score": 1.0, + "content": "does not involve storing previous training data.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + }, + { + "type": "text", + "bbox": [ + 131, + 487, + 504, + 509 + ], + "lines": [ + { + "bbox": [ + 129, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 129, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "3. FearNet achieves state-of-the-art results on large image and audio datasets with a relatively", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 141, + 497, + 466, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 497, + 466, + 510 + ], + "score": 1.0, + "content": "small memory footprint, demonstrating how dual-memory models can be scaled.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 52.5 + }, + { + "type": "title", + "bbox": [ + 108, + 524, + 211, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 213, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 213, + 540 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Catastrophic forgetting in DNNs occurs due to the plasticity-stability dilemma (Abraham & Robins,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "2005). If the network is too plastic, older memories will quickly be overwritten; however, if the", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "score": 1.0, + "content": "network is too stable, it is unable to learn new data. This problem was recognized almost 30 years", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "score": 1.0, + "content": "ago (McCloskey & Cohen, 1989). In French (1999), methods developed in the 1980s and 1990s", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "are extensively discussed, and French argued that mitigating catastrophic forgetting would require", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "having two separate memory centers: one for the long-term storage of older memories and another", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 616, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 630 + ], + "score": 1.0, + "content": "to quickly process new information as it comes in. He also theorized that this type of dual-memory", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 626, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 505, + 642 + ], + "score": 1.0, + "content": "system would be capable of consolidating memories from the fast learning memory center to long-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 638, + 160, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 160, + 651 + ], + "score": 1.0, + "content": "term storage.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 59 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Catastrophic forgetting often occurs when a system is trained on non-iid data. One strategy for re-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "ducing this phenomenon is to mix old examples with new examples, which simulates iid conditions.", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "For example, if the system learns ten classes in a study session and then needs to learn 10 new", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "classes in a later study session, one solution could be to mix examples from the first study session", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "into the later study session. This method is known as rehearsal, and it is one of the earliest methods", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "for reducing catastrophic forgetting (Hetherington & Seidenberg, 1989). Rehearsal essentially uses", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "an external memory to strengthen the model’s representations for examples learned previously, so", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 67 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 104, + 83, + 506, + 140 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 324, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 325, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 325, + 155 + ], + "score": 1.0, + "content": "Problem Formulation: Here, incremental class learn-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 325, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 166, + 165 + ], + "score": 1.0, + "content": "ing consists of", + "type": "text" + }, + { + "bbox": [ + 167, + 155, + 175, + 164 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 154, + 271, + 165 + ], + "score": 1.0, + "content": "study-sessions. At time", + "type": "text" + }, + { + "bbox": [ + 272, + 155, + 276, + 164 + ], + "score": 0.67, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 154, + 325, + 165 + ], + "score": 1.0, + "content": ", the learner", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 325, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 201, + 177 + ], + "score": 1.0, + "content": "receives a batch of data", + "type": "text" + }, + { + "bbox": [ + 201, + 166, + 213, + 176 + ], + "score": 0.87, + "content": "B _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 165, + 280, + 177 + ], + "score": 1.0, + "content": ", which contains", + "type": "text" + }, + { + "bbox": [ + 280, + 166, + 292, + 176 + ], + "score": 0.86, + "content": "N _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 165, + 325, + 177 + ], + "score": 1.0, + "content": "labeled", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 170, + 330, + 196 + ], + "spans": [ + { + "bbox": [ + 102, + 170, + 202, + 196 + ], + "score": 1.0, + "content": "training samples, i.e.,", + "type": "text" + }, + { + "bbox": [ + 203, + 176, + 291, + 192 + ], + "score": 0.92, + "content": "B _ { t } ~ = ~ \\{ ( \\mathbf { x } _ { j } , y _ { j } ) \\} _ { j = 1 } ^ { N _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 170, + 330, + 196 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 190, + 326, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 141, + 204 + ], + "score": 0.92, + "content": "\\mathbf { x } _ { j } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 190, + 326, + 204 + ], + "score": 1.0, + "content": "is the input feature vector to be classified and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 326, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 117, + 215 + ], + "score": 0.86, + "content": "y _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 202, + 326, + 216 + ], + "score": 1.0, + "content": "is its corresponding label. The number of training", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 325, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 141, + 225 + ], + "score": 1.0, + "content": "samples", + "type": "text" + }, + { + "bbox": [ + 142, + 214, + 155, + 225 + ], + "score": 0.87, + "content": "N _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 214, + 325, + 225 + ], + "score": 1.0, + "content": "may vary between sessions, and the data", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 224, + 325, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 325, + 236 + ], + "score": 1.0, + "content": "inside a study-session is not assumed to be indepen-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 325, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 325, + 248 + ], + "score": 1.0, + "content": "dent and identically distributed (iid). During a study", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 247, + 325, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 325, + 258 + ], + "score": 1.0, + "content": "session, the learner only has access to its current batch,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 257, + 325, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 325, + 269 + ], + "score": 1.0, + "content": "but it may use its own memory to store information", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 268, + 325, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 325, + 279 + ], + "score": 1.0, + "content": "from prior study sessions. We refer to the first session", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 279, + 325, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 325, + 291 + ], + "score": 1.0, + "content": "as the model’s “base-knowledge,” which contains ex-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 290, + 325, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 163, + 302 + ], + "score": 1.0, + "content": "emplars from", + "type": "text" + }, + { + "bbox": [ + 163, + 290, + 195, + 301 + ], + "score": 0.9, + "content": "M \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 290, + 325, + 302 + ], + "score": 1.0, + "content": "classes. The batches learned in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 301, + 325, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 325, + 313 + ], + "score": 1.0, + "content": "all subsequent sessions contain only one class, i.e., all", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 312, + 275, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 117, + 325 + ], + "score": 0.83, + "content": "y _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 312, + 275, + 324 + ], + "score": 1.0, + "content": "will be identical within those sessions.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12.5, + "bbox_fs": [ + 102, + 143, + 330, + 325 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 329, + 305, + 340 + ], + "lines": [ + { + "bbox": [ + 107, + 329, + 307, + 341 + ], + "spans": [ + { + "bbox": [ + 107, + 329, + 307, + 341 + ], + "score": 1.0, + "content": "Novel Contributions: Our contributions include:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 334, + 159, + 502, + 334 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 159, + 502, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 334, + 159, + 502, + 334 + ], + "spans": [ + { + "bbox": [ + 334, + 159, + 502, + 334 + ], + "score": 0.967, + "type": "image", + "image_path": "9ec0369767a96c98cf47fe8ba08e1c364ce5daf78138cb6ec61cc3ee815bb4b4.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 334, + 159, + 502, + 171.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 334, + 171.5, + 502, + 184.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 334, + 184.0, + 502, + 196.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 334, + 196.5, + 502, + 209.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 334, + 209.0, + 502, + 221.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 334, + 221.5, + 502, + 234.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 334, + 234.0, + 502, + 246.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 334, + 246.5, + 502, + 259.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 334, + 259.0, + 502, + 271.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 334, + 271.5, + 502, + 284.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 334, + 284.0, + 502, + 296.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 334, + 296.5, + 502, + 309.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 334, + 309.0, + 502, + 321.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 334, + 321.5, + 502, + 334.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 332, + 345, + 504, + 400 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 332, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 332, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 1: FearNet consists of three brain-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 331, + 354, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 331, + 354, + 505, + 369 + ], + "score": 1.0, + "content": "inspired modules based on 1) mPFC (long-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 331, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 331, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "term storage), 2) HC (recent storage), and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 331, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 331, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "3) BLA for determining whether to use", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 332, + 388, + 428, + 400 + ], + "spans": [ + { + "bbox": [ + 332, + 388, + 428, + 400 + ], + "score": 1.0, + "content": "mPFC or HC for recall.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + } + ], + "index": 35.75 + }, + { + "type": "text", + "bbox": [ + 131, + 351, + 323, + 406 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 325, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 325, + 363 + ], + "score": 1.0, + "content": "1. FearNet’s architecture includes three neu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 140, + 360, + 325, + 375 + ], + "spans": [ + { + "bbox": [ + 140, + 360, + 325, + 375 + ], + "score": 1.0, + "content": "ral networks: one inspired by the hip-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 373, + 325, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 325, + 384 + ], + "score": 1.0, + "content": "pocampal complex (HC) for recent memo-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 384, + 325, + 396 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 325, + 396 + ], + "score": 1.0, + "content": "ries, one inspired by the medial prefrontal", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 395, + 325, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 395, + 325, + 407 + ], + "score": 1.0, + "content": "cortex (mPFC) for long-term storage, and one", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 141, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "inspired by the basolateral amygdala (BLA) that determines whether to use HC or mPFC", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 141, + 416, + 184, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 416, + 184, + 429 + ], + "score": 1.0, + "content": "for recall.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 38, + "bbox_fs": [ + 130, + 350, + 325, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 136, + 406, + 504, + 428 + ], + "lines": [], + "index": 46.5, + "bbox_fs": [ + 141, + 405, + 505, + 429 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 130, + 439, + 504, + 483 + ], + "lines": [ + { + "bbox": [ + 129, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 129, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "2. Motivated by memory replay during sleep, FearNet employs a generative autoencoder for", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 141, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 141, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "pseudorehearsal, which mitigates catastrophic forgetting by generating previously learned", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 142, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 142, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "examples that are replayed alongside novel information during consolidation. This process", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 142, + 472, + 332, + 484 + ], + "spans": [ + { + "bbox": [ + 142, + 472, + 332, + 484 + ], + "score": 1.0, + "content": "does not involve storing previous training data.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5, + "bbox_fs": [ + 129, + 439, + 505, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 487, + 504, + 509 + ], + "lines": [ + { + "bbox": [ + 129, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 129, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "3. FearNet achieves state-of-the-art results on large image and audio datasets with a relatively", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 141, + 497, + 466, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 497, + 466, + 510 + ], + "score": 1.0, + "content": "small memory footprint, demonstrating how dual-memory models can be scaled.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 52.5, + "bbox_fs": [ + 129, + 487, + 505, + 510 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 524, + 211, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 213, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 213, + 540 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Catastrophic forgetting in DNNs occurs due to the plasticity-stability dilemma (Abraham & Robins,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "2005). If the network is too plastic, older memories will quickly be overwritten; however, if the", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "score": 1.0, + "content": "network is too stable, it is unable to learn new data. This problem was recognized almost 30 years", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "score": 1.0, + "content": "ago (McCloskey & Cohen, 1989). In French (1999), methods developed in the 1980s and 1990s", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "are extensively discussed, and French argued that mitigating catastrophic forgetting would require", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "having two separate memory centers: one for the long-term storage of older memories and another", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 616, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 630 + ], + "score": 1.0, + "content": "to quickly process new information as it comes in. He also theorized that this type of dual-memory", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 626, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 505, + 642 + ], + "score": 1.0, + "content": "system would be capable of consolidating memories from the fast learning memory center to long-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 638, + 160, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 160, + 651 + ], + "score": 1.0, + "content": "term storage.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 59, + "bbox_fs": [ + 104, + 550, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Catastrophic forgetting often occurs when a system is trained on non-iid data. One strategy for re-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "ducing this phenomenon is to mix old examples with new examples, which simulates iid conditions.", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "For example, if the system learns ten classes in a study session and then needs to learn 10 new", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "classes in a later study session, one solution could be to mix examples from the first study session", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "into the later study session. This method is known as rehearsal, and it is one of the earliest methods", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "for reducing catastrophic forgetting (Hetherington & Seidenberg, 1989). Rehearsal essentially uses", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "an external memory to strengthen the model’s representations for examples learned previously, so", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "that they are not overwritten when learning data from new classes. Rehearsal reduces forgetting,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "but performance is still worse than offline models. Moreover, rehearsal requires storing all of the", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "training data. Robins (1995) argued that storing of training examples was inefficient and of “little", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "interest,” so he introduced pseudorehearsal. Rather than replaying past training data, in pseudore-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "hearsal, the algorithm generates new examples for a given class. In Robins (1995), this was done", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "by creating random input vectors, having the network assign them a label, and then mixing them", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "into the new training data. This idea was revived in Draelos et al. (2017), where a generative au-", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "toencoder was used to create pseudo-examples for unsupervised incremental learning. This method", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "inspired FearNet’s approach to memory consolidation. Pseudorehearsal is related to memory replay", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "that occurs in mammalian brains, which involves reactivation of recently encoded memories in HC", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 448, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 448, + 205 + ], + "score": 1.0, + "content": "so that they can be integrated into long-term storage in mPFC (Rasch & Born, 2013).", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 67, + "bbox_fs": [ + 105, + 655, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "that they are not overwritten when learning data from new classes. Rehearsal reduces forgetting,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "but performance is still worse than offline models. Moreover, rehearsal requires storing all of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "training data. Robins (1995) argued that storing of training examples was inefficient and of “little", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "interest,” so he introduced pseudorehearsal. Rather than replaying past training data, in pseudore-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "hearsal, the algorithm generates new examples for a given class. In Robins (1995), this was done", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "by creating random input vectors, having the network assign them a label, and then mixing them", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "into the new training data. This idea was revived in Draelos et al. (2017), where a generative au-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "toencoder was used to create pseudo-examples for unsupervised incremental learning. This method", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "inspired FearNet’s approach to memory consolidation. Pseudorehearsal is related to memory replay", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "that occurs in mammalian brains, which involves reactivation of recently encoded memories in HC", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 448, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 448, + 205 + ], + "score": 1.0, + "content": "so that they can be integrated into long-term storage in mPFC (Rasch & Born, 2013).", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "Recently there has been renewed interest in solving catastrophic forgetting in supervised learning.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "Many new methods are designed to mitigate catastrophic forgetting when each study session con-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "tains a permuted version of the entire training dataset (see Goodfellow et al. (2013)). Unlike incre-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "mental class learning, all labels are contained in each study session. PathNet uses an evolutionary", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "algorithm to find the optimal path through a large DNN, and then freezes the weights along that path", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "(Fernando et al., 2017). It assumes all classes are seen in each study session, and it is not capable of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "incremental class learning. Elastic Weight Consolidation (EWC) employs a regularization scheme", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "that redirects plasticity to the weights that are least important to previously learned study sessions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "(Kirkpatrick et al., 2017). After EWC learns a study session, it uses the training data to build a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "Fisher matrix that determines the importance of each feature to the classification task it just learned.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 453, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 453, + 331 + ], + "score": 1.0, + "content": "EWC was shown to work poorly at incremental class learning in Kemker et al. (2018).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "The Fixed Expansion Layer (FEL) model mitigates catastrophic forgetting by using sparse updates", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "(Coop et al., 2013). FEL uses two hidden layers, where the second hidden layer (i.e., the FEL layer)", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "has connectivity constraints. The FEL layer is much larger than the first hidden layer, is sparsely", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 368, + 506, + 381 + ], + "score": 1.0, + "content": "populated with excitatory and inhibitory weights, and is not updated during training. This limits", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "learning of dense shared representations, which reduces the risk of learning interfering with old", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 444, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 444, + 403 + ], + "score": 1.0, + "content": "memories. FEL requires a large number of units to work well (Kemker et al., 2018).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 324, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 325, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 325, + 421 + ], + "score": 1.0, + "content": "Gepperth & Karaoguz (2016) introduced a new ap-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 418, + 325, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 325, + 431 + ], + "score": 1.0, + "content": "proach for incremental learning, which we call Gepp-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 429, + 325, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 325, + 441 + ], + "score": 1.0, + "content": "Net. GeppNet uses a self-organizing map (SOM) to re-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 440, + 325, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 325, + 452 + ], + "score": 1.0, + "content": "organize the input onto a two-dimensional lattice. This", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 452, + 325, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 325, + 462 + ], + "score": 1.0, + "content": "serves as a long-term memory, which is fed into a sim-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 462, + 325, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 325, + 473 + ], + "score": 1.0, + "content": "ple linear layer for classification. After the SOM is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 473, + 325, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 325, + 486 + ], + "score": 1.0, + "content": "initialized, it can only be updated if the input is suffi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 483, + 325, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 325, + 497 + ], + "score": 1.0, + "content": "ciently novel. This prevents the model from forgetting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 495, + 325, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 325, + 507 + ], + "score": 1.0, + "content": "older data too quickly. GeppNet also uses rehearsal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 505, + 325, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 325, + 519 + ], + "score": 1.0, + "content": "using all previous training data. A variant of Gepp-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 517, + 325, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 325, + 529 + ], + "score": 1.0, + "content": "Net, GeppNet+STM, uses a fixed-size memory buffer", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 528, + 325, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 325, + 539 + ], + "score": 1.0, + "content": "to store novel examples. When this buffer is full, it", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 539, + 325, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 325, + 551 + ], + "score": 1.0, + "content": "replaces the oldest example. During pre-defined in-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 548, + 325, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 325, + 563 + ], + "score": 1.0, + "content": "tervals, the buffer is used to train the model. Gepp-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 560, + 325, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 149, + 572 + ], + "score": 0.69, + "content": "\\mathbf { \\Gamma } _ { \\mathbf { N e t + S T M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 560, + 325, + 573 + ], + "score": 1.0, + "content": "is better at retaining base-knowledge since", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 572, + 325, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 325, + 583 + ], + "score": 1.0, + "content": "it only trains during its consolidation phase, but the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 581, + 325, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 325, + 596 + ], + "score": 1.0, + "content": "STM-free version learns new data better because it up-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 594, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 289, + 606 + ], + "score": 1.0, + "content": "dates the model on every novel labeled input.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 36.5 + }, + { + "type": "image", + "bbox": [ + 333, + 410, + 503, + 538 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 333, + 410, + 503, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 333, + 410, + 503, + 538 + ], + "spans": [ + { + "bbox": [ + 333, + 410, + 503, + 538 + ], + "score": 0.971, + "type": "image", + "image_path": "9487165ad014d2cb3f67829861104cc2ddefd93cc570c6b1da29573fe8bec15b.jpg" + } + ] + } + ], + "index": 50, + "virtual_lines": [ + { + "bbox": [ + 333, + 410, + 503, + 424.22222222222223 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 333, + 424.22222222222223, + 503, + 438.44444444444446 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 333, + 438.44444444444446, + 503, + 452.6666666666667 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 333, + 452.6666666666667, + 503, + 466.8888888888889 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 333, + 466.8888888888889, + 503, + 481.11111111111114 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 333, + 481.11111111111114, + 503, + 495.33333333333337 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 333, + 495.33333333333337, + 503, + 509.5555555555556 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 333, + 509.5555555555556, + 503, + 523.7777777777778 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 333, + 523.7777777777778, + 503, + 538.0 + ], + "spans": [], + "index": 54 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 332, + 549, + 504, + 604 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 331, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 331, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "Figure 2: iCaRL’s performance depends", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 331, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 331, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "heavily on the number of exemplars per", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 331, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 331, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "class (EPC) that it stores. Reducing EPC", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 331, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 331, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "from 20 (blue) to 1 (red) severely impairs", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 331, + 593, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 331, + 593, + 482, + 605 + ], + "score": 1.0, + "content": "its ability to recall older information.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 57 + } + ], + "index": 53.5 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 721 + ], + "lines": [ + { + "bbox": [ + 104, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "iCaRL (Rebuffi et al., 2017) is an incremental class learning framework. Rather than directly using", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 104, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "a DNN for classification, iCaRL uses it for supervised representation learning. During a study", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 406, + 645 + ], + "score": 1.0, + "content": "session, iCaRL updates a DNN using the study session’s data and a set of", + "type": "text" + }, + { + "bbox": [ + 406, + 633, + 414, + 642 + ], + "score": 0.78, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "stored examples from", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 174, + 656 + ], + "score": 1.0, + "content": "earlier sessions", + "type": "text" + }, + { + "bbox": [ + 174, + 644, + 225, + 655 + ], + "score": 0.89, + "content": "( J = 2 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "for CIFAR-100 in their paper), which is a kind of rehearsal. After", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 191, + 667 + ], + "score": 1.0, + "content": "a study session, the", + "type": "text" + }, + { + "bbox": [ + 191, + 655, + 199, + 665 + ], + "score": 0.77, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "examples retained are carefully chosen using herding. After learning the", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 104, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 245, + 678 + ], + "score": 1.0, + "content": "entire dataset, iCaRL has retained", + "type": "text" + }, + { + "bbox": [ + 245, + 665, + 265, + 677 + ], + "score": 0.91, + "content": "J / T", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 664, + 370, + 678 + ], + "score": 1.0, + "content": "exemplars per class (e.g.,", + "type": "text" + }, + { + "bbox": [ + 371, + 665, + 415, + 677 + ], + "score": 0.92, + "content": "J / \\bar { T } = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "for CIFAR-100). The", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "DNN in iCaRL is then used to compute an embedding for each stored example, and then the mean", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "embedding for each class seen is computed. To classify a new instance, the DNN is used to compute", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "an embedding for it, and then the class with the nearest mean embedding is assigned. iCaRL’s", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 105, + 709, + 468, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 468, + 722 + ], + "score": 1.0, + "content": "performance is heavily influenced by the number of examples it stores, as shown in Fig. 2.", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 64.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 203 + ], + "lines": [], + "index": 5, + "bbox_fs": [ + 105, + 81, + 506, + 205 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "Recently there has been renewed interest in solving catastrophic forgetting in supervised learning.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "Many new methods are designed to mitigate catastrophic forgetting when each study session con-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "tains a permuted version of the entire training dataset (see Goodfellow et al. (2013)). Unlike incre-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "mental class learning, all labels are contained in each study session. PathNet uses an evolutionary", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "algorithm to find the optimal path through a large DNN, and then freezes the weights along that path", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "(Fernando et al., 2017). It assumes all classes are seen in each study session, and it is not capable of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "incremental class learning. Elastic Weight Consolidation (EWC) employs a regularization scheme", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "that redirects plasticity to the weights that are least important to previously learned study sessions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "(Kirkpatrick et al., 2017). After EWC learns a study session, it uses the training data to build a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "Fisher matrix that determines the importance of each feature to the classification task it just learned.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 453, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 453, + 331 + ], + "score": 1.0, + "content": "EWC was shown to work poorly at incremental class learning in Kemker et al. (2018).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 208, + 506, + 331 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "The Fixed Expansion Layer (FEL) model mitigates catastrophic forgetting by using sparse updates", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "(Coop et al., 2013). FEL uses two hidden layers, where the second hidden layer (i.e., the FEL layer)", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "has connectivity constraints. The FEL layer is much larger than the first hidden layer, is sparsely", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 368, + 506, + 381 + ], + "score": 1.0, + "content": "populated with excitatory and inhibitory weights, and is not updated during training. This limits", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "learning of dense shared representations, which reduces the risk of learning interfering with old", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 444, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 444, + 403 + ], + "score": 1.0, + "content": "memories. FEL requires a large number of units to work well (Kemker et al., 2018).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 335, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 324, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 325, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 325, + 421 + ], + "score": 1.0, + "content": "Gepperth & Karaoguz (2016) introduced a new ap-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 418, + 325, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 325, + 431 + ], + "score": 1.0, + "content": "proach for incremental learning, which we call Gepp-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 429, + 325, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 325, + 441 + ], + "score": 1.0, + "content": "Net. GeppNet uses a self-organizing map (SOM) to re-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 440, + 325, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 325, + 452 + ], + "score": 1.0, + "content": "organize the input onto a two-dimensional lattice. This", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 452, + 325, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 325, + 462 + ], + "score": 1.0, + "content": "serves as a long-term memory, which is fed into a sim-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 462, + 325, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 325, + 473 + ], + "score": 1.0, + "content": "ple linear layer for classification. After the SOM is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 473, + 325, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 325, + 486 + ], + "score": 1.0, + "content": "initialized, it can only be updated if the input is suffi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 483, + 325, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 325, + 497 + ], + "score": 1.0, + "content": "ciently novel. This prevents the model from forgetting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 495, + 325, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 325, + 507 + ], + "score": 1.0, + "content": "older data too quickly. GeppNet also uses rehearsal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 505, + 325, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 325, + 519 + ], + "score": 1.0, + "content": "using all previous training data. A variant of Gepp-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 517, + 325, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 325, + 529 + ], + "score": 1.0, + "content": "Net, GeppNet+STM, uses a fixed-size memory buffer", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 528, + 325, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 325, + 539 + ], + "score": 1.0, + "content": "to store novel examples. When this buffer is full, it", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 539, + 325, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 325, + 551 + ], + "score": 1.0, + "content": "replaces the oldest example. During pre-defined in-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 548, + 325, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 325, + 563 + ], + "score": 1.0, + "content": "tervals, the buffer is used to train the model. Gepp-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 560, + 325, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 149, + 572 + ], + "score": 0.69, + "content": "\\mathbf { \\Gamma } _ { \\mathbf { N e t + S T M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 560, + 325, + 573 + ], + "score": 1.0, + "content": "is better at retaining base-knowledge since", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 572, + 325, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 325, + 583 + ], + "score": 1.0, + "content": "it only trains during its consolidation phase, but the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 581, + 325, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 325, + 596 + ], + "score": 1.0, + "content": "STM-free version learns new data better because it up-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 594, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 289, + 606 + ], + "score": 1.0, + "content": "dates the model on every novel labeled input.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 406, + 325, + 606 + ] + }, + { + "type": "image", + "bbox": [ + 333, + 410, + 503, + 538 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 333, + 410, + 503, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 333, + 410, + 503, + 538 + ], + "spans": [ + { + "bbox": [ + 333, + 410, + 503, + 538 + ], + "score": 0.971, + "type": "image", + "image_path": "9487165ad014d2cb3f67829861104cc2ddefd93cc570c6b1da29573fe8bec15b.jpg" + } + ] + } + ], + "index": 50, + "virtual_lines": [ + { + "bbox": [ + 333, + 410, + 503, + 424.22222222222223 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 333, + 424.22222222222223, + 503, + 438.44444444444446 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 333, + 438.44444444444446, + 503, + 452.6666666666667 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 333, + 452.6666666666667, + 503, + 466.8888888888889 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 333, + 466.8888888888889, + 503, + 481.11111111111114 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 333, + 481.11111111111114, + 503, + 495.33333333333337 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 333, + 495.33333333333337, + 503, + 509.5555555555556 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 333, + 509.5555555555556, + 503, + 523.7777777777778 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 333, + 523.7777777777778, + 503, + 538.0 + ], + "spans": [], + "index": 54 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 332, + 549, + 504, + 604 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 331, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 331, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "Figure 2: iCaRL’s performance depends", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 331, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 331, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "heavily on the number of exemplars per", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 331, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 331, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "class (EPC) that it stores. Reducing EPC", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 331, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 331, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "from 20 (blue) to 1 (red) severely impairs", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 331, + 593, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 331, + 593, + 482, + 605 + ], + "score": 1.0, + "content": "its ability to recall older information.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 57 + } + ], + "index": 53.5 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 721 + ], + "lines": [ + { + "bbox": [ + 104, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "iCaRL (Rebuffi et al., 2017) is an incremental class learning framework. Rather than directly using", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 104, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "a DNN for classification, iCaRL uses it for supervised representation learning. During a study", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 406, + 645 + ], + "score": 1.0, + "content": "session, iCaRL updates a DNN using the study session’s data and a set of", + "type": "text" + }, + { + "bbox": [ + 406, + 633, + 414, + 642 + ], + "score": 0.78, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "stored examples from", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 174, + 656 + ], + "score": 1.0, + "content": "earlier sessions", + "type": "text" + }, + { + "bbox": [ + 174, + 644, + 225, + 655 + ], + "score": 0.89, + "content": "( J = 2 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "for CIFAR-100 in their paper), which is a kind of rehearsal. After", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 191, + 667 + ], + "score": 1.0, + "content": "a study session, the", + "type": "text" + }, + { + "bbox": [ + 191, + 655, + 199, + 665 + ], + "score": 0.77, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "examples retained are carefully chosen using herding. After learning the", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 104, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 245, + 678 + ], + "score": 1.0, + "content": "entire dataset, iCaRL has retained", + "type": "text" + }, + { + "bbox": [ + 245, + 665, + 265, + 677 + ], + "score": 0.91, + "content": "J / T", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 664, + 370, + 678 + ], + "score": 1.0, + "content": "exemplars per class (e.g.,", + "type": "text" + }, + { + "bbox": [ + 371, + 665, + 415, + 677 + ], + "score": 0.92, + "content": "J / \\bar { T } = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "for CIFAR-100). The", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "DNN in iCaRL is then used to compute an embedding for each stored example, and then the mean", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "embedding for each class seen is computed. To classify a new instance, the DNN is used to compute", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "an embedding for it, and then the class with the nearest mean embedding is assigned. iCaRL’s", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 105, + 709, + 468, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 468, + 722 + ], + "score": 1.0, + "content": "performance is heavily influenced by the number of examples it stores, as shown in Fig. 2.", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 64.5, + "bbox_fs": [ + 104, + 609, + 506, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 408, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 410, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 410, + 96 + ], + "score": 1.0, + "content": "3 MAMMALIAN MEMORY: NEUROSCIENCE AND MODELS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 505, + 119 + ], + "score": 1.0, + "content": "FearNet is heavily inspired by the dual-memory model of mammalian memory (McClelland et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 505, + 131 + ], + "score": 1.0, + "content": "1995), which has considerable experimental support from neuroscience (Frankland et al., 2004;", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "score": 1.0, + "content": "Takashima et al., 2006; Kitamura et al., 2017; Bontempi et al., 1999; Taupin & Gage, 2002; Gais", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "et al., 2007). This theory proposes that HC and mPFC operate as complementary memory systems,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 150, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 164 + ], + "score": 1.0, + "content": "where HC is responsible for recalling recent memories and mPFC is responsible for recalling remote", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "(mature) memories. GeppNet is the most recent DNN to be based on this theory, but it was also", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "independently explored in the 1990s in French (1997) and Ans & Rousset (1997). In this section,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 351, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 351, + 195 + ], + "score": 1.0, + "content": "we review some of the evidence for the dual-memory model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 200, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "One of the major reasons why HC is thought to be responsible for recent memories is that if HC is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "score": 1.0, + "content": "bilaterally destroyed, then anterograde amnesia occurs with old memories for semantic information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 223, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 504, + 235 + ], + "score": 1.0, + "content": "preserved. One mechanism HC may use to facilitate creating new memories is adult neurogenesis.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 233, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 506, + 246 + ], + "score": 1.0, + "content": "This occurs in HC’s dentate gyrus (Altman, 1963; Eriksson et al., 1998). The new neurons have", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 244, + 412, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 412, + 258 + ], + "score": 1.0, + "content": "higher initial plasticity, but it reduces as time progresses (Deng et al., 2010).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 274 + ], + "score": 1.0, + "content": "In contrast, mPFC is responsible for the recall of remote (long-term) memories (Bontempi et al.,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "1999). Taupin & Gage (2002) and Gais et al. (2007) showed that mPFC plays a strong role in mem-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "ory consolidation during REM sleep. McClelland et al. (1995) and Euston et al. (2012) theorized", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 295, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 506, + 307 + ], + "score": 1.0, + "content": "that, during sleep, HC reactivates recent memories to prevent forgetting which causes these recent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "memories to replay in mPFC as well, with dreams possibly being caused by this process. After", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "memories are transferred from HC to mPFC, evidence suggests that corresponding memory in HC", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 327, + 195, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 195, + 340 + ], + "score": 1.0, + "content": "is erased (Poe, 2017).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Recently, Kitamura et al. (2017) performed contextual fear conditioning (CFC) experiments in mice", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "to trace the formation and consolidation of recent memories to long-term storage. CFC experiments", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "score": 1.0, + "content": "involve shocking mice while subjecting them to various visual stimuli (i.e., colored lights). They", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "found that BLA, which is responsible for regulating the brain’s fear response, would shift where", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "it retrieved the corresponding memory from (HC or mPFC) as that memory was consolidated over", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 399, + 476, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 476, + 412 + ], + "score": 1.0, + "content": "time. FearNet follows the memory consolidation theory proposed by Kitamura et al. (2017).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 108, + 428, + 243, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 245, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 245, + 443 + ], + "score": 1.0, + "content": "4 THE FEARNET MODEL", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "score": 1.0, + "content": "FearNet has two complementary memory centers, 1) a short-term memory system that immediately", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "learns new information for recent recall (HC) and 2) a DNN for the storage of remote memories", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "(mPFC). FearNet also has a separate BLA network that determines which memory center contains", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "the associated memory required for prediction. During sleep phases, FearNet uses a generative", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "score": 1.0, + "content": "model to consolidate data from HC to mPFC through pseudorehearsal. Pseudocode for FearNet is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 508, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 522 + ], + "score": 1.0, + "content": "provided in the supplemental material. Because the focus of our work is not representation learning,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 520, + 432, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 432, + 531 + ], + "score": 1.0, + "content": "we use pre-trained ResNet embeddings to obtain features that are fed to FearNet.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 546, + 245, + 557 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 245, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 245, + 558 + ], + "score": 1.0, + "content": "4.1 DUAL-MEMORY STORAGE", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 567, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 504, + 579 + ], + "score": 1.0, + "content": "FearNet’s HC model is a variant of a probabilistic neural network (Specht, 1990). HC computes class", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "conditional probabilities using stored training examples. Formally, HC estimates the probability that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 590, + 292, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 198, + 602 + ], + "score": 1.0, + "content": "an input feature vector", + "type": "text" + }, + { + "bbox": [ + 199, + 591, + 207, + 599 + ], + "score": 0.55, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 590, + 273, + 602 + ], + "score": 1.0, + "content": "belongs to class", + "type": "text" + }, + { + "bbox": [ + 273, + 590, + 280, + 599 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 590, + 292, + 602 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 615, + 364, + 643 + ], + "lines": [ + { + "bbox": [ + 247, + 615, + 364, + 643 + ], + "spans": [ + { + "bbox": [ + 247, + 615, + 364, + 643 + ], + "score": 0.93, + "content": "P _ { H C } \\left( C = k | \\mathbf { x } \\right) = \\frac { \\beta _ { k } } { \\sum _ { k ^ { \\prime } } \\beta _ { k ^ { \\prime } } }", + "type": "interline_equation", + "image_path": "e2c3ab335fbf48b2d06228c1c54eb2baaca69897d20b8d1313af78ffb2537d86.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 615, + 364, + 629.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 247, + 629.0, + 364, + 643.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 649, + 451, + 678 + ], + "lines": [ + { + "bbox": [ + 154, + 649, + 451, + 678 + ], + "spans": [ + { + "bbox": [ + 154, + 649, + 451, + 678 + ], + "score": 0.54, + "content": "\\beta _ { k } = \\left\\{ \\begin{array} { c c } { \\big ( \\epsilon + \\operatorname* { m i n } _ { j } \\| \\mathbf x - \\mathbf u _ { k , j } \\| _ { 2 } \\big ) ^ { - 1 } } & { \\mathrm { i f ~ H C ~ c o n t a i n s ~ i n s t a n c e s ~ o f ~ c l a s s ~ } k } \\\\ { 0 } & { \\mathrm { o t h e r w i s e } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "fdf5250852e0be7651be399d9809bb0f6acf66a44b78ca29facace7bfd947b5a.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 154, + 649, + 451, + 658.6666666666666 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 154, + 658.6666666666666, + 451, + 668.3333333333333 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 154, + 668.3333333333333, + 451, + 677.9999999999999 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 133, + 694 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 682, + 157, + 692 + ], + "score": 0.9, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 680, + 290, + 694 + ], + "score": 1.0, + "content": "is a regularization parameter and", + "type": "text" + }, + { + "bbox": [ + 290, + 683, + 308, + 694 + ], + "score": 0.9, + "content": "{ \\mathbf { u } } _ { k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 680, + 333, + 694 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 333, + 682, + 340, + 693 + ], + "score": 0.82, + "content": "j ^ { : }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 680, + 479, + 694 + ], + "score": 1.0, + "content": "’th stored exemplar in HC for class", + "type": "text" + }, + { + "bbox": [ + 479, + 682, + 485, + 691 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 680, + 505, + 694 + ], + "score": 1.0, + "content": ". All", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 693, + 396, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 396, + 705 + ], + "score": 1.0, + "content": "exemplars are removed from HC after they are consolidated into mPFC.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "FearNet’s mPFC is implemented using a DNN trained both to reconstruct its input using a symmetric", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 298, + 734 + ], + "score": 1.0, + "content": "encoder-decoder (autoencoder) and to compute", + "type": "text" + }, + { + "bbox": [ + 298, + 720, + 376, + 732 + ], + "score": 0.67, + "content": "P _ { m P F C } \\left( C = k | \\mathbf { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 721, + 506, + 734 + ], + "score": 1.0, + "content": ". The autoencoder enables us to", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 408, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 410, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 410, + 96 + ], + "score": 1.0, + "content": "3 MAMMALIAN MEMORY: NEUROSCIENCE AND MODELS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 505, + 119 + ], + "score": 1.0, + "content": "FearNet is heavily inspired by the dual-memory model of mammalian memory (McClelland et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 505, + 131 + ], + "score": 1.0, + "content": "1995), which has considerable experimental support from neuroscience (Frankland et al., 2004;", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "score": 1.0, + "content": "Takashima et al., 2006; Kitamura et al., 2017; Bontempi et al., 1999; Taupin & Gage, 2002; Gais", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "et al., 2007). This theory proposes that HC and mPFC operate as complementary memory systems,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 150, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 164 + ], + "score": 1.0, + "content": "where HC is responsible for recalling recent memories and mPFC is responsible for recalling remote", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "(mature) memories. GeppNet is the most recent DNN to be based on this theory, but it was also", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "independently explored in the 1990s in French (1997) and Ans & Rousset (1997). In this section,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 351, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 351, + 195 + ], + "score": 1.0, + "content": "we review some of the evidence for the dual-memory model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 107, + 506, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 200, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "One of the major reasons why HC is thought to be responsible for recent memories is that if HC is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "score": 1.0, + "content": "bilaterally destroyed, then anterograde amnesia occurs with old memories for semantic information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 223, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 504, + 235 + ], + "score": 1.0, + "content": "preserved. One mechanism HC may use to facilitate creating new memories is adult neurogenesis.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 233, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 506, + 246 + ], + "score": 1.0, + "content": "This occurs in HC’s dentate gyrus (Altman, 1963; Eriksson et al., 1998). The new neurons have", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 244, + 412, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 412, + 258 + ], + "score": 1.0, + "content": "higher initial plasticity, but it reduces as time progresses (Deng et al., 2010).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 201, + 506, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 274 + ], + "score": 1.0, + "content": "In contrast, mPFC is responsible for the recall of remote (long-term) memories (Bontempi et al.,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "1999). Taupin & Gage (2002) and Gais et al. (2007) showed that mPFC plays a strong role in mem-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "ory consolidation during REM sleep. McClelland et al. (1995) and Euston et al. (2012) theorized", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 295, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 506, + 307 + ], + "score": 1.0, + "content": "that, during sleep, HC reactivates recent memories to prevent forgetting which causes these recent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "memories to replay in mPFC as well, with dreams possibly being caused by this process. After", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "memories are transferred from HC to mPFC, evidence suggests that corresponding memory in HC", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 327, + 195, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 195, + 340 + ], + "score": 1.0, + "content": "is erased (Poe, 2017).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 262, + 506, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Recently, Kitamura et al. (2017) performed contextual fear conditioning (CFC) experiments in mice", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "to trace the formation and consolidation of recent memories to long-term storage. CFC experiments", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "score": 1.0, + "content": "involve shocking mice while subjecting them to various visual stimuli (i.e., colored lights). They", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "found that BLA, which is responsible for regulating the brain’s fear response, would shift where", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "it retrieved the corresponding memory from (HC or mPFC) as that memory was consolidated over", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 399, + 476, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 476, + 412 + ], + "score": 1.0, + "content": "time. FearNet follows the memory consolidation theory proposed by Kitamura et al. (2017).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 344, + 506, + 412 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 428, + 243, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 245, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 245, + 443 + ], + "score": 1.0, + "content": "4 THE FEARNET MODEL", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "score": 1.0, + "content": "FearNet has two complementary memory centers, 1) a short-term memory system that immediately", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "learns new information for recent recall (HC) and 2) a DNN for the storage of remote memories", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "(mPFC). FearNet also has a separate BLA network that determines which memory center contains", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "the associated memory required for prediction. During sleep phases, FearNet uses a generative", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "score": 1.0, + "content": "model to consolidate data from HC to mPFC through pseudorehearsal. Pseudocode for FearNet is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 508, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 522 + ], + "score": 1.0, + "content": "provided in the supplemental material. Because the focus of our work is not representation learning,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 520, + 432, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 432, + 531 + ], + "score": 1.0, + "content": "we use pre-trained ResNet embeddings to obtain features that are fed to FearNet.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 453, + 506, + 531 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 546, + 245, + 557 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 245, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 245, + 558 + ], + "score": 1.0, + "content": "4.1 DUAL-MEMORY STORAGE", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 567, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 504, + 579 + ], + "score": 1.0, + "content": "FearNet’s HC model is a variant of a probabilistic neural network (Specht, 1990). HC computes class", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "conditional probabilities using stored training examples. Formally, HC estimates the probability that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 590, + 292, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 198, + 602 + ], + "score": 1.0, + "content": "an input feature vector", + "type": "text" + }, + { + "bbox": [ + 199, + 591, + 207, + 599 + ], + "score": 0.55, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 590, + 273, + 602 + ], + "score": 1.0, + "content": "belongs to class", + "type": "text" + }, + { + "bbox": [ + 273, + 590, + 280, + 599 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 590, + 292, + 602 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 566, + 505, + 602 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 615, + 364, + 643 + ], + "lines": [ + { + "bbox": [ + 247, + 615, + 364, + 643 + ], + "spans": [ + { + "bbox": [ + 247, + 615, + 364, + 643 + ], + "score": 0.93, + "content": "P _ { H C } \\left( C = k | \\mathbf { x } \\right) = \\frac { \\beta _ { k } } { \\sum _ { k ^ { \\prime } } \\beta _ { k ^ { \\prime } } }", + "type": "interline_equation", + "image_path": "e2c3ab335fbf48b2d06228c1c54eb2baaca69897d20b8d1313af78ffb2537d86.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 615, + 364, + 629.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 247, + 629.0, + 364, + 643.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 649, + 451, + 678 + ], + "lines": [ + { + "bbox": [ + 154, + 649, + 451, + 678 + ], + "spans": [ + { + "bbox": [ + 154, + 649, + 451, + 678 + ], + "score": 0.54, + "content": "\\beta _ { k } = \\left\\{ \\begin{array} { c c } { \\big ( \\epsilon + \\operatorname* { m i n } _ { j } \\| \\mathbf x - \\mathbf u _ { k , j } \\| _ { 2 } \\big ) ^ { - 1 } } & { \\mathrm { i f ~ H C ~ c o n t a i n s ~ i n s t a n c e s ~ o f ~ c l a s s ~ } k } \\\\ { 0 } & { \\mathrm { o t h e r w i s e } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "fdf5250852e0be7651be399d9809bb0f6acf66a44b78ca29facace7bfd947b5a.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 154, + 649, + 451, + 658.6666666666666 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 154, + 658.6666666666666, + 451, + 668.3333333333333 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 154, + 668.3333333333333, + 451, + 677.9999999999999 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 133, + 694 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 682, + 157, + 692 + ], + "score": 0.9, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 680, + 290, + 694 + ], + "score": 1.0, + "content": "is a regularization parameter and", + "type": "text" + }, + { + "bbox": [ + 290, + 683, + 308, + 694 + ], + "score": 0.9, + "content": "{ \\mathbf { u } } _ { k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 680, + 333, + 694 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 333, + 682, + 340, + 693 + ], + "score": 0.82, + "content": "j ^ { : }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 680, + 479, + 694 + ], + "score": 1.0, + "content": "’th stored exemplar in HC for class", + "type": "text" + }, + { + "bbox": [ + 479, + 682, + 485, + 691 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 680, + 505, + 694 + ], + "score": 1.0, + "content": ". All", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 693, + 396, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 396, + 705 + ], + "score": 1.0, + "content": "exemplars are removed from HC after they are consolidated into mPFC.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 680, + 505, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "FearNet’s mPFC is implemented using a DNN trained both to reconstruct its input using a symmetric", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 298, + 734 + ], + "score": 1.0, + "content": "encoder-decoder (autoencoder) and to compute", + "type": "text" + }, + { + "bbox": [ + 298, + 720, + 376, + 732 + ], + "score": 0.67, + "content": "P _ { m P F C } \\left( C = k | \\mathbf { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 721, + 506, + 734 + ], + "score": 1.0, + "content": ". The autoencoder enables us to", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 709, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 86, + 495, + 197 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 86, + 495, + 197 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 86, + 495, + 197 + ], + "spans": [ + { + "bbox": [ + 112, + 86, + 495, + 197 + ], + "score": 0.966, + "type": "image", + "image_path": "0a31f3f9e3363e7a1f2b1d9c767fa264123bab51cadfb522dfebe045d764880d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 86, + 495, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 123.0, + 495, + 160.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 160.0, + 495, + 197.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 209, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "score": 1.0, + "content": "Figure 3: The mPFC and BLA sub-systems in FearNet. mPFC is responsible for the long-term", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "storage of remote memories. BLA is used during prediction time to determine if the memory should", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 230, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 290, + 245 + ], + "score": 1.0, + "content": "be recalled from short- or long-term memory.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 261, + 496, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 497, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 497, + 274 + ], + "score": 1.0, + "content": "use pseudorehearsal, which is described in more detail in Sec. 4.2. The loss function for mPFC is", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 277, + 363, + 290 + ], + "lines": [ + { + "bbox": [ + 247, + 277, + 363, + 290 + ], + "spans": [ + { + "bbox": [ + 247, + 277, + 363, + 290 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { m P F C } = \\mathcal { L } _ { c l a s s } + \\mathcal { L } _ { r e c o n } ,", + "type": "interline_equation", + "image_path": "ad90f62032b729b0de886c36fc2533201c3b200ff65e31051b2e5f2c560fbad0.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 247, + 277, + 363, + 290 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 133, + 309 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 295, + 160, + 306 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { c l a s s }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 294, + 317, + 309 + ], + "score": 1.0, + "content": "is the supervised classification loss and", + "type": "text" + }, + { + "bbox": [ + 317, + 295, + 346, + 306 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { r e c o n }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 294, + 506, + 309 + ], + "score": 1.0, + "content": "is the unsupervised reconstruction loss,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 306, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 227, + 319 + ], + "score": 1.0, + "content": "as illustrated in Fig. 3(a). For", + "type": "text" + }, + { + "bbox": [ + 228, + 306, + 254, + 317 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { c l a s s }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 306, + 381, + 319 + ], + "score": 1.0, + "content": ", we use standard softmax loss.", + "type": "text" + }, + { + "bbox": [ + 381, + 306, + 410, + 317 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { r e c o n }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 306, + 506, + 319 + ], + "score": 1.0, + "content": "is the weighted sum of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 316, + 441, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 441, + 330 + ], + "score": 1.0, + "content": "mean squared error (MSE) reconstruction losses from each layer, which is given by", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 332, + 412, + 369 + ], + "lines": [ + { + "bbox": [ + 199, + 332, + 412, + 369 + ], + "spans": [ + { + "bbox": [ + 199, + 332, + 412, + 369 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { r e c o n } = \\sum _ { j = 0 } ^ { M } \\sum _ { i = 0 } ^ { H _ { j } - 1 } \\left. h _ { e n c o d e r , ( i , j ) } - h _ { d e c o d e r , ( i , j ) } \\right. _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "4fb0b30c9bb5b5142a56cca59c9db179b0d074e64254bf766159a438f9aca0d9.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 199, + 332, + 412, + 350.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 199, + 350.5, + 412, + 369.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 371, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 104, + 368, + 503, + 388 + ], + "spans": [ + { + "bbox": [ + 104, + 368, + 133, + 388 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 372, + 145, + 382 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 368, + 272, + 388 + ], + "score": 1.0, + "content": "is the number of mPFC layers,", + "type": "text" + }, + { + "bbox": [ + 273, + 372, + 286, + 384 + ], + "score": 0.9, + "content": "H _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 368, + 441, + 388 + ], + "score": 1.0, + "content": "is the number of hidden units in layer", + "type": "text" + }, + { + "bbox": [ + 442, + 373, + 447, + 383 + ], + "score": 0.74, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 368, + 451, + 388 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 452, + 372, + 503, + 385 + ], + "score": 0.77, + "content": "h _ { e n c o d e r , ( i , j ) }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 383, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 124, + 398 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 384, + 176, + 397 + ], + "score": 0.88, + "content": "h _ { d e c o d e r , ( i , j ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 383, + 372, + 398 + ], + "score": 1.0, + "content": "are the outputs of the encoder/decoder at layer", + "type": "text" + }, + { + "bbox": [ + 373, + 384, + 379, + 395 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 383, + 452, + 398 + ], + "score": 1.0, + "content": "respectively, and", + "type": "text" + }, + { + "bbox": [ + 452, + 384, + 463, + 395 + ], + "score": 0.88, + "content": "\\lambda _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 383, + 506, + 398 + ], + "score": 1.0, + "content": "is the re-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 393, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 407 + ], + "score": 1.0, + "content": "construction weight for that layer. mPFC is similar to a Ladder Network (Rasmus et al., 2015),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "which combines classification and reconstruction to improve regularization, especially during low-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 414, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 181, + 430 + ], + "score": 1.0, + "content": "shot learning. The", + "type": "text" + }, + { + "bbox": [ + 181, + 416, + 192, + 428 + ], + "score": 0.88, + "content": "\\lambda _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 414, + 378, + 430 + ], + "score": 1.0, + "content": "hyperparameters were found empirically, with", + "type": "text" + }, + { + "bbox": [ + 378, + 416, + 389, + 427 + ], + "score": 0.88, + "content": "\\lambda _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 414, + 506, + 430 + ], + "score": 1.0, + "content": "being largest and decreasing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 425, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 440 + ], + "score": 1.0, + "content": "for deeper layers (see supplementary material). This prioritizes the reconstruction task, which makes", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "the generated pseudo-examples more realistic. When training is completed during a study session,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 449, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 505, + 460 + ], + "score": 1.0, + "content": "all of the data in HC is pushed through the encoder to extract a dense feature representation of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 334, + 473 + ], + "score": 1.0, + "content": "original data, and then we compute a mean feature vector", + "type": "text" + }, + { + "bbox": [ + 334, + 461, + 345, + 471 + ], + "score": 0.86, + "content": "\\mu _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 459, + 435, + 473 + ], + "score": 1.0, + "content": "and covariance matrix", + "type": "text" + }, + { + "bbox": [ + 436, + 460, + 448, + 471 + ], + "score": 0.89, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "for each class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 107, + 473, + 112, + 480 + ], + "score": 0.47, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 470, + 506, + 483 + ], + "score": 1.0, + "content": ". These are stored and used to generate pseudo-examples during consolidation (see Sec. 4.2). We", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 462, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 462, + 493 + ], + "score": 1.0, + "content": "study FearNet’s performance as a function of how much data is stored in HC in Sec. 6.2.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 106, + 506, + 354, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 356, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 356, + 519 + ], + "score": 1.0, + "content": "4.2 PSEUDOREHEARSAL FOR MEMORY CONSOLIDATION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 526, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 540 + ], + "score": 1.0, + "content": "During FearNet’s sleep phase, the original inputs stored in HC are transferred to mPFC using", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "pseudo-examples created by an autoencoder. This process is known as intrinsic replay, and it was", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 548, + 331, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 331, + 562 + ], + "score": 1.0, + "content": "used by Draelos et al. (2017) for unsupervised learning.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 389, + 578 + ], + "score": 1.0, + "content": "Using the class statistics from the encoder, pseudo-examples for class", + "type": "text" + }, + { + "bbox": [ + 390, + 568, + 396, + 576 + ], + "score": 0.62, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "are generated by sampling", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 197, + 590 + ], + "score": 1.0, + "content": "a Gaussian with mean", + "type": "text" + }, + { + "bbox": [ + 197, + 578, + 208, + 588 + ], + "score": 0.85, + "content": "\\mu _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 576, + 300, + 590 + ], + "score": 1.0, + "content": "and covariance matrix", + "type": "text" + }, + { + "bbox": [ + 300, + 577, + 313, + 588 + ], + "score": 0.88, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 576, + 351, + 590 + ], + "score": 1.0, + "content": "to obtain", + "type": "text" + }, + { + "bbox": [ + 352, + 577, + 377, + 588 + ], + "score": 0.91, + "content": "\\hat { \\mathbf { x } } _ { r a n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 576, + 406, + 590 + ], + "score": 1.0, + "content": ". Then,", + "type": "text" + }, + { + "bbox": [ + 407, + 577, + 432, + 588 + ], + "score": 0.91, + "content": "\\hat { \\mathbf { x } } _ { r a n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "is passed through", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "the decoder to generate a pseudo-example. To create a balanced training set, for each class that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 235, + 612 + ], + "score": 1.0, + "content": "mPFC has learned, we generate", + "type": "text" + }, + { + "bbox": [ + 235, + 599, + 252, + 611 + ], + "score": 0.91, + "content": "\\lceil m \\rceil", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 599, + 354, + 612 + ], + "score": 1.0, + "content": "pseudo-examples, where", + "type": "text" + }, + { + "bbox": [ + 354, + 600, + 365, + 609 + ], + "score": 0.72, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "is the average number of examples", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 609, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 621 + ], + "score": 1.0, + "content": "per class stored in HC. The pseudo-examples are mixed with the data in HC, and the mixture is used", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 620, + 470, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 470, + 632 + ], + "score": 1.0, + "content": "to fine-tune mPFC using backpropagation. After consolidation, all units in HC are deleted.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 281, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 282, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 282, + 657 + ], + "score": 1.0, + "content": "4.3 NETWORK SELECTION USING BLA", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "During prediction, FearNet uses the BLA network (Fig. 3(b)) to determine whether to classify an", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "input x using HC or mPFC. This can be challenging because if HC has only been trained on one", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "class, it will put all of its probability mass on that class, whereas mPFC will likely be less confident.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 230, + 711 + ], + "score": 1.0, + "content": "The output of BLA is given by", + "type": "text" + }, + { + "bbox": [ + 230, + 699, + 254, + 711 + ], + "score": 0.77, + "content": "A \\left( \\mathbf { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "and will be a value between 0 and 1, with a 1 indicating mPFC", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "should be used. BLA is trained after each study session using only the data in HC and with pseudo-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "examples generated with mPFC, using the same procedure described in Sec. 4.2. Instead of using", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 86, + 495, + 197 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 86, + 495, + 197 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 86, + 495, + 197 + ], + "spans": [ + { + "bbox": [ + 112, + 86, + 495, + 197 + ], + "score": 0.966, + "type": "image", + "image_path": "0a31f3f9e3363e7a1f2b1d9c767fa264123bab51cadfb522dfebe045d764880d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 86, + 495, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 123.0, + 495, + 160.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 160.0, + 495, + 197.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 209, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "score": 1.0, + "content": "Figure 3: The mPFC and BLA sub-systems in FearNet. mPFC is responsible for the long-term", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "storage of remote memories. BLA is used during prediction time to determine if the memory should", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 230, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 290, + 245 + ], + "score": 1.0, + "content": "be recalled from short- or long-term memory.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 261, + 496, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 497, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 497, + 274 + ], + "score": 1.0, + "content": "use pseudorehearsal, which is described in more detail in Sec. 4.2. The loss function for mPFC is", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 261, + 497, + 274 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 277, + 363, + 290 + ], + "lines": [ + { + "bbox": [ + 247, + 277, + 363, + 290 + ], + "spans": [ + { + "bbox": [ + 247, + 277, + 363, + 290 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { m P F C } = \\mathcal { L } _ { c l a s s } + \\mathcal { L } _ { r e c o n } ,", + "type": "interline_equation", + "image_path": "ad90f62032b729b0de886c36fc2533201c3b200ff65e31051b2e5f2c560fbad0.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 247, + 277, + 363, + 290 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 133, + 309 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 295, + 160, + 306 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { c l a s s }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 294, + 317, + 309 + ], + "score": 1.0, + "content": "is the supervised classification loss and", + "type": "text" + }, + { + "bbox": [ + 317, + 295, + 346, + 306 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { r e c o n }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 294, + 506, + 309 + ], + "score": 1.0, + "content": "is the unsupervised reconstruction loss,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 306, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 227, + 319 + ], + "score": 1.0, + "content": "as illustrated in Fig. 3(a). For", + "type": "text" + }, + { + "bbox": [ + 228, + 306, + 254, + 317 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { c l a s s }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 306, + 381, + 319 + ], + "score": 1.0, + "content": ", we use standard softmax loss.", + "type": "text" + }, + { + "bbox": [ + 381, + 306, + 410, + 317 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { r e c o n }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 306, + 506, + 319 + ], + "score": 1.0, + "content": "is the weighted sum of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 316, + 441, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 441, + 330 + ], + "score": 1.0, + "content": "mean squared error (MSE) reconstruction losses from each layer, which is given by", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 294, + 506, + 330 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 332, + 412, + 369 + ], + "lines": [ + { + "bbox": [ + 199, + 332, + 412, + 369 + ], + "spans": [ + { + "bbox": [ + 199, + 332, + 412, + 369 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { r e c o n } = \\sum _ { j = 0 } ^ { M } \\sum _ { i = 0 } ^ { H _ { j } - 1 } \\left. h _ { e n c o d e r , ( i , j ) } - h _ { d e c o d e r , ( i , j ) } \\right. _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "4fb0b30c9bb5b5142a56cca59c9db179b0d074e64254bf766159a438f9aca0d9.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 199, + 332, + 412, + 350.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 199, + 350.5, + 412, + 369.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 371, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 104, + 368, + 503, + 388 + ], + "spans": [ + { + "bbox": [ + 104, + 368, + 133, + 388 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 372, + 145, + 382 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 368, + 272, + 388 + ], + "score": 1.0, + "content": "is the number of mPFC layers,", + "type": "text" + }, + { + "bbox": [ + 273, + 372, + 286, + 384 + ], + "score": 0.9, + "content": "H _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 368, + 441, + 388 + ], + "score": 1.0, + "content": "is the number of hidden units in layer", + "type": "text" + }, + { + "bbox": [ + 442, + 373, + 447, + 383 + ], + "score": 0.74, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 368, + 451, + 388 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 452, + 372, + 503, + 385 + ], + "score": 0.77, + "content": "h _ { e n c o d e r , ( i , j ) }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 383, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 124, + 398 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 384, + 176, + 397 + ], + "score": 0.88, + "content": "h _ { d e c o d e r , ( i , j ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 383, + 372, + 398 + ], + "score": 1.0, + "content": "are the outputs of the encoder/decoder at layer", + "type": "text" + }, + { + "bbox": [ + 373, + 384, + 379, + 395 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 383, + 452, + 398 + ], + "score": 1.0, + "content": "respectively, and", + "type": "text" + }, + { + "bbox": [ + 452, + 384, + 463, + 395 + ], + "score": 0.88, + "content": "\\lambda _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 383, + 506, + 398 + ], + "score": 1.0, + "content": "is the re-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 393, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 407 + ], + "score": 1.0, + "content": "construction weight for that layer. mPFC is similar to a Ladder Network (Rasmus et al., 2015),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "which combines classification and reconstruction to improve regularization, especially during low-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 414, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 181, + 430 + ], + "score": 1.0, + "content": "shot learning. The", + "type": "text" + }, + { + "bbox": [ + 181, + 416, + 192, + 428 + ], + "score": 0.88, + "content": "\\lambda _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 414, + 378, + 430 + ], + "score": 1.0, + "content": "hyperparameters were found empirically, with", + "type": "text" + }, + { + "bbox": [ + 378, + 416, + 389, + 427 + ], + "score": 0.88, + "content": "\\lambda _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 414, + 506, + 430 + ], + "score": 1.0, + "content": "being largest and decreasing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 425, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 440 + ], + "score": 1.0, + "content": "for deeper layers (see supplementary material). This prioritizes the reconstruction task, which makes", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "the generated pseudo-examples more realistic. When training is completed during a study session,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 449, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 505, + 460 + ], + "score": 1.0, + "content": "all of the data in HC is pushed through the encoder to extract a dense feature representation of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 334, + 473 + ], + "score": 1.0, + "content": "original data, and then we compute a mean feature vector", + "type": "text" + }, + { + "bbox": [ + 334, + 461, + 345, + 471 + ], + "score": 0.86, + "content": "\\mu _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 459, + 435, + 473 + ], + "score": 1.0, + "content": "and covariance matrix", + "type": "text" + }, + { + "bbox": [ + 436, + 460, + 448, + 471 + ], + "score": 0.89, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "for each class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 107, + 473, + 112, + 480 + ], + "score": 0.47, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 470, + 506, + 483 + ], + "score": 1.0, + "content": ". These are stored and used to generate pseudo-examples during consolidation (see Sec. 4.2). We", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 462, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 462, + 493 + ], + "score": 1.0, + "content": "study FearNet’s performance as a function of how much data is stored in HC in Sec. 6.2.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 368, + 506, + 493 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 506, + 354, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 356, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 356, + 519 + ], + "score": 1.0, + "content": "4.2 PSEUDOREHEARSAL FOR MEMORY CONSOLIDATION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 526, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 540 + ], + "score": 1.0, + "content": "During FearNet’s sleep phase, the original inputs stored in HC are transferred to mPFC using", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "pseudo-examples created by an autoencoder. This process is known as intrinsic replay, and it was", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 548, + 331, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 331, + 562 + ], + "score": 1.0, + "content": "used by Draelos et al. (2017) for unsupervised learning.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 525, + 505, + 562 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 389, + 578 + ], + "score": 1.0, + "content": "Using the class statistics from the encoder, pseudo-examples for class", + "type": "text" + }, + { + "bbox": [ + 390, + 568, + 396, + 576 + ], + "score": 0.62, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "are generated by sampling", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 197, + 590 + ], + "score": 1.0, + "content": "a Gaussian with mean", + "type": "text" + }, + { + "bbox": [ + 197, + 578, + 208, + 588 + ], + "score": 0.85, + "content": "\\mu _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 576, + 300, + 590 + ], + "score": 1.0, + "content": "and covariance matrix", + "type": "text" + }, + { + "bbox": [ + 300, + 577, + 313, + 588 + ], + "score": 0.88, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 576, + 351, + 590 + ], + "score": 1.0, + "content": "to obtain", + "type": "text" + }, + { + "bbox": [ + 352, + 577, + 377, + 588 + ], + "score": 0.91, + "content": "\\hat { \\mathbf { x } } _ { r a n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 576, + 406, + 590 + ], + "score": 1.0, + "content": ". Then,", + "type": "text" + }, + { + "bbox": [ + 407, + 577, + 432, + 588 + ], + "score": 0.91, + "content": "\\hat { \\mathbf { x } } _ { r a n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "is passed through", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "the decoder to generate a pseudo-example. To create a balanced training set, for each class that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 235, + 612 + ], + "score": 1.0, + "content": "mPFC has learned, we generate", + "type": "text" + }, + { + "bbox": [ + 235, + 599, + 252, + 611 + ], + "score": 0.91, + "content": "\\lceil m \\rceil", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 599, + 354, + 612 + ], + "score": 1.0, + "content": "pseudo-examples, where", + "type": "text" + }, + { + "bbox": [ + 354, + 600, + 365, + 609 + ], + "score": 0.72, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "is the average number of examples", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 609, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 621 + ], + "score": 1.0, + "content": "per class stored in HC. The pseudo-examples are mixed with the data in HC, and the mixture is used", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 620, + 470, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 470, + 632 + ], + "score": 1.0, + "content": "to fine-tune mPFC using backpropagation. After consolidation, all units in HC are deleted.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 565, + 506, + 632 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 281, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 282, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 282, + 657 + ], + "score": 1.0, + "content": "4.3 NETWORK SELECTION USING BLA", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "During prediction, FearNet uses the BLA network (Fig. 3(b)) to determine whether to classify an", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "input x using HC or mPFC. This can be challenging because if HC has only been trained on one", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "class, it will put all of its probability mass on that class, whereas mPFC will likely be less confident.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 230, + 711 + ], + "score": 1.0, + "content": "The output of BLA is given by", + "type": "text" + }, + { + "bbox": [ + 230, + 699, + 254, + 711 + ], + "score": 0.77, + "content": "A \\left( \\mathbf { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "and will be a value between 0 and 1, with a 1 indicating mPFC", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "should be used. BLA is trained after each study session using only the data in HC and with pseudo-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "examples generated with mPFC, using the same procedure described in Sec. 4.2. Instead of using", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "solely BLA to determine which network to use, we found that combining its output with those of", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 389, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 319, + 106 + ], + "score": 1.0, + "content": "mPFC and HC improved results. The predicted class", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 319, + 94, + 326, + 105 + ], + "score": 0.85, + "content": "\\hat { y }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 326, + 93, + 389, + 106 + ], + "score": 1.0, + "content": "is computed as", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "solely BLA to determine which network to use, we found that combining its output with those of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 389, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 319, + 106 + ], + "score": 1.0, + "content": "mPFC and HC improved results. The predicted class", + "type": "text" + }, + { + "bbox": [ + 319, + 94, + 326, + 105 + ], + "score": 0.85, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 93, + 389, + 106 + ], + "score": 1.0, + "content": "is computed as", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 108, + 452, + 135 + ], + "lines": [ + { + "bbox": [ + 152, + 108, + 452, + 135 + ], + "spans": [ + { + "bbox": [ + 152, + 108, + 452, + 135 + ], + "score": 0.87, + "content": "\\hat { y } = \\left\\{ \\begin{array} { c c } { \\underset { \\mathrm { a r g } } { \\arg \\operatorname* { m a x } } _ { k ^ { \\prime } } P _ { H C } \\left( C = k ^ { \\prime } | \\mathbf { x } \\right) } & { \\mathrm { i f } \\ \\psi > \\operatorname* { m a x } _ { k } P _ { m P F C } \\left( C = k | \\mathbf { x } \\right) } \\\\ { \\underset { \\mathrm { o t h e r w i s e } } { \\arg \\operatorname* { m a x } } } & { \\mathrm { o t h e r w i s e } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "88d3befb10904759c6a11b9ae16fb1bec767d6cd04a2ec0924ee914bf9c57b6f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 152, + 108, + 452, + 117.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 152, + 117.0, + 452, + 126.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 152, + 126.0, + 452, + 135.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 133, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 135, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 135, + 150 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 145, + 402, + 165 + ], + "lines": [ + { + "bbox": [ + 209, + 145, + 402, + 165 + ], + "spans": [ + { + "bbox": [ + 209, + 145, + 402, + 165 + ], + "score": 0.93, + "content": "\\psi = \\left( 1 - A \\left( \\mathbf { x } \\right) \\right) ^ { - 1 } \\operatorname* { m a x } _ { k } P _ { H C } \\left( C = k | \\mathbf { x } \\right) A \\left( \\mathbf { x } \\right)", + "type": "interline_equation", + "image_path": "ee123dba6e6e1954887949ef395ac365b193a08bcf15858e2f72a9f14fa6edc6.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 209, + 145, + 402, + 165 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 114, + 178 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "is the probability of the class according to HC weighted by the confidence that the associated", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "memory is actually stored in HC. BLA has the same number of layers/units as the mPFC encoder,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 478, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 478, + 200 + ], + "score": 1.0, + "content": "and uses a logistic output unit. We discuss alternative BLA models in supplemental material.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 215, + 244, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 245, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 245, + 230 + ], + "score": 1.0, + "content": "5 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "Evaluating Incremental Learning Performance. To evaluate how well the incrementally trained", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "models perform compared to an offline model, we use the three metrics proposed in Kemker et al.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 262, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 238, + 273 + ], + "score": 1.0, + "content": "(2018). After each study session", + "type": "text" + }, + { + "bbox": [ + 238, + 263, + 243, + 272 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 262, + 395, + 273 + ], + "score": 1.0, + "content": "in which a model learned a new class", + "type": "text" + }, + { + "bbox": [ + 395, + 262, + 402, + 272 + ], + "score": 0.77, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 262, + 505, + 273 + ], + "score": 1.0, + "content": ", we compute the model’s", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 237, + 286 + ], + "score": 1.0, + "content": "test accuracy on the new class", + "type": "text" + }, + { + "bbox": [ + 237, + 273, + 271, + 285 + ], + "score": 0.91, + "content": "( \\alpha _ { n e w , t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 273, + 431, + 286 + ], + "score": 1.0, + "content": ", the accuracy on the base-knowledge", + "type": "text" + }, + { + "bbox": [ + 432, + 273, + 466, + 285 + ], + "score": 0.91, + "content": "( \\alpha _ { b a s e , t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 273, + 505, + 286 + ], + "score": 1.0, + "content": ", and the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 305, + 297 + ], + "score": 1.0, + "content": "accuracy of all of the test data seen to this point", + "type": "text" + }, + { + "bbox": [ + 306, + 284, + 334, + 296 + ], + "score": 0.92, + "content": "( \\alpha _ { a l l , t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 284, + 377, + 297 + ], + "score": 1.0, + "content": ". After all", + "type": "text" + }, + { + "bbox": [ + 378, + 284, + 387, + 294 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "study sessions are complete,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 295, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 295, + 352, + 312 + ], + "score": 1.0, + "content": "a model’s ability to retain the base-knowledge is given by", + "type": "text" + }, + { + "bbox": [ + 352, + 295, + 472, + 312 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\Omega _ { b a s e } \\ = \\ \\frac { 1 } { T - 1 } \\sum _ { t = 2 } ^ { T } \\frac { \\alpha _ { b a s e , t } } { \\alpha _ { o f f l i n e } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 311, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 107, + 312, + 142, + 323 + ], + "score": 0.83, + "content": "\\alpha _ { o f f l i n e }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 311, + 506, + 323 + ], + "score": 1.0, + "content": "is the accuracy of a multi-layer perceptron (MLP) trained offline (i.e., it is given all of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "training data at once). The model’s ability to immediately recall new information is measured by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 139, + 332, + 507, + 349 + ], + "spans": [ + { + "bbox": [ + 139, + 332, + 507, + 349 + ], + "score": 1.0, + "content": "1T −1 PTt=2 αnew,t. Finally, we measure how well the model does on all available test data", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 347, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 238, + 367 + ], + "score": 1.0, + "content": "with Ωall = T −1 P t=2 αoffline", + "type": "text" + }, + { + "bbox": [ + 208, + 347, + 232, + 358 + ], + "score": 1.0, + "content": "αall,t", + "type": "text" + }, + { + "bbox": [ + 236, + 348, + 258, + 363 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 258, + 349, + 277, + 361 + ], + "score": 0.9, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 348, + 506, + 363 + ], + "score": 1.0, + "content": "metric shows how well new memories are integrated into", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 360, + 504, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 479, + 378 + ], + "score": 1.0, + "content": "the model over time. For all of the metrics, higher values indicate superior performance. Both", + "type": "text" + }, + { + "bbox": [ + 480, + 363, + 504, + 374 + ], + "score": 0.89, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 124, + 387 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 374, + 143, + 385 + ], + "score": 0.91, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "are relative to an offline MLP model, so a value of 1 indicates that a model has similar", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "performance to the offline baseline. This allows results across datasets to be better compared. Note", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 124, + 410 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 125, + 396, + 169, + 407 + ], + "score": 0.92, + "content": "\\Omega _ { b a s e } > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 394, + 188, + 410 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 188, + 396, + 227, + 407 + ], + "score": 0.91, + "content": "\\Omega _ { a l l } > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "only if the incremental learning algorithm is more accurate than the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "offline model, which can occur due to better regularization strategies employed by different models.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 303, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 304, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 304, + 435 + ], + "score": 1.0, + "content": "Datasets. We evaluate all of the models on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 434, + 304, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 304, + 446 + ], + "score": 1.0, + "content": "three benchmark datasets (Table 1): CIFAR-100,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 443, + 304, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 304, + 459 + ], + "score": 1.0, + "content": "CUB-200, and AudioSet. CIFAR-100 is a pop-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 456, + 304, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 304, + 468 + ], + "score": 1.0, + "content": "ular image classification dataset containing 100", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 468, + 304, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 304, + 479 + ], + "score": 1.0, + "content": "mutually-exclusive object categories, and it was", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 479, + 303, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 303, + 489 + ], + "score": 1.0, + "content": "used in Rebuffi et al. (2017) to evaluate iCaRL.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 489, + 304, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 171, + 501 + ], + "score": 1.0, + "content": "All images are", + "type": "text" + }, + { + "bbox": [ + 171, + 489, + 208, + 500 + ], + "score": 0.9, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 489, + 304, + 501 + ], + "score": 1.0, + "content": "pixels. CUB-200 is a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 500, + 304, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 304, + 512 + ], + "score": 1.0, + "content": "fine-grained image classification dataset contain-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 511, + 304, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 304, + 523 + ], + "score": 1.0, + "content": "ing high resolution images of 200 different bird", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 311, + 425, + 505, + 501 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 311, + 425, + 505, + 501 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 311, + 425, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 311, + 425, + 505, + 501 + ], + "score": 0.977, + "html": "
CIFAR-100CUB-200AudioSet
Classification TaskRGB ImageRGB ImageAudio
Classes100200100
Feature Shape2.0482,0481,280
Train Samples50,0005,99428,779
Test Samples10,0005,7945,523
Train Samples/Class50029-30250-300
Test Samples/Class10011-3043-62
", + "type": "table", + "image_path": "7b43a068859b87a51d7d3d60c90cf1c5df884c63275c18183b9e8bf8eed8109f.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 311, + 425, + 505, + 437.6666666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 311, + 437.6666666666667, + 505, + 450.33333333333337 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 311, + 450.33333333333337, + 505, + 463.00000000000006 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 311, + 463.00000000000006, + 505, + 475.66666666666674 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 311, + 475.66666666666674, + 505, + 488.3333333333334 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 311, + 488.3333333333334, + 505, + 501.0000000000001 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 345, + 510, + 470, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 344, + 510, + 471, + 523 + ], + "spans": [ + { + "bbox": [ + 344, + 510, + 471, + 523 + ], + "score": 1.0, + "content": "Table 1: Dataset Specifications", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + } + ], + "index": 38.75 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "species (Welinder et al., 2010). We use the 2011 version of the dataset. AudioSet is an audio clas-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "sification dataset (Gemmeke et al., 2017). We use the variant of AudioSet used by Kemker et al.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "(2018), which contains a 100 class subset such that none of the classes were super- or sub-classes", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "of one another. Also, since the AudioSet data samples can have more than one class, the chosen", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 566, + 355, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 355, + 578 + ], + "score": 1.0, + "content": "samples had only one of the 100 classes chosen in this subset.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "For CIFAR-100 and CUB-200, we extract ResNet-50 image embeddings as the input to each of the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "models, where ResNet-50 was pre-trained on ImageNet (He et al., 2016). We use the output after", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "the mean pooling layer and normalize the features to unit length. For AudioSet, we use the audio", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "CNN embeddings produced by pre-training the model on the YouTube-8M dataset (Abu-El-Haija", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "et al., 2016). We use the pre-extracted AudioSet feature embeddings, which represent ten second", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 639, + 384, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 384, + 650 + ], + "score": 1.0, + "content": "sound clips (i.e., ten 128-dimensional vectors concatenated in order).", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Comparison Models. We compare FearNet to FEL, GeppNet, GeppNet+STM, iCaRL, and an one-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "nearest neighbor (1-NN). FEL, GeppNet, and GeppNet+STM were chosen due to their previously", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "reported efficacy at incremental class learning in Kemker et al. (2018). iCARL is explicitly designed", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "for incremental class learning, and represents the state-of-the-art on this problem. We compare", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "against 1-NN due to its similarity to our HC model. 1-NN does not forget any previously observed", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "examples, but it tends to have worse generalization error than parametric methods and requires", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 721, + 229, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 229, + 733 + ], + "score": 1.0, + "content": "storing all of the training data.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 56 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 108, + 452, + 135 + ], + "lines": [ + { + "bbox": [ + 152, + 108, + 452, + 135 + ], + "spans": [ + { + "bbox": [ + 152, + 108, + 452, + 135 + ], + "score": 0.87, + "content": "\\hat { y } = \\left\\{ \\begin{array} { c c } { \\underset { \\mathrm { a r g } } { \\arg \\operatorname* { m a x } } _ { k ^ { \\prime } } P _ { H C } \\left( C = k ^ { \\prime } | \\mathbf { x } \\right) } & { \\mathrm { i f } \\ \\psi > \\operatorname* { m a x } _ { k } P _ { m P F C } \\left( C = k | \\mathbf { x } \\right) } \\\\ { \\underset { \\mathrm { o t h e r w i s e } } { \\arg \\operatorname* { m a x } } } & { \\mathrm { o t h e r w i s e } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "88d3befb10904759c6a11b9ae16fb1bec767d6cd04a2ec0924ee914bf9c57b6f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 152, + 108, + 452, + 117.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 152, + 117.0, + 452, + 126.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 152, + 126.0, + 452, + 135.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 133, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 135, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 135, + 150 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 136, + 135, + 150 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 145, + 402, + 165 + ], + "lines": [ + { + "bbox": [ + 209, + 145, + 402, + 165 + ], + "spans": [ + { + "bbox": [ + 209, + 145, + 402, + 165 + ], + "score": 0.93, + "content": "\\psi = \\left( 1 - A \\left( \\mathbf { x } \\right) \\right) ^ { - 1 } \\operatorname* { m a x } _ { k } P _ { H C } \\left( C = k | \\mathbf { x } \\right) A \\left( \\mathbf { x } \\right)", + "type": "interline_equation", + "image_path": "ee123dba6e6e1954887949ef395ac365b193a08bcf15858e2f72a9f14fa6edc6.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 209, + 145, + 402, + 165 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 114, + 178 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "is the probability of the class according to HC weighted by the confidence that the associated", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "memory is actually stored in HC. BLA has the same number of layers/units as the mPFC encoder,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 478, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 478, + 200 + ], + "score": 1.0, + "content": "and uses a logistic output unit. We discuss alternative BLA models in supplemental material.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 166, + 506, + 200 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 215, + 244, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 245, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 245, + 230 + ], + "score": 1.0, + "content": "5 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "Evaluating Incremental Learning Performance. To evaluate how well the incrementally trained", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "models perform compared to an offline model, we use the three metrics proposed in Kemker et al.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 262, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 238, + 273 + ], + "score": 1.0, + "content": "(2018). After each study session", + "type": "text" + }, + { + "bbox": [ + 238, + 263, + 243, + 272 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 262, + 395, + 273 + ], + "score": 1.0, + "content": "in which a model learned a new class", + "type": "text" + }, + { + "bbox": [ + 395, + 262, + 402, + 272 + ], + "score": 0.77, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 262, + 505, + 273 + ], + "score": 1.0, + "content": ", we compute the model’s", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 237, + 286 + ], + "score": 1.0, + "content": "test accuracy on the new class", + "type": "text" + }, + { + "bbox": [ + 237, + 273, + 271, + 285 + ], + "score": 0.91, + "content": "( \\alpha _ { n e w , t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 273, + 431, + 286 + ], + "score": 1.0, + "content": ", the accuracy on the base-knowledge", + "type": "text" + }, + { + "bbox": [ + 432, + 273, + 466, + 285 + ], + "score": 0.91, + "content": "( \\alpha _ { b a s e , t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 273, + 505, + 286 + ], + "score": 1.0, + "content": ", and the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 305, + 297 + ], + "score": 1.0, + "content": "accuracy of all of the test data seen to this point", + "type": "text" + }, + { + "bbox": [ + 306, + 284, + 334, + 296 + ], + "score": 0.92, + "content": "( \\alpha _ { a l l , t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 284, + 377, + 297 + ], + "score": 1.0, + "content": ". After all", + "type": "text" + }, + { + "bbox": [ + 378, + 284, + 387, + 294 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "study sessions are complete,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 295, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 295, + 352, + 312 + ], + "score": 1.0, + "content": "a model’s ability to retain the base-knowledge is given by", + "type": "text" + }, + { + "bbox": [ + 352, + 295, + 472, + 312 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\Omega _ { b a s e } \\ = \\ \\frac { 1 } { T - 1 } \\sum _ { t = 2 } ^ { T } \\frac { \\alpha _ { b a s e , t } } { \\alpha _ { o f f l i n e } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 311, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 107, + 312, + 142, + 323 + ], + "score": 0.83, + "content": "\\alpha _ { o f f l i n e }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 311, + 506, + 323 + ], + "score": 1.0, + "content": "is the accuracy of a multi-layer perceptron (MLP) trained offline (i.e., it is given all of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "training data at once). The model’s ability to immediately recall new information is measured by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 139, + 332, + 507, + 349 + ], + "spans": [ + { + "bbox": [ + 139, + 332, + 507, + 349 + ], + "score": 1.0, + "content": "1T −1 PTt=2 αnew,t. Finally, we measure how well the model does on all available test data", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 347, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 238, + 367 + ], + "score": 1.0, + "content": "with Ωall = T −1 P t=2 αoffline", + "type": "text" + }, + { + "bbox": [ + 208, + 347, + 232, + 358 + ], + "score": 1.0, + "content": "αall,t", + "type": "text" + }, + { + "bbox": [ + 236, + 348, + 258, + 363 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 258, + 349, + 277, + 361 + ], + "score": 0.9, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 348, + 506, + 363 + ], + "score": 1.0, + "content": "metric shows how well new memories are integrated into", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 360, + 504, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 479, + 378 + ], + "score": 1.0, + "content": "the model over time. For all of the metrics, higher values indicate superior performance. Both", + "type": "text" + }, + { + "bbox": [ + 480, + 363, + 504, + 374 + ], + "score": 0.89, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 124, + 387 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 374, + 143, + 385 + ], + "score": 0.91, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "are relative to an offline MLP model, so a value of 1 indicates that a model has similar", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "performance to the offline baseline. This allows results across datasets to be better compared. Note", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 124, + 410 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 125, + 396, + 169, + 407 + ], + "score": 0.92, + "content": "\\Omega _ { b a s e } > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 394, + 188, + 410 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 188, + 396, + 227, + 407 + ], + "score": 0.91, + "content": "\\Omega _ { a l l } > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "only if the incremental learning algorithm is more accurate than the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "offline model, which can occur due to better regularization strategies employed by different models.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 240, + 507, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 303, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 304, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 304, + 435 + ], + "score": 1.0, + "content": "Datasets. We evaluate all of the models on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 434, + 304, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 304, + 446 + ], + "score": 1.0, + "content": "three benchmark datasets (Table 1): CIFAR-100,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 443, + 304, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 304, + 459 + ], + "score": 1.0, + "content": "CUB-200, and AudioSet. CIFAR-100 is a pop-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 456, + 304, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 304, + 468 + ], + "score": 1.0, + "content": "ular image classification dataset containing 100", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 468, + 304, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 304, + 479 + ], + "score": 1.0, + "content": "mutually-exclusive object categories, and it was", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 479, + 303, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 303, + 489 + ], + "score": 1.0, + "content": "used in Rebuffi et al. (2017) to evaluate iCaRL.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 489, + 304, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 171, + 501 + ], + "score": 1.0, + "content": "All images are", + "type": "text" + }, + { + "bbox": [ + 171, + 489, + 208, + 500 + ], + "score": 0.9, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 489, + 304, + 501 + ], + "score": 1.0, + "content": "pixels. CUB-200 is a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 500, + 304, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 304, + 512 + ], + "score": 1.0, + "content": "fine-grained image classification dataset contain-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 511, + 304, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 304, + 523 + ], + "score": 1.0, + "content": "ing high resolution images of 200 different bird", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 423, + 304, + 523 + ] + }, + { + "type": "table", + "bbox": [ + 311, + 425, + 505, + 501 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 311, + 425, + 505, + 501 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 311, + 425, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 311, + 425, + 505, + 501 + ], + "score": 0.977, + "html": "
CIFAR-100CUB-200AudioSet
Classification TaskRGB ImageRGB ImageAudio
Classes100200100
Feature Shape2.0482,0481,280
Train Samples50,0005,99428,779
Test Samples10,0005,7945,523
Train Samples/Class50029-30250-300
Test Samples/Class10011-3043-62
", + "type": "table", + "image_path": "7b43a068859b87a51d7d3d60c90cf1c5df884c63275c18183b9e8bf8eed8109f.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 311, + 425, + 505, + 437.6666666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 311, + 437.6666666666667, + 505, + 450.33333333333337 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 311, + 450.33333333333337, + 505, + 463.00000000000006 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 311, + 463.00000000000006, + 505, + 475.66666666666674 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 311, + 475.66666666666674, + 505, + 488.3333333333334 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 311, + 488.3333333333334, + 505, + 501.0000000000001 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 345, + 510, + 470, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 344, + 510, + 471, + 523 + ], + "spans": [ + { + "bbox": [ + 344, + 510, + 471, + 523 + ], + "score": 1.0, + "content": "Table 1: Dataset Specifications", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + } + ], + "index": 38.75 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "species (Welinder et al., 2010). We use the 2011 version of the dataset. AudioSet is an audio clas-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "sification dataset (Gemmeke et al., 2017). We use the variant of AudioSet used by Kemker et al.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "(2018), which contains a 100 class subset such that none of the classes were super- or sub-classes", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "of one another. Also, since the AudioSet data samples can have more than one class, the chosen", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 566, + 355, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 355, + 578 + ], + "score": 1.0, + "content": "samples had only one of the 100 classes chosen in this subset.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 522, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "For CIFAR-100 and CUB-200, we extract ResNet-50 image embeddings as the input to each of the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "models, where ResNet-50 was pre-trained on ImageNet (He et al., 2016). We use the output after", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "the mean pooling layer and normalize the features to unit length. For AudioSet, we use the audio", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "CNN embeddings produced by pre-training the model on the YouTube-8M dataset (Abu-El-Haija", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "et al., 2016). We use the pre-extracted AudioSet feature embeddings, which represent ten second", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 639, + 384, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 384, + 650 + ], + "score": 1.0, + "content": "sound clips (i.e., ten 128-dimensional vectors concatenated in order).", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 582, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Comparison Models. We compare FearNet to FEL, GeppNet, GeppNet+STM, iCaRL, and an one-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "nearest neighbor (1-NN). FEL, GeppNet, and GeppNet+STM were chosen due to their previously", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "reported efficacy at incremental class learning in Kemker et al. (2018). iCARL is explicitly designed", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "for incremental class learning, and represents the state-of-the-art on this problem. We compare", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "against 1-NN due to its similarity to our HC model. 1-NN does not forget any previously observed", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "examples, but it tends to have worse generalization error than parametric methods and requires", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 721, + 229, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 229, + 733 + ], + "score": 1.0, + "content": "storing all of the training data.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 56, + "bbox_fs": [ + 104, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 85, + 500, + 199 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 85, + 500, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 85, + 500, + 199 + ], + "spans": [ + { + "bbox": [ + 107, + 85, + 500, + 199 + ], + "score": 0.971, + "type": "image", + "image_path": "8fe87b21db3278cee40742062caccd69283383f3df858d8c12bed9bd6756eed5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 85, + 500, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 123.0, + 500, + 161.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 161.0, + 500, + 199.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 185, + 211, + 424, + 223 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 210, + 427, + 225 + ], + "spans": [ + { + "bbox": [ + 184, + 210, + 427, + 225 + ], + "score": 1.0, + "content": "Figure 4: Mean-class test accuracy of all classes seen so far.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 505, + 258 + ], + "score": 1.0, + "content": "In each of our experiments, all models take the same feature embedding as input for a given dataset.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "score": 1.0, + "content": "This required modifying iCaRL by turning its CNN into a fully connected network. We performed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "a hyperparameter search for each model/dataset combination to tune the number of units and layers", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 280, + 228, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 228, + 292 + ], + "score": 1.0, + "content": "(see Supplemental Materials).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "Training Parameters. FearNet was implemented in Tensorflow. For mPFC and BLA, each fully", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "connected layer uses an exponential linear unit activation function (Clevert et al., 2016). The output", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "of the encoder also connects to a softmax output layer. Xavier initialization is used to initialize all", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "weight layers (Glorot & Bengio, 2010), and all of the biases are initialized to one. BLA’s architecture", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 340, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 480, + 353 + ], + "score": 1.0, + "content": "is identical to mPFC’s encoder, except it has a logistic output unit, instead of a softmax layer.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "mPFC and BLA were trained using NAdam. We train mPFC on the base-knowledge set for 1,000", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "epochs, consolidate HC over to mPFC for 60 epochs, and train BLA for 20 epochs. Because mPFC’s", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 342, + 392 + ], + "score": 1.0, + "content": "decoder is vital to preserving memories, its learning rate is", + "type": "text" + }, + { + "bbox": [ + 342, + 379, + 369, + 391 + ], + "score": 0.76, + "content": "1 / 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "times lower than the encoder. We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "performed a hyperparameter search for each dataset and model, varying the model shape (64-1,024", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "units), depth (2-4 layers), and how often to sleep (see Sec. 6.2). Across datasets, mPFC and BLA", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "performed best with two hidden layers, but the number of units per layer varied across datasets. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "specific values used for each dataset are given in supplemental material. In preliminary experiments,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "we found no benefit to adding weight decay to mPFC, likely because the reconstruction task helps", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 446, + 193, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 193, + 457 + ], + "score": 1.0, + "content": "regularize the model.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 476, + 256, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 258, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 258, + 491 + ], + "score": 1.0, + "content": "6 EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 504, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "Unless otherwise noted, each class is only seen in one unique study-session and the first base-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "knowledge study session contains half the classes in the dataset. We perform additional experiments", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "to study how changing the number of base-knowledge classes affects performance in Sec. 6.2. Un-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 536, + 413, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 413, + 549 + ], + "score": 1.0, + "content": "less otherwise noted, FearNet sleeps every 10 study sessions across datasets.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 108, + 565, + 276, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 278, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 278, + 578 + ], + "score": 1.0, + "content": "6.1 STATE-OF-THE-ART COMPARISON", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "Table 2 shows incremental class learning summary results for all six methods. FearNet achieves", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 140, + 612 + ], + "score": 1.0, + "content": "the best", + "type": "text" + }, + { + "bbox": [ + 140, + 599, + 164, + 609 + ], + "score": 0.9, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 597, + 182, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 183, + 599, + 201, + 609 + ], + "score": 0.9, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "on all three datasets. Fig. 4 shows that FearNet more closely resembles the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 608, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 273, + 623 + ], + "score": 1.0, + "content": "offline MLP baseline than other methods.", + "type": "text" + }, + { + "bbox": [ + 274, + 610, + 297, + 621 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 608, + 506, + 623 + ], + "score": 1.0, + "content": "measures test accuracy on the most recently trained", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 393, + 633 + ], + "score": 1.0, + "content": "class. 1 For FearNet, this measures the performance of HC and BLA.", + "type": "text" + }, + { + "bbox": [ + 394, + 621, + 417, + 632 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "does not account for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "score": 1.0, + "content": "how well the class was consolidated into mPFC which happens later during a sleep phase; however,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 641, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 125, + 653 + ], + "score": 0.9, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 641, + 308, + 656 + ], + "score": 1.0, + "content": "does account for this. FEL achieves a high", + "type": "text" + }, + { + "bbox": [ + 309, + 642, + 332, + 653 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 641, + 506, + 656 + ], + "score": 1.0, + "content": "score because it is able to achieve nearly", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 654, + 504, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 504, + 665 + ], + "score": 1.0, + "content": "perfect test accuracy on every new class it learns, but this results in forgetting more quickly than", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "FearNet. 1-NN is similar to our HC model; but on its own, it fails to generalize as well as FearNet,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 687 + ], + "score": 1.0, + "content": "is memory inefficient, and is slow to make predictions. The final mean-class test accuracy for the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 685, + 504, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 282, + 698 + ], + "score": 1.0, + "content": "offline MLP used to normalize the metrics is", + "type": "text" + }, + { + "bbox": [ + 283, + 686, + 310, + 696 + ], + "score": 0.9, + "content": "6 9 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 685, + 375, + 698 + ], + "score": 1.0, + "content": "for CIFAR-100,", + "type": "text" + }, + { + "bbox": [ + 376, + 686, + 403, + 696 + ], + "score": 0.79, + "content": "5 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 685, + 477, + 698 + ], + "score": 1.0, + "content": "for CUB-200, and", + "type": "text" + }, + { + "bbox": [ + 477, + 686, + 504, + 696 + ], + "score": 0.84, + "content": "4 5 . 8 \\%", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 697, + 163, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 163, + 707 + ], + "score": 1.0, + "content": "for AudioSet.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 117, + 721, + 437, + 732 + ], + "lines": [ + { + "bbox": [ + 122, + 719, + 435, + 734 + ], + "spans": [ + { + "bbox": [ + 122, + 721, + 145, + 731 + ], + "score": 0.82, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 719, + 210, + 734 + ], + "score": 1.0, + "content": "is not scaled with", + "type": "text" + }, + { + "bbox": [ + 210, + 723, + 243, + 732 + ], + "score": 0.81, + "content": "\\alpha _ { o f f l i n e }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 719, + 379, + 734 + ], + "score": 1.0, + "content": ", so it does not have the same scale as", + "type": "text" + }, + { + "bbox": [ + 379, + 722, + 402, + 731 + ], + "score": 0.91, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 719, + 418, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 418, + 721, + 435, + 731 + ], + "score": 0.89, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 85, + 500, + 199 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 85, + 500, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 85, + 500, + 199 + ], + "spans": [ + { + "bbox": [ + 107, + 85, + 500, + 199 + ], + "score": 0.971, + "type": "image", + "image_path": "8fe87b21db3278cee40742062caccd69283383f3df858d8c12bed9bd6756eed5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 85, + 500, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 123.0, + 500, + 161.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 161.0, + 500, + 199.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 185, + 211, + 424, + 223 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 210, + 427, + 225 + ], + "spans": [ + { + "bbox": [ + 184, + 210, + 427, + 225 + ], + "score": 1.0, + "content": "Figure 4: Mean-class test accuracy of all classes seen so far.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 505, + 258 + ], + "score": 1.0, + "content": "In each of our experiments, all models take the same feature embedding as input for a given dataset.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "score": 1.0, + "content": "This required modifying iCaRL by turning its CNN into a fully connected network. We performed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "a hyperparameter search for each model/dataset combination to tune the number of units and layers", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 280, + 228, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 228, + 292 + ], + "score": 1.0, + "content": "(see Supplemental Materials).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 247, + 506, + 292 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "Training Parameters. FearNet was implemented in Tensorflow. For mPFC and BLA, each fully", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "connected layer uses an exponential linear unit activation function (Clevert et al., 2016). The output", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "of the encoder also connects to a softmax output layer. Xavier initialization is used to initialize all", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "weight layers (Glorot & Bengio, 2010), and all of the biases are initialized to one. BLA’s architecture", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 340, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 480, + 353 + ], + "score": 1.0, + "content": "is identical to mPFC’s encoder, except it has a logistic output unit, instead of a softmax layer.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 297, + 505, + 353 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "mPFC and BLA were trained using NAdam. We train mPFC on the base-knowledge set for 1,000", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "epochs, consolidate HC over to mPFC for 60 epochs, and train BLA for 20 epochs. Because mPFC’s", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 342, + 392 + ], + "score": 1.0, + "content": "decoder is vital to preserving memories, its learning rate is", + "type": "text" + }, + { + "bbox": [ + 342, + 379, + 369, + 391 + ], + "score": 0.76, + "content": "1 / 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "times lower than the encoder. We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "performed a hyperparameter search for each dataset and model, varying the model shape (64-1,024", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "units), depth (2-4 layers), and how often to sleep (see Sec. 6.2). Across datasets, mPFC and BLA", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "performed best with two hidden layers, but the number of units per layer varied across datasets. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "specific values used for each dataset are given in supplemental material. In preliminary experiments,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "we found no benefit to adding weight decay to mPFC, likely because the reconstruction task helps", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 446, + 193, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 193, + 457 + ], + "score": 1.0, + "content": "regularize the model.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 358, + 506, + 457 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 476, + 256, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 258, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 258, + 491 + ], + "score": 1.0, + "content": "6 EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 504, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "Unless otherwise noted, each class is only seen in one unique study-session and the first base-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "knowledge study session contains half the classes in the dataset. We perform additional experiments", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "to study how changing the number of base-knowledge classes affects performance in Sec. 6.2. Un-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 536, + 413, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 413, + 549 + ], + "score": 1.0, + "content": "less otherwise noted, FearNet sleeps every 10 study sessions across datasets.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 503, + 506, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 565, + 276, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 278, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 278, + 578 + ], + "score": 1.0, + "content": "6.1 STATE-OF-THE-ART COMPARISON", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "Table 2 shows incremental class learning summary results for all six methods. FearNet achieves", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 140, + 612 + ], + "score": 1.0, + "content": "the best", + "type": "text" + }, + { + "bbox": [ + 140, + 599, + 164, + 609 + ], + "score": 0.9, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 597, + 182, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 183, + 599, + 201, + 609 + ], + "score": 0.9, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "on all three datasets. Fig. 4 shows that FearNet more closely resembles the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 608, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 273, + 623 + ], + "score": 1.0, + "content": "offline MLP baseline than other methods.", + "type": "text" + }, + { + "bbox": [ + 274, + 610, + 297, + 621 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 608, + 506, + 623 + ], + "score": 1.0, + "content": "measures test accuracy on the most recently trained", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 393, + 633 + ], + "score": 1.0, + "content": "class. 1 For FearNet, this measures the performance of HC and BLA.", + "type": "text" + }, + { + "bbox": [ + 394, + 621, + 417, + 632 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "does not account for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "score": 1.0, + "content": "how well the class was consolidated into mPFC which happens later during a sleep phase; however,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 641, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 125, + 653 + ], + "score": 0.9, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 641, + 308, + 656 + ], + "score": 1.0, + "content": "does account for this. FEL achieves a high", + "type": "text" + }, + { + "bbox": [ + 309, + 642, + 332, + 653 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 641, + 506, + 656 + ], + "score": 1.0, + "content": "score because it is able to achieve nearly", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 654, + 504, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 504, + 665 + ], + "score": 1.0, + "content": "perfect test accuracy on every new class it learns, but this results in forgetting more quickly than", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "FearNet. 1-NN is similar to our HC model; but on its own, it fails to generalize as well as FearNet,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 687 + ], + "score": 1.0, + "content": "is memory inefficient, and is slow to make predictions. The final mean-class test accuracy for the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 685, + 504, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 282, + 698 + ], + "score": 1.0, + "content": "offline MLP used to normalize the metrics is", + "type": "text" + }, + { + "bbox": [ + 283, + 686, + 310, + 696 + ], + "score": 0.9, + "content": "6 9 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 685, + 375, + 698 + ], + "score": 1.0, + "content": "for CIFAR-100,", + "type": "text" + }, + { + "bbox": [ + 376, + 686, + 403, + 696 + ], + "score": 0.79, + "content": "5 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 685, + 477, + 698 + ], + "score": 1.0, + "content": "for CUB-200, and", + "type": "text" + }, + { + "bbox": [ + 477, + 686, + 504, + 696 + ], + "score": 0.84, + "content": "4 5 . 8 \\%", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 697, + 163, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 163, + 707 + ], + "score": 1.0, + "content": "for AudioSet.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 587, + 506, + 707 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 80, + 479, + 173 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 132, + 80, + 479, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 80, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 80, + 479, + 173 + ], + "score": 0.982, + "html": "
ModelCIFAR-100 Sbase Ωnew allCUB-200 SnewAudioSet ΩnewMean Sbase
1-Nearest Neighbor||0.8780.648Sbase |0.746Ωaul 0.434Sbase 10.6550.269Sall 0.613Sall
GeppNet+STM0.8660.879 0.408 0.8000.7640.694 0.204 0.6450.9410.861[0.760 0.8570.729 0.769
GeppNet0.8330.529 0.7540.7270.558 0.6450.9320.3720.8310.759
FEL0.7070.999 0.6190.7020.976 0.6410.4910.499 1.000 0.4560.879
iCaRL0.7460.807 0.7490.942 0.5470.8640.740 0.4870.7330.633 0.8010.572 0.782
FearNet0.9270.824 0.9470.9240.598 0.8910.9620.455 0.9320.9380.923
", + "type": "table", + "image_path": "5077ab9e33df318c7465f84dfb058d56ebfd4e8d5ff8b148c9284ab3455b4f29.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 132, + 80, + 479, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 132, + 111.0, + 479, + 142.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 132, + 142.0, + 479, + 173.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 181, + 505, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 470, + 193 + ], + "score": 1.0, + "content": "Table 2: State-of-the-art comparison on CIFAR-100, CUB-200, and AudioSet. The best", + "type": "text" + }, + { + "bbox": [ + 470, + 181, + 489, + 192 + ], + "score": 0.89, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 190, + 448, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 207, + 205 + ], + "score": 1.0, + "content": "each dataset are in bold.", + "type": "text" + }, + { + "bbox": [ + 207, + 192, + 231, + 204 + ], + "score": 0.91, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 190, + 249, + 205 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 192, + 267, + 203 + ], + "score": 0.91, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 190, + 448, + 205 + ], + "score": 1.0, + "content": "are normalized by the offline MLP baseline.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 158, + 221, + 453, + 284 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 158, + 221, + 453, + 284 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 158, + 221, + 453, + 284 + ], + "spans": [ + { + "bbox": [ + 158, + 221, + 453, + 284 + ], + "score": 0.977, + "html": "
CIFAR-100CUB-200AudioSet
OracleWith BLAOracleWith BLAOracleWith BLA
Sbase0.9650.9270.9680.9240.9700.962
Snew0.9120.8240.7290.5980.7010.455
Ωau1.0020.9470.9360.8910.9720.932
", + "type": "table", + "image_path": "4e1f9a40e0b8823b06ee244e2bd219f64045a9e9ab6f5474709a28c15f84182b.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 158, + 221, + 453, + 242.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 158, + 242.0, + 453, + 263.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 158, + 263.0, + 453, + 284.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 291, + 505, + 314 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 290, + 504, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 504, + 304 + ], + "score": 1.0, + "content": "Table 3: FearNet performance when the location of the associated memory is known using an oracle", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 302, + 184, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 184, + 315 + ], + "score": 1.0, + "content": "versus using BLA.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 108, + 334, + 252, + 346 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 253, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 253, + 347 + ], + "score": 1.0, + "content": "6.2 ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "Novelty Detection with BLA. We evaluated the performance of BLA by comparing it to an oracle", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "version of FearNet, i.e., a version that knew if the relevant memory was stored in either mPFC or", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "HC. Table 3 shows that FearNet’s BLA does a good job at predicting which network to use; however,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 390, + 474, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 168, + 403 + ], + "score": 1.0, + "content": "the decrease in", + "type": "text" + }, + { + "bbox": [ + 168, + 391, + 192, + 402 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 390, + 474, + 403 + ], + "score": 1.0, + "content": "suggests BLA is sometimes using mPFC when it should be using HC.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 364, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 365, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 365, + 421 + ], + "score": 1.0, + "content": "When should the model sleep? To study how the frequency", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 418, + 365, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 365, + 431 + ], + "score": 1.0, + "content": "of memory consolidation affects FearNet’s performance, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 428, + 364, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 364, + 443 + ], + "score": 1.0, + "content": "trained FearNet on CUB-200 and varied the sleep frequency", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 440, + 365, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 365, + 452 + ], + "score": 1.0, + "content": "from 1-15 study sessions. When FearNet increases the number", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 451, + 364, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 364, + 463 + ], + "score": 1.0, + "content": "of classes it learns before sleeping (Fig. 5), it is better able to re-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 462, + 364, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 364, + 474 + ], + "score": 1.0, + "content": "tain its base-knowledge, but this reduces its ability to recall new", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "score": 1.0, + "content": "information. In humans, sleep deprivation is known to impair", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 483, + 365, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 365, + 497 + ], + "score": 1.0, + "content": "new learning (Yoo et al., 2007), and that forgetting occurs during", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 495, + 365, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 365, + 508 + ], + "score": 1.0, + "content": "sleep (Poe, 2017). Each time FearNet sleeps, the mPFC weights", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 364, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 364, + 518 + ], + "score": 1.0, + "content": "are perturbed which can cause it to gradually forget older memo-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 517, + 365, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 365, + 529 + ], + "score": 1.0, + "content": "ries. Sleeping less causes HC’s recall performance to deteriorate.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 375, + 411, + 500, + 504 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 375, + 411, + 500, + 504 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 375, + 411, + 500, + 504 + ], + "spans": [ + { + "bbox": [ + 375, + 411, + 500, + 504 + ], + "score": 0.966, + "type": "image", + "image_path": "2f78923b9aae75d6dcc03f23fbd92ef51d126b3d916bf7c065180de4a5bced14.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 375, + 411, + 500, + 457.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 375, + 457.5, + 500, + 504.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 371, + 515, + 504, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 372, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 372, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "Figure 5: FearNet performance", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 372, + 527, + 504, + 538 + ], + "spans": [ + { + "bbox": [ + 372, + 527, + 504, + 538 + ], + "score": 1.0, + "content": "as the sleep frequency decreases.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 339, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 339, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 339, + 556 + ], + "score": 1.0, + "content": "Multi-Modal Incremental Learning. As shown in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 556, + 339, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 339, + 568 + ], + "score": 1.0, + "content": "Sec. 6.1, FearNet can incrementally learn and retain in-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 567, + 339, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 339, + 579 + ], + "score": 1.0, + "content": "formation from a single dataset, but how does it perform", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 577, + 339, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 339, + 590 + ], + "score": 1.0, + "content": "if new inputs differ greatly from previously learned ones?", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 588, + 339, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 339, + 600 + ], + "score": 1.0, + "content": "This scenario is one of the first shown to cause catas-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 600, + 339, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 339, + 612 + ], + "score": 1.0, + "content": "trophic forgetting in MLPs. To study this, we trained", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 610, + 339, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 339, + 623 + ], + "score": 1.0, + "content": "FearNet to incrementally learn CIFAR-100 and AudioSet,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 621, + 339, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 339, + 633 + ], + "score": 1.0, + "content": "which after training is a 200-way classification problem.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 632, + 339, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 339, + 644 + ], + "score": 1.0, + "content": "To do this, AudioSet’s features are zero-padded to make", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 643, + 339, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 339, + 655 + ], + "score": 1.0, + "content": "them the same length as CIFAR-100s. Table 4 shows", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 654, + 340, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 340, + 668 + ], + "score": 1.0, + "content": "the performance of FearNet for three separate training", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "score": 1.0, + "content": "paradigms: 1) FearNet learns CIFAR-100 as the base-", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35.5 + }, + { + "type": "table", + "bbox": [ + 349, + 547, + 502, + 609 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 349, + 547, + 502, + 609 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 349, + 547, + 502, + 609 + ], + "spans": [ + { + "bbox": [ + 349, + 547, + 502, + 609 + ], + "score": 0.974, + "html": "
Base-Knowledge CIFAR-100 AudioSet 50/50 Mix
Sbase|0.995 0.8450.837
Snew0.693 0.9030.822
Saul0.854 0.6340.820
", + "type": "table", + "image_path": "4613cabd06ec069008010343d0d65e0d6b4aafe7f22fc7b00b3698023edb4d4e.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 349, + 547, + 502, + 578.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 349, + 578.0, + 502, + 609.0 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 347, + 617, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 345, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 345, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "Table 4: Multi-modal incremental", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 345, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 345, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "learning experiment. FearNet was", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 346, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 346, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "trained with various base-knowledge", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 345, + 651, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 345, + 651, + 505, + 662 + ], + "score": 1.0, + "content": "sets (column-header) and then incre-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 345, + 662, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 345, + 662, + 504, + 673 + ], + "score": 1.0, + "content": "mentally trained on all remaining data.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "knowledge and then incrementally learns AudioSet; 2) FearNet learns AudioSet as the base-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "knowledge and then incrementally learns CIFAR-100; and 3) the base-knowledge contains a 50/50", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "split from both datasets with FearNet incrementally learning the remaining classes. Our results", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "suggest FearNet is capable of incrementally learning multi-modal information, if the model has a", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "good starting point (high base-knowledge); however, if the model starts with lower base-knowledge", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 80, + 479, + 173 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 132, + 80, + 479, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 80, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 80, + 479, + 173 + ], + "score": 0.982, + "html": "
ModelCIFAR-100 Sbase Ωnew allCUB-200 SnewAudioSet ΩnewMean Sbase
1-Nearest Neighbor||0.8780.648Sbase |0.746Ωaul 0.434Sbase 10.6550.269Sall 0.613Sall
GeppNet+STM0.8660.879 0.408 0.8000.7640.694 0.204 0.6450.9410.861[0.760 0.8570.729 0.769
GeppNet0.8330.529 0.7540.7270.558 0.6450.9320.3720.8310.759
FEL0.7070.999 0.6190.7020.976 0.6410.4910.499 1.000 0.4560.879
iCaRL0.7460.807 0.7490.942 0.5470.8640.740 0.4870.7330.633 0.8010.572 0.782
FearNet0.9270.824 0.9470.9240.598 0.8910.9620.455 0.9320.9380.923
", + "type": "table", + "image_path": "5077ab9e33df318c7465f84dfb058d56ebfd4e8d5ff8b148c9284ab3455b4f29.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 132, + 80, + 479, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 132, + 111.0, + 479, + 142.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 132, + 142.0, + 479, + 173.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 181, + 505, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 470, + 193 + ], + "score": 1.0, + "content": "Table 2: State-of-the-art comparison on CIFAR-100, CUB-200, and AudioSet. The best", + "type": "text" + }, + { + "bbox": [ + 470, + 181, + 489, + 192 + ], + "score": 0.89, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 190, + 448, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 207, + 205 + ], + "score": 1.0, + "content": "each dataset are in bold.", + "type": "text" + }, + { + "bbox": [ + 207, + 192, + 231, + 204 + ], + "score": 0.91, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 190, + 249, + 205 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 192, + 267, + 203 + ], + "score": 0.91, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 190, + 448, + 205 + ], + "score": 1.0, + "content": "are normalized by the offline MLP baseline.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 158, + 221, + 453, + 284 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 158, + 221, + 453, + 284 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 158, + 221, + 453, + 284 + ], + "spans": [ + { + "bbox": [ + 158, + 221, + 453, + 284 + ], + "score": 0.977, + "html": "
CIFAR-100CUB-200AudioSet
OracleWith BLAOracleWith BLAOracleWith BLA
Sbase0.9650.9270.9680.9240.9700.962
Snew0.9120.8240.7290.5980.7010.455
Ωau1.0020.9470.9360.8910.9720.932
", + "type": "table", + "image_path": "4e1f9a40e0b8823b06ee244e2bd219f64045a9e9ab6f5474709a28c15f84182b.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 158, + 221, + 453, + 242.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 158, + 242.0, + 453, + 263.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 158, + 263.0, + 453, + 284.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 291, + 505, + 314 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 290, + 504, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 504, + 304 + ], + "score": 1.0, + "content": "Table 3: FearNet performance when the location of the associated memory is known using an oracle", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 302, + 184, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 184, + 315 + ], + "score": 1.0, + "content": "versus using BLA.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 108, + 334, + 252, + 346 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 253, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 253, + 347 + ], + "score": 1.0, + "content": "6.2 ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "Novelty Detection with BLA. We evaluated the performance of BLA by comparing it to an oracle", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "version of FearNet, i.e., a version that knew if the relevant memory was stored in either mPFC or", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "HC. Table 3 shows that FearNet’s BLA does a good job at predicting which network to use; however,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 390, + 474, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 168, + 403 + ], + "score": 1.0, + "content": "the decrease in", + "type": "text" + }, + { + "bbox": [ + 168, + 391, + 192, + 402 + ], + "score": 0.91, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 390, + 474, + 403 + ], + "score": 1.0, + "content": "suggests BLA is sometimes using mPFC when it should be using HC.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 357, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 364, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 365, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 365, + 421 + ], + "score": 1.0, + "content": "When should the model sleep? To study how the frequency", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 418, + 365, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 365, + 431 + ], + "score": 1.0, + "content": "of memory consolidation affects FearNet’s performance, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 428, + 364, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 364, + 443 + ], + "score": 1.0, + "content": "trained FearNet on CUB-200 and varied the sleep frequency", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 440, + 365, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 365, + 452 + ], + "score": 1.0, + "content": "from 1-15 study sessions. When FearNet increases the number", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 451, + 364, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 364, + 463 + ], + "score": 1.0, + "content": "of classes it learns before sleeping (Fig. 5), it is better able to re-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 462, + 364, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 364, + 474 + ], + "score": 1.0, + "content": "tain its base-knowledge, but this reduces its ability to recall new", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "score": 1.0, + "content": "information. In humans, sleep deprivation is known to impair", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 483, + 365, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 365, + 497 + ], + "score": 1.0, + "content": "new learning (Yoo et al., 2007), and that forgetting occurs during", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 495, + 365, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 365, + 508 + ], + "score": 1.0, + "content": "sleep (Poe, 2017). Each time FearNet sleeps, the mPFC weights", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 364, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 364, + 518 + ], + "score": 1.0, + "content": "are perturbed which can cause it to gradually forget older memo-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 517, + 365, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 365, + 529 + ], + "score": 1.0, + "content": "ries. Sleeping less causes HC’s recall performance to deteriorate.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 406, + 365, + 529 + ] + }, + { + "type": "image", + "bbox": [ + 375, + 411, + 500, + 504 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 375, + 411, + 500, + 504 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 375, + 411, + 500, + 504 + ], + "spans": [ + { + "bbox": [ + 375, + 411, + 500, + 504 + ], + "score": 0.966, + "type": "image", + "image_path": "2f78923b9aae75d6dcc03f23fbd92ef51d126b3d916bf7c065180de4a5bced14.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 375, + 411, + 500, + 457.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 375, + 457.5, + 500, + 504.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 371, + 515, + 504, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 372, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 372, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "Figure 5: FearNet performance", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 372, + 527, + 504, + 538 + ], + "spans": [ + { + "bbox": [ + 372, + 527, + 504, + 538 + ], + "score": 1.0, + "content": "as the sleep frequency decreases.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 339, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 339, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 339, + 556 + ], + "score": 1.0, + "content": "Multi-Modal Incremental Learning. As shown in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 556, + 339, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 339, + 568 + ], + "score": 1.0, + "content": "Sec. 6.1, FearNet can incrementally learn and retain in-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 567, + 339, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 339, + 579 + ], + "score": 1.0, + "content": "formation from a single dataset, but how does it perform", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 577, + 339, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 339, + 590 + ], + "score": 1.0, + "content": "if new inputs differ greatly from previously learned ones?", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 588, + 339, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 339, + 600 + ], + "score": 1.0, + "content": "This scenario is one of the first shown to cause catas-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 600, + 339, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 339, + 612 + ], + "score": 1.0, + "content": "trophic forgetting in MLPs. To study this, we trained", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 610, + 339, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 339, + 623 + ], + "score": 1.0, + "content": "FearNet to incrementally learn CIFAR-100 and AudioSet,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 621, + 339, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 339, + 633 + ], + "score": 1.0, + "content": "which after training is a 200-way classification problem.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 632, + 339, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 339, + 644 + ], + "score": 1.0, + "content": "To do this, AudioSet’s features are zero-padded to make", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 643, + 339, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 339, + 655 + ], + "score": 1.0, + "content": "them the same length as CIFAR-100s. Table 4 shows", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 654, + 340, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 340, + 668 + ], + "score": 1.0, + "content": "the performance of FearNet for three separate training", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "score": 1.0, + "content": "paradigms: 1) FearNet learns CIFAR-100 as the base-", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 545, + 340, + 677 + ] + }, + { + "type": "table", + "bbox": [ + 349, + 547, + 502, + 609 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 349, + 547, + 502, + 609 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 349, + 547, + 502, + 609 + ], + "spans": [ + { + "bbox": [ + 349, + 547, + 502, + 609 + ], + "score": 0.974, + "html": "
Base-Knowledge CIFAR-100 AudioSet 50/50 Mix
Sbase|0.995 0.8450.837
Snew0.693 0.9030.822
Saul0.854 0.6340.820
", + "type": "table", + "image_path": "4613cabd06ec069008010343d0d65e0d6b4aafe7f22fc7b00b3698023edb4d4e.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 349, + 547, + 502, + 578.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 349, + 578.0, + 502, + 609.0 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 347, + 617, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 345, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 345, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "Table 4: Multi-modal incremental", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 345, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 345, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "learning experiment. FearNet was", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 346, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 346, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "trained with various base-knowledge", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 345, + 651, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 345, + 651, + 505, + 662 + ], + "score": 1.0, + "content": "sets (column-header) and then incre-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 345, + 662, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 345, + 662, + 504, + 673 + ], + "score": 1.0, + "content": "mentally trained on all remaining data.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46, + "bbox_fs": [ + 345, + 616, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "knowledge and then incrementally learns AudioSet; 2) FearNet learns AudioSet as the base-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "knowledge and then incrementally learns CIFAR-100; and 3) the base-knowledge contains a 50/50", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "split from both datasets with FearNet incrementally learning the remaining classes. Our results", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "suggest FearNet is capable of incrementally learning multi-modal information, if the model has a", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "good starting point (high base-knowledge); however, if the model starts with lower base-knowledge", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "score": 1.0, + "content": "performance (e.g., AudioSet), the model struggles to learn new information incrementally (see Sup-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 260, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 260, + 106 + ], + "score": 1.0, + "content": "plemental Material for detailed plots).", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 51, + "bbox_fs": [ + 105, + 677, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "score": 1.0, + "content": "performance (e.g., AudioSet), the model struggles to learn new information incrementally (see Sup-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 260, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 260, + 106 + ], + "score": 1.0, + "content": "plemental Material for detailed plots).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "Base-Knowledge Effect on Performance. In this section, we examine how the size of the base-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "score": 1.0, + "content": "knowledge (i.e., number of classes) affects FearNet’s performance on CUB-200. To do this, we", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "varied the size of the base-knowledge from 10-150 classes, with the remaining classes learned in-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "crementally. Detailed plots are provided in the Supplemental Material. As the base-knowledge size", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "increases, there is a noticeable increase in overall model performance because 1) mPFC has a bet-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 505, + 178 + ], + "score": 1.0, + "content": "ter learned representation from a larger quantity of data and 2) there are not as many incremental", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 483, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 483, + 189 + ], + "score": 1.0, + "content": "learning steps remaining for the dataset, so the base-knowledge performance is less perturbed.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 225, + 190, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 192, + 240 + ], + "score": 1.0, + "content": "7 DISCUSSION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "FearNet’s mPFC is trained to both discriminate examples and also generate new examples. While", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "the main use of mPFC’s generative abilities is to enable psuedorehearsal, this ability may also help", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "score": 1.0, + "content": "make the model more robust to catastrophic forgetting. Gillies (1991) observed that unsupervised", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "networks are more robust (but not immune) to catastrophic forgetting because there are no target out-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "puts to be forgotten. Since the pseudoexample generator is learned as a unsupervised reconstruction", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 389, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 389, + 331 + ], + "score": 1.0, + "content": "task, this could explain why FearNet is slow to forget old information.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 324, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 325, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 325, + 347 + ], + "score": 1.0, + "content": "Table 5 shows the memory requirements for each", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 325, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 325, + 358 + ], + "score": 1.0, + "content": "model in Sec. 6.1 for learning CIFAR-100 and a hypo-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 325, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 325, + 368 + ], + "score": 1.0, + "content": "thetical extrapolation for learning 1,000 classes. This", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 325, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 325, + 380 + ], + "score": 1.0, + "content": "chart accounts for a fixed model capacity and stor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 378, + 325, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 325, + 392 + ], + "score": 1.0, + "content": "age of any data or class statistics. FearNet’s memory", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 389, + 325, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 325, + 402 + ], + "score": 1.0, + "content": "footprint is comparatively small because it only stores", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 400, + 325, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 325, + 412 + ], + "score": 1.0, + "content": "class statistics rather than some or all of the raw train-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 324, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 324, + 424 + ], + "score": 1.0, + "content": "ing data, which makes it better suited for deployment.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 332, + 336, + 511, + 419 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 332, + 336, + 511, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 332, + 336, + 511, + 419 + ], + "spans": [ + { + "bbox": [ + 332, + 336, + 511, + 419 + ], + "score": 0.979, + "html": "
Model100 Classes1,000 Classes
1-NN4.1 GB40.9 GB
GeppNet+STM4.1 GB41.0 GB
GeppNet4.1 GB41.0 GB
FEL272.5MB395.0 MB
iCaRL17.6 MB166.0 MB
FearNet10.7MB74.4 MB
", + "type": "table", + "image_path": "941bde1bb3259bfb43548610b60ce1a256f30619d2d088c1f9c3da493fe7a5e3.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 332, + 336, + 511, + 349.8333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 332, + 349.8333333333333, + 511, + 363.66666666666663 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 332, + 363.66666666666663, + 511, + 377.49999999999994 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 332, + 377.49999999999994, + 511, + 391.33333333333326 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 332, + 391.33333333333326, + 511, + 405.1666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 332, + 405.1666666666666, + 511, + 418.9999999999999 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 429, + 324, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 325, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 325, + 442 + ], + "score": 1.0, + "content": "An open question is how to deal with storage and up-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 439, + 325, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 325, + 451 + ], + "score": 1.0, + "content": "dating of class statistics if classes are seen in more than", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 450, + 326, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 326, + 464 + ], + "score": 1.0, + "content": "one study sessions. One possibility is to use a running", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 461, + 325, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 325, + 475 + ], + "score": 1.0, + "content": "update for the class means and covariances, but it may", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 332, + 428, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 331, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 331, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "Table 5: Memory requirements to train", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 331, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 331, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "CIFAR-100 and the amount of memory", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 331, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 331, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "that would be required if these models", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 331, + 460, + 465, + 472 + ], + "spans": [ + { + "bbox": [ + 331, + 460, + 465, + 472 + ], + "score": 1.0, + "content": "were trained up to 1,000 classes.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 108, + 473, + 490, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 491, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 491, + 485 + ], + "score": 1.0, + "content": "be better to favor the data from the most recent study session due to learning in the autoencoder.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "FearNet assumed that the output of the mPFC encoder was normally distributed for each class, which", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "may not be the case. It would be interesting to consider modeling the classes with a more complex", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "model, e.g., a Gaussian Mixture Model. Robins (1995) showed that pseudorehearsal worked rea-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 523, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 506, + 534 + ], + "score": 1.0, + "content": "sonably well with randomly generated vectors because they were associated with the weights of a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "given class. Replaying these vectors strengthened their corresponding weights, which could be what", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 545, + 397, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 397, + 557 + ], + "score": 1.0, + "content": "is happening with the pseudo-examples generated by FearNet’s decoder.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 344, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 345, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 345, + 573 + ], + "score": 1.0, + "content": "The largest impact on model size is the stored covariance", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 572, + 345, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 136, + 583 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 136, + 572, + 149, + 583 + ], + "score": 0.88, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 572, + 345, + 583 + ], + "score": 1.0, + "content": "for each class. We tested a variant of FearNet", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 582, + 345, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 189, + 595 + ], + "score": 1.0, + "content": "that used a diagonal", + "type": "text" + }, + { + "bbox": [ + 189, + 583, + 202, + 594 + ], + "score": 0.9, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 582, + 345, + 595 + ], + "score": 1.0, + "content": "instead of a full covariance matrix.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 593, + 345, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 345, + 606 + ], + "score": 1.0, + "content": "Table 6 shows that performance degrades, but FearNet still", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 605, + 136, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 136, + 617 + ], + "score": 1.0, + "content": "works.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + }, + { + "type": "table", + "bbox": [ + 352, + 563, + 507, + 635 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 352, + 563, + 507, + 635 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 352, + 563, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 352, + 563, + 507, + 635 + ], + "score": 0.976, + "html": "
Full CovarianceDiagonal Covariance
Sbase0.9420.781
Snew0.8050.877
Ωall0.9590.800
Model Size10.7MB3.8 MB
", + "type": "table", + "image_path": "10db10d0235e7c758aa4bcb903a7aab3d4d4006fb61a06906bea9436e2ed2912.jpg" + } + ] + } + ], + "index": 52, + "virtual_lines": [ + { + "bbox": [ + 352, + 563, + 507, + 577.4 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 352, + 577.4, + 507, + 591.8 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 352, + 591.8, + 507, + 606.1999999999999 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 352, + 606.1999999999999, + 507, + 620.5999999999999 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 352, + 620.5999999999999, + 507, + 634.9999999999999 + ], + "spans": [], + "index": 61 + } + ] + } + ], + "index": 52 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 343, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 345, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 345, + 634 + ], + "score": 1.0, + "content": "FearNet can be adapted to other paradigms, such as unsu-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 632, + 345, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 345, + 645 + ], + "score": 1.0, + "content": "pervised learning and regression. For unsupervised learn-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 643, + 345, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 345, + 657 + ], + "score": 1.0, + "content": "ing, FearNet’s mPFC already does a form of it implicitly.", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 655, + 345, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 345, + 667 + ], + "score": 1.0, + "content": "For regression, this would require changing mPFC’s loss", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 665, + 345, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 345, + 678 + ], + "score": 1.0, + "content": "function and may require grouping input feature vectors", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 676, + 345, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 345, + 689 + ], + "score": 1.0, + "content": "into similar collections. FearNet could also be adapted to", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 688, + 345, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 345, + 700 + ], + "score": 1.0, + "content": "perform the supervised data permutation experiment per-", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 352, + 644, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 351, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 351, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "Table 6: Using a diagonal covariance", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 352, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 352, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "matrix for FearNet’s class statistics", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 351, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 351, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "instead of a full covariance matrix on", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 352, + 677, + 404, + 687 + ], + "spans": [ + { + "bbox": [ + 352, + 677, + 404, + 687 + ], + "score": 1.0, + "content": "CIFAR-100.", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 63.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 713 + ], + "score": 1.0, + "content": "formed by Goodfellow et al. (2013) and Kirkpatrick et al. (2017). This would likely require storing", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "statistics from previous permutations and classes. FearNet would sleep between learning different", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 721, + 438, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 438, + 733 + ], + "score": 1.0, + "content": "permutations; however, if the number of classes was high, recent recall may suffer.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 67 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 104, + 81, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "Base-Knowledge Effect on Performance. In this section, we examine how the size of the base-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "score": 1.0, + "content": "knowledge (i.e., number of classes) affects FearNet’s performance on CUB-200. To do this, we", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "varied the size of the base-knowledge from 10-150 classes, with the remaining classes learned in-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "crementally. Detailed plots are provided in the Supplemental Material. As the base-knowledge size", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "increases, there is a noticeable increase in overall model performance because 1) mPFC has a bet-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 505, + 178 + ], + "score": 1.0, + "content": "ter learned representation from a larger quantity of data and 2) there are not as many incremental", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 483, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 483, + 189 + ], + "score": 1.0, + "content": "learning steps remaining for the dataset, so the base-knowledge performance is less perturbed.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 110, + 505, + 189 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 225, + 190, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 192, + 240 + ], + "score": 1.0, + "content": "7 DISCUSSION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "FearNet’s mPFC is trained to both discriminate examples and also generate new examples. While", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "the main use of mPFC’s generative abilities is to enable psuedorehearsal, this ability may also help", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "score": 1.0, + "content": "make the model more robust to catastrophic forgetting. Gillies (1991) observed that unsupervised", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "networks are more robust (but not immune) to catastrophic forgetting because there are no target out-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "puts to be forgotten. Since the pseudoexample generator is learned as a unsupervised reconstruction", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 389, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 389, + 331 + ], + "score": 1.0, + "content": "task, this could explain why FearNet is slow to forget old information.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 263, + 506, + 331 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 324, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 325, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 325, + 347 + ], + "score": 1.0, + "content": "Table 5 shows the memory requirements for each", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 325, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 325, + 358 + ], + "score": 1.0, + "content": "model in Sec. 6.1 for learning CIFAR-100 and a hypo-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 325, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 325, + 368 + ], + "score": 1.0, + "content": "thetical extrapolation for learning 1,000 classes. This", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 325, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 325, + 380 + ], + "score": 1.0, + "content": "chart accounts for a fixed model capacity and stor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 378, + 325, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 325, + 392 + ], + "score": 1.0, + "content": "age of any data or class statistics. FearNet’s memory", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 389, + 325, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 325, + 402 + ], + "score": 1.0, + "content": "footprint is comparatively small because it only stores", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 400, + 325, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 325, + 412 + ], + "score": 1.0, + "content": "class statistics rather than some or all of the raw train-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 324, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 324, + 424 + ], + "score": 1.0, + "content": "ing data, which makes it better suited for deployment.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 334, + 325, + 424 + ] + }, + { + "type": "table", + "bbox": [ + 332, + 336, + 511, + 419 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 332, + 336, + 511, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 332, + 336, + 511, + 419 + ], + "spans": [ + { + "bbox": [ + 332, + 336, + 511, + 419 + ], + "score": 0.979, + "html": "
Model100 Classes1,000 Classes
1-NN4.1 GB40.9 GB
GeppNet+STM4.1 GB41.0 GB
GeppNet4.1 GB41.0 GB
FEL272.5MB395.0 MB
iCaRL17.6 MB166.0 MB
FearNet10.7MB74.4 MB
", + "type": "table", + "image_path": "941bde1bb3259bfb43548610b60ce1a256f30619d2d088c1f9c3da493fe7a5e3.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 332, + 336, + 511, + 349.8333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 332, + 349.8333333333333, + 511, + 363.66666666666663 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 332, + 363.66666666666663, + 511, + 377.49999999999994 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 332, + 377.49999999999994, + 511, + 391.33333333333326 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 332, + 391.33333333333326, + 511, + 405.1666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 332, + 405.1666666666666, + 511, + 418.9999999999999 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 429, + 324, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 325, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 325, + 442 + ], + "score": 1.0, + "content": "An open question is how to deal with storage and up-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 439, + 325, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 325, + 451 + ], + "score": 1.0, + "content": "dating of class statistics if classes are seen in more than", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 450, + 326, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 326, + 464 + ], + "score": 1.0, + "content": "one study sessions. One possibility is to use a running", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 461, + 325, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 325, + 475 + ], + "score": 1.0, + "content": "update for the class means and covariances, but it may", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 427, + 326, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 332, + 428, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 331, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 331, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "Table 5: Memory requirements to train", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 331, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 331, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "CIFAR-100 and the amount of memory", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 331, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 331, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "that would be required if these models", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 331, + 460, + 465, + 472 + ], + "spans": [ + { + "bbox": [ + 331, + 460, + 465, + 472 + ], + "score": 1.0, + "content": "were trained up to 1,000 classes.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 331, + 426, + 505, + 472 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 473, + 490, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 491, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 491, + 485 + ], + "score": 1.0, + "content": "be better to favor the data from the most recent study session due to learning in the autoencoder.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 471, + 491, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "FearNet assumed that the output of the mPFC encoder was normally distributed for each class, which", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "may not be the case. It would be interesting to consider modeling the classes with a more complex", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "model, e.g., a Gaussian Mixture Model. Robins (1995) showed that pseudorehearsal worked rea-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 523, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 506, + 534 + ], + "score": 1.0, + "content": "sonably well with randomly generated vectors because they were associated with the weights of a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "given class. Replaying these vectors strengthened their corresponding weights, which could be what", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 545, + 397, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 397, + 557 + ], + "score": 1.0, + "content": "is happening with the pseudo-examples generated by FearNet’s decoder.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 489, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 344, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 345, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 345, + 573 + ], + "score": 1.0, + "content": "The largest impact on model size is the stored covariance", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 572, + 345, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 136, + 583 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 136, + 572, + 149, + 583 + ], + "score": 0.88, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 572, + 345, + 583 + ], + "score": 1.0, + "content": "for each class. We tested a variant of FearNet", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 582, + 345, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 189, + 595 + ], + "score": 1.0, + "content": "that used a diagonal", + "type": "text" + }, + { + "bbox": [ + 189, + 583, + 202, + 594 + ], + "score": 0.9, + "content": "\\Sigma _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 582, + 345, + 595 + ], + "score": 1.0, + "content": "instead of a full covariance matrix.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 593, + 345, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 345, + 606 + ], + "score": 1.0, + "content": "Table 6 shows that performance degrades, but FearNet still", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 605, + 136, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 136, + 617 + ], + "score": 1.0, + "content": "works.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 561, + 345, + 617 + ] + }, + { + "type": "table", + "bbox": [ + 352, + 563, + 507, + 635 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 352, + 563, + 507, + 635 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 352, + 563, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 352, + 563, + 507, + 635 + ], + "score": 0.976, + "html": "
Full CovarianceDiagonal Covariance
Sbase0.9420.781
Snew0.8050.877
Ωall0.9590.800
Model Size10.7MB3.8 MB
", + "type": "table", + "image_path": "10db10d0235e7c758aa4bcb903a7aab3d4d4006fb61a06906bea9436e2ed2912.jpg" + } + ] + } + ], + "index": 52, + "virtual_lines": [ + { + "bbox": [ + 352, + 563, + 507, + 577.4 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 352, + 577.4, + 507, + 591.8 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 352, + 591.8, + 507, + 606.1999999999999 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 352, + 606.1999999999999, + 507, + 620.5999999999999 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 352, + 620.5999999999999, + 507, + 634.9999999999999 + ], + "spans": [], + "index": 61 + } + ] + } + ], + "index": 52 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 343, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 345, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 345, + 634 + ], + "score": 1.0, + "content": "FearNet can be adapted to other paradigms, such as unsu-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 632, + 345, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 345, + 645 + ], + "score": 1.0, + "content": "pervised learning and regression. For unsupervised learn-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 643, + 345, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 345, + 657 + ], + "score": 1.0, + "content": "ing, FearNet’s mPFC already does a form of it implicitly.", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 655, + 345, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 345, + 667 + ], + "score": 1.0, + "content": "For regression, this would require changing mPFC’s loss", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 665, + 345, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 345, + 678 + ], + "score": 1.0, + "content": "function and may require grouping input feature vectors", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 676, + 345, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 345, + 689 + ], + "score": 1.0, + "content": "into similar collections. FearNet could also be adapted to", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 688, + 345, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 345, + 700 + ], + "score": 1.0, + "content": "perform the supervised data permutation experiment per-", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 57, + "bbox_fs": [ + 105, + 621, + 345, + 700 + ] + }, + { + "type": "text", + "bbox": [ + 352, + 644, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 351, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 351, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "Table 6: Using a diagonal covariance", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 352, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 352, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "matrix for FearNet’s class statistics", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 351, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 351, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "instead of a full covariance matrix on", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 352, + 677, + 404, + 687 + ], + "spans": [ + { + "bbox": [ + 352, + 677, + 404, + 687 + ], + "score": 1.0, + "content": "CIFAR-100.", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 63.5, + "bbox_fs": [ + 351, + 642, + 505, + 687 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 713 + ], + "score": 1.0, + "content": "formed by Goodfellow et al. (2013) and Kirkpatrick et al. (2017). This would likely require storing", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "statistics from previous permutations and classes. FearNet would sleep between learning different", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 721, + 438, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 438, + 733 + ], + "score": 1.0, + "content": "permutations; however, if the number of classes was high, recent recall may suffer.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 67, + "bbox_fs": [ + 105, + 697, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 195, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "score": 1.0, + "content": "8 CONCLUSION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "In this paper, we proposed a brain-inspired framework capable of incrementally learning data with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "score": 1.0, + "content": "different modalities and object classes. FearNet outperforms existing methods for incremental class", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 129, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 504, + 140 + ], + "score": 1.0, + "content": "learning on large image and audio classification benchmarks, demonstrating that FearNet is capable", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "of recalling and consolidating recently learned information while also retaining old information. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 148, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 164 + ], + "score": 1.0, + "content": "addition, we showed that FearNet is more memory efficient, making it ideal for platforms where size,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 160, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 504, + 173 + ], + "score": 1.0, + "content": "weight, and power requirements are limited. Future work will include 1) integrating BLA directly", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "into the model (versus training it independently); 2) replacing HC with a semi-parametric model; 3)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "score": 1.0, + "content": "learning the feature embedding from raw inputs; and 4) replacing the pseduorehearsal mechanism", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "with a generative model that does not require the storage of class statistics, which would be more", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 205, + 179, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 179, + 217 + ], + "score": 1.0, + "content": "memory efficient.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 107, + 233, + 175, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 176, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 176, + 246 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 504, + 264 + ], + "score": 1.0, + "content": "Wickliffe C Abraham and Anthony Robins. Memory retention–the synaptic stability versus plastic-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 262, + 351, + 275 + ], + "spans": [ + { + "bbox": [ + 116, + 262, + 351, + 275 + ], + "score": 1.0, + "content": "ity dilemma. Trends in Neurosciences, 28(2):73–78, 2005.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 503, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 372, + 294 + ], + "score": 1.0, + "content": "Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, et al. Youtube-", + "type": "text" + }, + { + "bbox": [ + 373, + 282, + 387, + 292 + ], + "score": 0.27, + "content": "s \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 281, + 505, + 294 + ], + "score": 1.0, + "content": ": A large-scale video classifi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 292, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 116, + 292, + 295, + 304 + ], + "score": 1.0, + "content": "cation benchmark. arXiv:1609.08675, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 104, + 311, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Joseph Altman. Autoradiographic investigation of cell proliferation in the brains of rats and cats.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 322, + 310, + 334 + ], + "spans": [ + { + "bbox": [ + 116, + 322, + 310, + 334 + ], + "score": 1.0, + "content": "The Anatomical Record, 145(4):573–591, 1963.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 341, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "score": 1.0, + "content": "Bernard Ans and Stphane Rousset. Avoiding catastrophic forgetting by coupling two reverberating", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "neural networks. Comptes Rendus de l’Acadmie des Sciences - Series III - Sciences de la Vie, 320", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 363, + 279, + 375 + ], + "spans": [ + { + "bbox": [ + 116, + 363, + 279, + 375 + ], + "score": 1.0, + "content": "(12):989 – 997, 1997. ISSN 0764-4469.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 504, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "Bruno Bontempi, Catherine Laurent-Demir, Claude Destrade, and Robert Jaffard. Time-dependent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "reorganization of brain circuitry underlying long-term memory storage. Nature, 400(6745):671–", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 403, + 162, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 162, + 416 + ], + "score": 1.0, + "content": "675, 1999.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 423, + 503, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 435 + ], + "score": 1.0, + "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 434, + 351, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 434, + 351, + 447 + ], + "score": 1.0, + "content": "learning by exponential linear units (elus). In ICLR, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 453, + 504, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "Robert Coop, Aaron Mishtal, and Itamar Arel. Ensemble learning in fixed expansion layer networks", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "for mitigating catastrophic forgetting. IEEE Trans. on Neural Networks and Learning Systems,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 475, + 221, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 475, + 221, + 487 + ], + "score": 1.0, + "content": "24(10):1623–1634, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 105, + 493, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "Wei Deng, James B Aimone, and Fred H Gage. New neurons and new memories: how does adult", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 505, + 504, + 517 + ], + "spans": [ + { + "bbox": [ + 116, + 505, + 504, + 517 + ], + "score": 1.0, + "content": "hippocampal neurogenesis affect learning and memory? Nature Reviews Neuroscience, 11(5):", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 516, + 181, + 527 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 181, + 527 + ], + "score": 1.0, + "content": "339–350, 2010.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "score": 1.0, + "content": "Timothy J Draelos, Nadine E Miner, Christopher C Lamb, Jonathan A Cox, Craig M Vineyard,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Kristofor D Carlson, William M Severa, Conrad D James, and James B Aimone. Neurogenesis", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "deep learning: Extending deep networks to accommodate new classes. In International Joint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 568, + 356, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 356, + 580 + ], + "score": 1.0, + "content": "Conference on Neural Networks, pp. 526–533. IEEE, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 108, + 586, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 104, + 584, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 104, + 584, + 506, + 601 + ], + "score": 1.0, + "content": "Peter S Eriksson, Ekaterina Perfilieva, Thomas Bjork-Eriksson, Ann-Marie Alborn, Claes Nordborg, ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 597, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 506, + 611 + ], + "score": 1.0, + "content": "Daniel A Peterson, and Fred H Gage. Neurogenesis in the adult human hippocampus. Nature", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 608, + 257, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 257, + 621 + ], + "score": 1.0, + "content": "medicine, 4(11):1313–1317, 1998.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 627, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "score": 1.0, + "content": "David R Euston, Aaron J Gruber, and Bruce L McNaughton. The role of medial prefrontal cortex in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 639, + 374, + 651 + ], + "spans": [ + { + "bbox": [ + 116, + 639, + 374, + 651 + ], + "score": 1.0, + "content": "memory and decision making. Neuron, 76(6):1057–1070, 2012.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 680, + 289, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 289, + 691 + ], + "score": 1.0, + "content": "neural networks. arXiv:1701.08734, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Paul W Frankland, Bruno Bontempi, Lynn E Talton, Leszek Kaczmarek, and Alcino J Silva. The", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "involvement of the anterior cingulate cortex in remote contextual fear memory. Science, 304", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 211, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 211, + 732 + ], + "score": 1.0, + "content": "(5672):881–883, 2004.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 195, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "score": 1.0, + "content": "8 CONCLUSION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "In this paper, we proposed a brain-inspired framework capable of incrementally learning data with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "score": 1.0, + "content": "different modalities and object classes. FearNet outperforms existing methods for incremental class", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 129, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 504, + 140 + ], + "score": 1.0, + "content": "learning on large image and audio classification benchmarks, demonstrating that FearNet is capable", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "of recalling and consolidating recently learned information while also retaining old information. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 148, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 164 + ], + "score": 1.0, + "content": "addition, we showed that FearNet is more memory efficient, making it ideal for platforms where size,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 160, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 504, + 173 + ], + "score": 1.0, + "content": "weight, and power requirements are limited. Future work will include 1) integrating BLA directly", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "into the model (versus training it independently); 2) replacing HC with a semi-parametric model; 3)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "score": 1.0, + "content": "learning the feature embedding from raw inputs; and 4) replacing the pseduorehearsal mechanism", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "with a generative model that does not require the storage of class statistics, which would be more", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 205, + 179, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 179, + 217 + ], + "score": 1.0, + "content": "memory efficient.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 106, + 506, + 217 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 233, + 175, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 176, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 176, + 246 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 504, + 264 + ], + "score": 1.0, + "content": "Wickliffe C Abraham and Anthony Robins. Memory retention–the synaptic stability versus plastic-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 262, + 351, + 275 + ], + "spans": [ + { + "bbox": [ + 116, + 262, + 351, + 275 + ], + "score": 1.0, + "content": "ity dilemma. Trends in Neurosciences, 28(2):73–78, 2005.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 249, + 504, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 503, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 372, + 294 + ], + "score": 1.0, + "content": "Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, et al. Youtube-", + "type": "text" + }, + { + "bbox": [ + 373, + 282, + 387, + 292 + ], + "score": 0.27, + "content": "s \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 281, + 505, + 294 + ], + "score": 1.0, + "content": ": A large-scale video classifi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 292, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 116, + 292, + 295, + 304 + ], + "score": 1.0, + "content": "cation benchmark. arXiv:1609.08675, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 281, + 505, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 311, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Joseph Altman. Autoradiographic investigation of cell proliferation in the brains of rats and cats.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 322, + 310, + 334 + ], + "spans": [ + { + "bbox": [ + 116, + 322, + 310, + 334 + ], + "score": 1.0, + "content": "The Anatomical Record, 145(4):573–591, 1963.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 311, + 505, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 341, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "score": 1.0, + "content": "Bernard Ans and Stphane Rousset. Avoiding catastrophic forgetting by coupling two reverberating", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "neural networks. Comptes Rendus de l’Acadmie des Sciences - Series III - Sciences de la Vie, 320", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 363, + 279, + 375 + ], + "spans": [ + { + "bbox": [ + 116, + 363, + 279, + 375 + ], + "score": 1.0, + "content": "(12):989 – 997, 1997. ISSN 0764-4469.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 340, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 504, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "Bruno Bontempi, Catherine Laurent-Demir, Claude Destrade, and Robert Jaffard. Time-dependent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "reorganization of brain circuitry underlying long-term memory storage. Nature, 400(6745):671–", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 403, + 162, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 162, + 416 + ], + "score": 1.0, + "content": "675, 1999.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 382, + 505, + 416 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 423, + 503, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 435 + ], + "score": 1.0, + "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 434, + 351, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 434, + 351, + 447 + ], + "score": 1.0, + "content": "learning by exponential linear units (elus). In ICLR, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 422, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 453, + 504, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "Robert Coop, Aaron Mishtal, and Itamar Arel. Ensemble learning in fixed expansion layer networks", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "for mitigating catastrophic forgetting. IEEE Trans. on Neural Networks and Learning Systems,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 475, + 221, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 475, + 221, + 487 + ], + "score": 1.0, + "content": "24(10):1623–1634, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 452, + 506, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 493, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "Wei Deng, James B Aimone, and Fred H Gage. New neurons and new memories: how does adult", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 505, + 504, + 517 + ], + "spans": [ + { + "bbox": [ + 116, + 505, + 504, + 517 + ], + "score": 1.0, + "content": "hippocampal neurogenesis affect learning and memory? Nature Reviews Neuroscience, 11(5):", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 516, + 181, + 527 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 181, + 527 + ], + "score": 1.0, + "content": "339–350, 2010.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 494, + 505, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "score": 1.0, + "content": "Timothy J Draelos, Nadine E Miner, Christopher C Lamb, Jonathan A Cox, Craig M Vineyard,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Kristofor D Carlson, William M Severa, Conrad D James, and James B Aimone. Neurogenesis", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "deep learning: Extending deep networks to accommodate new classes. In International Joint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 568, + 356, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 356, + 580 + ], + "score": 1.0, + "content": "Conference on Neural Networks, pp. 526–533. IEEE, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 534, + 506, + 580 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 586, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 104, + 584, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 104, + 584, + 506, + 601 + ], + "score": 1.0, + "content": "Peter S Eriksson, Ekaterina Perfilieva, Thomas Bjork-Eriksson, Ann-Marie Alborn, Claes Nordborg, ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 597, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 506, + 611 + ], + "score": 1.0, + "content": "Daniel A Peterson, and Fred H Gage. Neurogenesis in the adult human hippocampus. Nature", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 608, + 257, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 257, + 621 + ], + "score": 1.0, + "content": "medicine, 4(11):1313–1317, 1998.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 104, + 584, + 506, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 627, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "score": 1.0, + "content": "David R Euston, Aaron J Gruber, and Bruce L McNaughton. The role of medial prefrontal cortex in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 639, + 374, + 651 + ], + "spans": [ + { + "bbox": [ + 116, + 639, + 374, + 651 + ], + "score": 1.0, + "content": "memory and decision making. Neuron, 76(6):1057–1070, 2012.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 626, + 505, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 680, + 289, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 289, + 691 + ], + "score": 1.0, + "content": "neural networks. arXiv:1701.08734, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 657, + 505, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Paul W Frankland, Bruno Bontempi, Lynn E Talton, Leszek Kaczmarek, and Alcino J Silva. The", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "involvement of the anterior cingulate cortex in remote contextual fear memory. Science, 304", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 211, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 211, + 732 + ], + "score": 1.0, + "content": "(5672):881–883, 2004.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 699, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "score": 1.0, + "content": "Robert M French. Pseudo-recurrent connectionist networks: An approach to the ‘sensitivity-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 364, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 364, + 105 + ], + "score": 1.0, + "content": "stability’ dilemma. Connection Science, 9(4):353–380, 1997.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "score": 1.0, + "content": "Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 200, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 200, + 136 + ], + "score": 1.0, + "content": "3(4):128–135, 1999.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "Steffen Gais, Genevieve Albouy, M ` elanie Boly, Thien Thanh Dang-Vu, Annabelle Darsaud, Martin ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "Desseilles, Geraldine Rauchs, Manuel Schabus, Virginie Sterpenich, Gilles Vandewalle, et al. ´", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 505, + 179 + ], + "score": 1.0, + "content": "Sleep transforms the cerebral trace of declarative memories. Proceedings of the National Academy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 177, + 285, + 188 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 285, + 188 + ], + "score": 1.0, + "content": "of Sciences, 104(47):18778–18783, 2007.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 195, + 504, + 230 + ], + "lines": [ + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "score": 1.0, + "content": "Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 116, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 218, + 321, + 230 + ], + "spans": [ + { + "bbox": [ + 116, + 218, + 321, + 230 + ], + "score": 1.0, + "content": "audio events. In ICASSP, New Orleans, LA, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 104, + 235, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 505, + 253 + ], + "score": 1.0, + "content": "Alexander Gepperth and Cem Karaoguz. A bio-inspired incremental learning architecture for ap-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 408, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 408, + 261 + ], + "score": 1.0, + "content": "plied perceptual problems. Cognitive Computation, 8(5):924–934, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 504, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "AJ Gillies. The Stability/Plasticity Dilemma in Self-organising Neural Networks. PhD thesis, MSc", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 279, + 443, + 292 + ], + "spans": [ + { + "bbox": [ + 117, + 279, + 443, + 292 + ], + "score": 1.0, + "content": "Thesis, Computer Science Department, University of Otago, New Zealand, 1991.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 504, + 332 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 116, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "networks. In Proceedings of the Thirteenth International Conference on Artificial Intelligence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 321, + 254, + 333 + ], + "spans": [ + { + "bbox": [ + 116, + 321, + 254, + 333 + ], + "score": 1.0, + "content": "and Statistics, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 104, + 339, + 504, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "score": 1.0, + "content": "Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical in-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "vestigation of catastrophic forgetting in gradient-based neural networks. arXiv:1312.6211, 2013.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 104, + 370, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 385 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 381, + 266, + 394 + ], + "spans": [ + { + "bbox": [ + 115, + 381, + 266, + 394 + ], + "score": 1.0, + "content": "nition. In CPVR, pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 504, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "P Hetherington and Mark S Seidenberg. Is there catastrophic interference in connectionist networks.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "In Proceedings of the 11th annual conference of the cognitive science society, volume 26, pp. 33.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 424, + 235, + 435 + ], + "spans": [ + { + "bbox": [ + 116, + 424, + 235, + 435 + ], + "score": 1.0, + "content": "Erlbaum Hillsdale, NJ, 1989.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 503, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 457 + ], + "score": 1.0, + "content": "Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 454, + 351, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 351, + 466 + ], + "score": 1.0, + "content": "catastrophic forgetting in neural networks. In AAAI, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, et al. Overcoming catastrophic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 485, + 503, + 496 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 503, + 496 + ], + "score": 1.0, + "content": "forgetting in neural networks. Proc. of the National Academy of Sciences, pp. 201611835, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 504, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 503, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 517 + ], + "score": 1.0, + "content": "Takashi Kitamura, Sachie K Ogawa, Dheeraj S Roy, Teruhiro Okuyama, Mark D Morrissey, Lil-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "lian M Smith, Roger L Redondo, and Susumu Tonegawa. Engrams and circuits crucial for systems", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 526, + 364, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 526, + 364, + 538 + ], + "score": 1.0, + "content": "consolidation of a memory. Science, 356(6333):73–78, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 504, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 115, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in Neural Information Processing Systems 25, pp. 1097–", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 567, + 167, + 579 + ], + "spans": [ + { + "bbox": [ + 116, + 567, + 167, + 579 + ], + "score": 1.0, + "content": "1105, 2012.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 104, + 586, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 504, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 504, + 600 + ], + "score": 1.0, + "content": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In Data Mining, 2008. ICDM’08.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 598, + 394, + 610 + ], + "spans": [ + { + "bbox": [ + 116, + 598, + 394, + 610 + ], + "score": 1.0, + "content": "Eighth IEEE International Conference on, pp. 413–422. IEEE, 2008.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 104, + 616, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 505, + 632 + ], + "score": 1.0, + "content": "James L McClelland, Bruce L McNaughton, and Randall C O’reilly. Why there are complementary", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 116, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "learning systems in the hippocampus and neocortex: insights from the successes and failures of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 640, + 468, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 640, + 468, + 652 + ], + "score": 1.0, + "content": "connectionist models of learning and memory. Psychological review, 102(3):419, 1995.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 503, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 670, + 471, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 471, + 683 + ], + "score": 1.0, + "content": "sequential learning problem. Psychology of learning and motivation, 24:109–165, 1989.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 689, + 443, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 443, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 443, + 703 + ], + "score": 1.0, + "content": "Gina R Poe. Sleep is for forgetting. Journal of Neuroscience, 37(3):464–473, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Bjorn Rasch and Jan Born. About sleeps role in memory. ¨ Physiological reviews, 93(2):681–766,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 143, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 143, + 732 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "score": 1.0, + "content": "Robert M French. Pseudo-recurrent connectionist networks: An approach to the ‘sensitivity-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 364, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 364, + 105 + ], + "score": 1.0, + "content": "stability’ dilemma. Connection Science, 9(4):353–380, 1997.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 504, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "score": 1.0, + "content": "Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 200, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 200, + 136 + ], + "score": 1.0, + "content": "3(4):128–135, 1999.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 506, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "Steffen Gais, Genevieve Albouy, M ` elanie Boly, Thien Thanh Dang-Vu, Annabelle Darsaud, Martin ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "Desseilles, Geraldine Rauchs, Manuel Schabus, Virginie Sterpenich, Gilles Vandewalle, et al. ´", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 505, + 179 + ], + "score": 1.0, + "content": "Sleep transforms the cerebral trace of declarative memories. Proceedings of the National Academy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 177, + 285, + 188 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 285, + 188 + ], + "score": 1.0, + "content": "of Sciences, 104(47):18778–18783, 2007.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 144, + 505, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 195, + 504, + 230 + ], + "lines": [ + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "score": 1.0, + "content": "Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 116, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 218, + 321, + 230 + ], + "spans": [ + { + "bbox": [ + 116, + 218, + 321, + 230 + ], + "score": 1.0, + "content": "audio events. In ICASSP, New Orleans, LA, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 194, + 505, + 230 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 104, + 235, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 505, + 253 + ], + "score": 1.0, + "content": "Alexander Gepperth and Cem Karaoguz. A bio-inspired incremental learning architecture for ap-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 408, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 408, + 261 + ], + "score": 1.0, + "content": "plied perceptual problems. Cognitive Computation, 8(5):924–934, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 235, + 505, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 504, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "AJ Gillies. The Stability/Plasticity Dilemma in Self-organising Neural Networks. PhD thesis, MSc", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 279, + 443, + 292 + ], + "spans": [ + { + "bbox": [ + 117, + 279, + 443, + 292 + ], + "score": 1.0, + "content": "Thesis, Computer Science Department, University of Otago, New Zealand, 1991.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 268, + 506, + 292 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 504, + 332 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 116, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "networks. In Proceedings of the Thirteenth International Conference on Artificial Intelligence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 321, + 254, + 333 + ], + "spans": [ + { + "bbox": [ + 116, + 321, + 254, + 333 + ], + "score": 1.0, + "content": "and Statistics, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 299, + 505, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 339, + 504, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "score": 1.0, + "content": "Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical in-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "vestigation of catastrophic forgetting in gradient-based neural networks. arXiv:1312.6211, 2013.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 339, + 505, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 370, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 385 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 381, + 266, + 394 + ], + "spans": [ + { + "bbox": [ + 115, + 381, + 266, + 394 + ], + "score": 1.0, + "content": "nition. In CPVR, pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 368, + 505, + 394 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 504, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "P Hetherington and Mark S Seidenberg. Is there catastrophic interference in connectionist networks.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "In Proceedings of the 11th annual conference of the cognitive science society, volume 26, pp. 33.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 424, + 235, + 435 + ], + "spans": [ + { + "bbox": [ + 116, + 424, + 235, + 435 + ], + "score": 1.0, + "content": "Erlbaum Hillsdale, NJ, 1989.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 400, + 505, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 503, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 457 + ], + "score": 1.0, + "content": "Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 454, + 351, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 351, + 466 + ], + "score": 1.0, + "content": "catastrophic forgetting in neural networks. In AAAI, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 441, + 505, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, et al. Overcoming catastrophic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 485, + 503, + 496 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 503, + 496 + ], + "score": 1.0, + "content": "forgetting in neural networks. Proc. of the National Academy of Sciences, pp. 201611835, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 473, + 505, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 504, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 503, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 517 + ], + "score": 1.0, + "content": "Takashi Kitamura, Sachie K Ogawa, Dheeraj S Roy, Teruhiro Okuyama, Mark D Morrissey, Lil-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "lian M Smith, Roger L Redondo, and Susumu Tonegawa. Engrams and circuits crucial for systems", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 526, + 364, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 526, + 364, + 538 + ], + "score": 1.0, + "content": "consolidation of a memory. Science, 356(6333):73–78, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 503, + 505, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 504, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 115, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in Neural Information Processing Systems 25, pp. 1097–", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 567, + 167, + 579 + ], + "spans": [ + { + "bbox": [ + 116, + 567, + 167, + 579 + ], + "score": 1.0, + "content": "1105, 2012.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 545, + 506, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 586, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 504, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 504, + 600 + ], + "score": 1.0, + "content": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In Data Mining, 2008. ICDM’08.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 598, + 394, + 610 + ], + "spans": [ + { + "bbox": [ + 116, + 598, + 394, + 610 + ], + "score": 1.0, + "content": "Eighth IEEE International Conference on, pp. 413–422. IEEE, 2008.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 587, + 504, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 104, + 616, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 505, + 632 + ], + "score": 1.0, + "content": "James L McClelland, Bruce L McNaughton, and Randall C O’reilly. Why there are complementary", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 116, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "learning systems in the hippocampus and neocortex: insights from the successes and failures of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 640, + 468, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 640, + 468, + 652 + ], + "score": 1.0, + "content": "connectionist models of learning and memory. Psychological review, 102(3):419, 1995.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 616, + 505, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 503, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 670, + 471, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 471, + 683 + ], + "score": 1.0, + "content": "sequential learning problem. Psychology of learning and motivation, 24:109–165, 1989.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 659, + 505, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 689, + 443, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 443, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 443, + 703 + ], + "score": 1.0, + "content": "Gina R Poe. Sleep is for forgetting. Journal of Neuroscience, 37(3):464–473, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 689, + 443, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Bjorn Rasch and Jan Born. About sleeps role in memory. ¨ Physiological reviews, 93(2):681–766,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 143, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 143, + 732 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 414, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 414, + 106 + ], + "score": 1.0, + "content": "supervised learning with ladder networks. In NIPS, pp. 3546–3554, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, and Christoph H Lampert. iCaRL: Incremental", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 335, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 335, + 135 + ], + "score": 1.0, + "content": "classifier and representation learning. In CVPR, 2017.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 503, + 164 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 153, + 181, + 164 + ], + "spans": [ + { + "bbox": [ + 117, + 153, + 181, + 164 + ], + "score": 1.0, + "content": "123–146, 1995.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 104, + 172, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "Peter J Rousseeuw and Katrien Van Driessen. A fast algorithm for the minimum covariance deter-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 341, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 341, + 195 + ], + "score": 1.0, + "content": "minant estimator. Technometrics, 41(3):212–223, 1999.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 109, + 201, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 200, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 505, + 215 + ], + "score": 1.0, + "content": "Bernhard Scholkopf, John C. Platt, John C. Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. ¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 116, + 213, + 505, + 225 + ], + "score": 1.0, + "content": "Estimating the support of a high-dimensional distribution. Neural Comput., 13(7):1443–1471,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 161, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 161, + 236 + ], + "score": 1.0, + "content": "July 2001.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 104, + 242, + 456, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 457, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 457, + 255 + ], + "score": 1.0, + "content": "Donald F Specht. Probabilistic neural networks. Neural networks, 3(1):109–118, 1990.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 506, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 262, + 504, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 504, + 272 + ], + "score": 1.0, + "content": "A Takashima, Karl Magnus Petersson, F Rutters, I Tendolkar, O Jensen, MJ Zwarts, BL Mc-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "Naughton, and G Fernandez. Declarative memory consolidation in humans: a prospective func-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 115, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "tional magnetic resonance imaging study. Proceedings of the National Academy of Sciences of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 295, + 330, + 307 + ], + "spans": [ + { + "bbox": [ + 116, + 295, + 330, + 307 + ], + "score": 1.0, + "content": "the United States of America, 103(3):756–761, 2006.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 502, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 312, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 327 + ], + "score": 1.0, + "content": "Philippe Taupin and Fred H Gage. Adult neurogenesis and neural stem cells of the central nervous", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 325, + 428, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 325, + 428, + 336 + ], + "score": 1.0, + "content": "system in mammals. Journal of neuroscience research, 69(6):745–749, 2002.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 108, + 342, + 502, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 504, + 356 + ], + "score": 1.0, + "content": "P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 353, + 478, + 366 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 478, + 366 + ], + "score": 1.0, + "content": "Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 108, + 372, + 503, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "Seung-Schik Yoo, Peter T Hu, Ninad Gujar, Ferenc A Jolesz, and Matthew P Walker. A deficit in the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 384, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 116, + 384, + 504, + 397 + ], + "score": 1.0, + "content": "ability to form new human memories without sleep. Nature Neuroscience, 10(3):385–392, 2007.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 108, + 416, + 268, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 270, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 270, + 431 + ], + "score": 1.0, + "content": "A SUPPLEMENTAL MATERIAL", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 441, + 256, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 257, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 257, + 453 + ], + "score": 1.0, + "content": "A.1 MODEL HYPERPARAMETERS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 461, + 504, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "Table S1 shows the training parameters for the FearNet model for each dataset. We also experi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "mented with various dropout rates, weight decay, and various activation functions; however, weight", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 484, + 298, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 298, + 496 + ], + "score": 1.0, + "content": "decay did not work well with FearNet’s mPFC.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 186, + 502, + 425, + 666 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 186, + 502, + 425, + 666 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 186, + 502, + 425, + 666 + ], + "spans": [ + { + "bbox": [ + 186, + 502, + 425, + 666 + ], + "score": 0.978, + "html": "
HyperparameterValues
Learning Rate2.10-3 450 (AudioSet & CIFAR-100)
Mini-Batch Size200 (CUB-200)
mPFCBase-Knowledge Epochs Memory Consolidation Epochs1,000
60
BLA Training Epochs20
CIFAR-100: [140,130]
Hidden Layer SizeCUB-200:[350,300]
AudioSet: [300,100]
Sleep Frequency10 (see Sec. 6.2)
Dropout Rate0.25
Unsupervised Loss Weights (入)[104,1.0,0.1]
Hidden Layer Activation Weight DecayExponential Linear Units 0.0
", + "type": "table", + "image_path": "23f51a66dbafa598dd87744059dff16e1ee9f7149f9010b481ef42a41a0dfd77.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 502, + 425, + 515.6666666666666 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 186, + 515.6666666666666, + 425, + 529.3333333333333 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 186, + 529.3333333333333, + 425, + 542.9999999999999 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 186, + 542.9999999999999, + 425, + 556.6666666666665 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 186, + 556.6666666666665, + 425, + 570.3333333333331 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 186, + 570.3333333333331, + 425, + 583.9999999999998 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 186, + 583.9999999999998, + 425, + 597.6666666666664 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 186, + 597.6666666666664, + 425, + 611.333333333333 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 186, + 611.333333333333, + 425, + 624.9999999999997 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 186, + 624.9999999999997, + 425, + 638.6666666666663 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 186, + 638.6666666666663, + 425, + 652.3333333333329 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 186, + 652.3333333333329, + 425, + 665.9999999999995 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 227, + 675, + 384, + 686 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 225, + 672, + 385, + 688 + ], + "spans": [ + { + "bbox": [ + 225, + 672, + 385, + 688 + ], + "score": 1.0, + "content": "Table S1: FearNet Training Parameters", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + } + ], + "index": 35.75 + }, + { + "type": "text", + "bbox": [ + 106, + 688, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "Table S2 shows the training parameters for the iCaRL framework used in this paper. We adapted", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the code from the author’s GitHub page for our own experiments. The ResNet-18 convolutional", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 710, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 722 + ], + "score": 1.0, + "content": "neural network was replaced with a fully-connected neural network. We experimented with various", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "regularization strategies to increase the initial base-knowledge accuracy with weight decay working", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 414, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 414, + 106 + ], + "score": 1.0, + "content": "supervised learning with ladder networks. In NIPS, pp. 3546–3554, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, and Christoph H Lampert. iCaRL: Incremental", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 335, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 335, + 135 + ], + "score": 1.0, + "content": "classifier and representation learning. In CVPR, 2017.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 112, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 503, + 164 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 153, + 181, + 164 + ], + "spans": [ + { + "bbox": [ + 117, + 153, + 181, + 164 + ], + "score": 1.0, + "content": "123–146, 1995.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 141, + 505, + 164 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 172, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "Peter J Rousseeuw and Katrien Van Driessen. A fast algorithm for the minimum covariance deter-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 341, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 341, + 195 + ], + "score": 1.0, + "content": "minant estimator. Technometrics, 41(3):212–223, 1999.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 171, + 505, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 201, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 200, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 505, + 215 + ], + "score": 1.0, + "content": "Bernhard Scholkopf, John C. Platt, John C. Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. ¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 116, + 213, + 505, + 225 + ], + "score": 1.0, + "content": "Estimating the support of a high-dimensional distribution. Neural Comput., 13(7):1443–1471,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 161, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 161, + 236 + ], + "score": 1.0, + "content": "July 2001.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 200, + 505, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 242, + 456, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 457, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 457, + 255 + ], + "score": 1.0, + "content": "Donald F Specht. Probabilistic neural networks. Neural networks, 3(1):109–118, 1990.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 241, + 457, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 506, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 262, + 504, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 504, + 272 + ], + "score": 1.0, + "content": "A Takashima, Karl Magnus Petersson, F Rutters, I Tendolkar, O Jensen, MJ Zwarts, BL Mc-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "Naughton, and G Fernandez. Declarative memory consolidation in humans: a prospective func-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 115, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "tional magnetic resonance imaging study. Proceedings of the National Academy of Sciences of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 295, + 330, + 307 + ], + "spans": [ + { + "bbox": [ + 116, + 295, + 330, + 307 + ], + "score": 1.0, + "content": "the United States of America, 103(3):756–761, 2006.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 262, + 506, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 502, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 312, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 327 + ], + "score": 1.0, + "content": "Philippe Taupin and Fred H Gage. Adult neurogenesis and neural stem cells of the central nervous", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 325, + 428, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 325, + 428, + 336 + ], + "score": 1.0, + "content": "system in mammals. Journal of neuroscience research, 69(6):745–749, 2002.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 312, + 505, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 342, + 502, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 504, + 356 + ], + "score": 1.0, + "content": "P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 353, + 478, + 366 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 478, + 366 + ], + "score": 1.0, + "content": "Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 343, + 504, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 372, + 503, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "Seung-Schik Yoo, Peter T Hu, Ninad Gujar, Ferenc A Jolesz, and Matthew P Walker. A deficit in the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 384, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 116, + 384, + 504, + 397 + ], + "score": 1.0, + "content": "ability to form new human memories without sleep. Nature Neuroscience, 10(3):385–392, 2007.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 372, + 505, + 397 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 416, + 268, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 270, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 270, + 431 + ], + "score": 1.0, + "content": "A SUPPLEMENTAL MATERIAL", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 441, + 256, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 257, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 257, + 453 + ], + "score": 1.0, + "content": "A.1 MODEL HYPERPARAMETERS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 461, + 504, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "Table S1 shows the training parameters for the FearNet model for each dataset. We also experi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "mented with various dropout rates, weight decay, and various activation functions; however, weight", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 484, + 298, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 298, + 496 + ], + "score": 1.0, + "content": "decay did not work well with FearNet’s mPFC.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 462, + 505, + 496 + ] + }, + { + "type": "table", + "bbox": [ + 186, + 502, + 425, + 666 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 186, + 502, + 425, + 666 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 186, + 502, + 425, + 666 + ], + "spans": [ + { + "bbox": [ + 186, + 502, + 425, + 666 + ], + "score": 0.978, + "html": "
HyperparameterValues
Learning Rate2.10-3 450 (AudioSet & CIFAR-100)
Mini-Batch Size200 (CUB-200)
mPFCBase-Knowledge Epochs Memory Consolidation Epochs1,000
60
BLA Training Epochs20
CIFAR-100: [140,130]
Hidden Layer SizeCUB-200:[350,300]
AudioSet: [300,100]
Sleep Frequency10 (see Sec. 6.2)
Dropout Rate0.25
Unsupervised Loss Weights (入)[104,1.0,0.1]
Hidden Layer Activation Weight DecayExponential Linear Units 0.0
", + "type": "table", + "image_path": "23f51a66dbafa598dd87744059dff16e1ee9f7149f9010b481ef42a41a0dfd77.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 502, + 425, + 515.6666666666666 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 186, + 515.6666666666666, + 425, + 529.3333333333333 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 186, + 529.3333333333333, + 425, + 542.9999999999999 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 186, + 542.9999999999999, + 425, + 556.6666666666665 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 186, + 556.6666666666665, + 425, + 570.3333333333331 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 186, + 570.3333333333331, + 425, + 583.9999999999998 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 186, + 583.9999999999998, + 425, + 597.6666666666664 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 186, + 597.6666666666664, + 425, + 611.333333333333 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 186, + 611.333333333333, + 425, + 624.9999999999997 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 186, + 624.9999999999997, + 425, + 638.6666666666663 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 186, + 638.6666666666663, + 425, + 652.3333333333329 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 186, + 652.3333333333329, + 425, + 665.9999999999995 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 227, + 675, + 384, + 686 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 225, + 672, + 385, + 688 + ], + "spans": [ + { + "bbox": [ + 225, + 672, + 385, + 688 + ], + "score": 1.0, + "content": "Table S1: FearNet Training Parameters", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + } + ], + "index": 35.75 + }, + { + "type": "text", + "bbox": [ + 106, + 688, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "Table S2 shows the training parameters for the iCaRL framework used in this paper. We adapted", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the code from the author’s GitHub page for our own experiments. The ResNet-18 convolutional", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 710, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 722 + ], + "score": 1.0, + "content": "neural network was replaced with a fully-connected neural network. We experimented with various", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "regularization strategies to increase the initial base-knowledge accuracy with weight decay working", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 686, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 199, + 101, + 412, + 205 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 102, + 82, + 478, + 95 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 81, + 479, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 479, + 96 + ], + "score": 1.0, + "content": "the best. The values that are given as a range of values are the hyperparameter search spaces.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 199, + 101, + 412, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 199, + 101, + 412, + 205 + ], + "spans": [ + { + "bbox": [ + 199, + 101, + 412, + 205 + ], + "score": 0.978, + "html": "
HyperparameterValues
Learning Rate2.10-3
Mini-Batch Size450
Exemplars per Class (EPC)20
Hidden Layer Size64-1024
Number of Hidden Layers2-4
Dropout Rate[0.5,0.75,1.00]
HiddenLayer ActivationReLU
Weight Decay[0.0,10-5,10-4,5 · 10-4]
", + "type": "table", + "image_path": "f546dfb66f0ea585630b26beca6f7294e314c617740dab7e07158173e53b780e.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 199, + 101, + 412, + 115.85714285714286 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 199, + 115.85714285714286, + 412, + 130.71428571428572 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 199, + 130.71428571428572, + 412, + 145.57142857142858 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 199, + 145.57142857142858, + 412, + 160.42857142857144 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 199, + 160.42857142857144, + 412, + 175.2857142857143 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 199, + 175.2857142857143, + 412, + 190.14285714285717 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 199, + 190.14285714285717, + 412, + 205.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 229, + 213, + 382, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 228, + 212, + 382, + 226 + ], + "spans": [ + { + "bbox": [ + 228, + 212, + 382, + 226 + ], + "score": 1.0, + "content": "Table S2: iCaRL Training Parameters", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 171, + 267, + 441, + 320 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 227, + 506, + 260 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "score": 1.0, + "content": "Table S3 shows the training parameters for GeppNet and GeppNet+STM. Parameters not listed here", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "score": 1.0, + "content": "are the default parameters defined by Gepperth & Karaoguz (2016). The values that are given as a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 323, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 323, + 262 + ], + "score": 1.0, + "content": "range of values are the hyperparameter search spaces.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "table_body", + "bbox": [ + 171, + 267, + 441, + 320 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 171, + 267, + 441, + 320 + ], + "spans": [ + { + "bbox": [ + 171, + 267, + 441, + 320 + ], + "score": 0.972, + "html": "
HyperparameterValues
SOM Lattice Shape (N)20-36
Non-Linearity Suppression Threshold (0)0.1-0.75
Incremental Class Learning Iterations (Tinc2 - Tinc1)[2,000,20,000]
", + "type": "table", + "image_path": "48863b19a866539f6560fc0e33e0d39a3c3e5fa2c77190ae47e50b65cce20fb3.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 171, + 267, + 441, + 284.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 171, + 284.6666666666667, + 441, + 302.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 171, + 302.33333333333337, + 441, + 320.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 225, + 328, + 387, + 340 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 223, + 327, + 388, + 342 + ], + "spans": [ + { + "bbox": [ + 223, + 327, + 388, + 342 + ], + "score": 1.0, + "content": "Table S3: GeppNet Training Parameters", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 103, + 341, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "Table S4 shows the training parameters for the Fixed Expansion Layer (FEL). The number of units", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 351, + 220, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 220, + 365 + ], + "score": 1.0, + "content": "in the FEL layer is given by", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 362, + 358, + 388 + ], + "lines": [ + { + "bbox": [ + 253, + 362, + 358, + 388 + ], + "spans": [ + { + "bbox": [ + 253, + 362, + 358, + 388 + ], + "score": 0.93, + "content": "\\mathrm { F E L ~ U n i t s } = { \\frac { H ^ { 2 } + H K } { K } }", + "type": "interline_equation", + "image_path": "5e73f5685a7868c3d14e025b929b9a9728b8e4a8c195152ec256afaeeefb4189.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 253, + 362, + 358, + 388 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 133, + 402 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 390, + 144, + 399 + ], + "score": 0.79, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 389, + 351, + 402 + ], + "score": 1.0, + "content": "is the number of units in the first hidden-layer and", + "type": "text" + }, + { + "bbox": [ + 351, + 390, + 361, + 399 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "is the maximum number of classes", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 400, + 500, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 500, + 414 + ], + "score": 1.0, + "content": "in the dataset. The values that are given as a range of values are the hyperparameter search spaces.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 223, + 419, + 387, + 492 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 223, + 419, + 387, + 492 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 223, + 419, + 387, + 492 + ], + "spans": [ + { + "bbox": [ + 223, + 419, + 387, + 492 + ], + "score": 0.977, + "html": "
HyperparameterValues
Hidden Layer Size (H)64-1800
FEL Layer Size Number of Hidden LayersSee Equation 6
Mini-Batch Size2 8
Initial Learning Rate10-2
", + "type": "table", + "image_path": "39b7f0e874056c8831308643934f5229027de7d5c569b5e0a1430cfe73e11b2b.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 223, + 419, + 387, + 433.6 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 223, + 433.6, + 387, + 448.20000000000005 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 223, + 448.20000000000005, + 387, + 462.80000000000007 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 223, + 462.80000000000007, + 387, + 477.4000000000001 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 223, + 477.4000000000001, + 387, + 492.0000000000001 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 234, + 501, + 378, + 513 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 233, + 500, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 233, + 500, + 378, + 514 + ], + "score": 1.0, + "content": "Table S4: FEL Training Parameters", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 343, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 343, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 343, + 535 + ], + "score": 1.0, + "content": "A.2 ICARL PERFORMANCE WITH MORE EXEMPLARS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "Table S5 provides additional experimental results for when there are more exemplars per class", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "(EPC) for the iCaRL framework. Rebuffi et al. (2017) used 20 EPC in their original paper; however,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "we increased the number to 100 EPC to see if storing more training data helped iCaRL. Although", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "a higher EPC does increase iCaRL performance, it still does not outperform FearNet. Note that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "CUB-200 only has about 30 training samples per class, so iCaRL is storing the entire training set", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 598, + 340, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 340, + 609 + ], + "score": 1.0, + "content": "for 100 EPC. Our main results use the default value of 20.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 633, + 202, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 204, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 204, + 646 + ], + "score": 1.0, + "content": "A.3 BLA VARIANTS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Our BLA model is a classifier that determines whether a prediction should be made using HC (recent", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "memory) or mPFC (remote memory). An alternative approach would be to use an outlier detection", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "algorithm that determines whether the data being processed by a sub-network is an outlier for that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "sub-network and should therefore be processed by the other sub-network. To explore this alternative", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "BLA formulation, we experimented with three outlier detection algorithms: 1) one-class support", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "vector machine (SVM) (Scholkopf et al., 2001), 2) determining if the data fits into a Gaussian dis- ¨", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "tribution using a minimum covariance determinant estimation (i.e., elliptical envelope) (Rousseeuw", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 199, + 101, + 412, + 205 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 102, + 82, + 478, + 95 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 81, + 479, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 479, + 96 + ], + "score": 1.0, + "content": "the best. The values that are given as a range of values are the hyperparameter search spaces.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 199, + 101, + 412, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 199, + 101, + 412, + 205 + ], + "spans": [ + { + "bbox": [ + 199, + 101, + 412, + 205 + ], + "score": 0.978, + "html": "
HyperparameterValues
Learning Rate2.10-3
Mini-Batch Size450
Exemplars per Class (EPC)20
Hidden Layer Size64-1024
Number of Hidden Layers2-4
Dropout Rate[0.5,0.75,1.00]
HiddenLayer ActivationReLU
Weight Decay[0.0,10-5,10-4,5 · 10-4]
", + "type": "table", + "image_path": "f546dfb66f0ea585630b26beca6f7294e314c617740dab7e07158173e53b780e.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 199, + 101, + 412, + 115.85714285714286 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 199, + 115.85714285714286, + 412, + 130.71428571428572 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 199, + 130.71428571428572, + 412, + 145.57142857142858 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 199, + 145.57142857142858, + 412, + 160.42857142857144 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 199, + 160.42857142857144, + 412, + 175.2857142857143 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 199, + 175.2857142857143, + 412, + 190.14285714285717 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 199, + 190.14285714285717, + 412, + 205.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 229, + 213, + 382, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 228, + 212, + 382, + 226 + ], + "spans": [ + { + "bbox": [ + 228, + 212, + 382, + 226 + ], + "score": 1.0, + "content": "Table S2: iCaRL Training Parameters", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 171, + 267, + 441, + 320 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 227, + 506, + 260 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "score": 1.0, + "content": "Table S3 shows the training parameters for GeppNet and GeppNet+STM. Parameters not listed here", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "score": 1.0, + "content": "are the default parameters defined by Gepperth & Karaoguz (2016). The values that are given as a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 323, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 323, + 262 + ], + "score": 1.0, + "content": "range of values are the hyperparameter search spaces.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "table_body", + "bbox": [ + 171, + 267, + 441, + 320 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 171, + 267, + 441, + 320 + ], + "spans": [ + { + "bbox": [ + 171, + 267, + 441, + 320 + ], + "score": 0.972, + "html": "
HyperparameterValues
SOM Lattice Shape (N)20-36
Non-Linearity Suppression Threshold (0)0.1-0.75
Incremental Class Learning Iterations (Tinc2 - Tinc1)[2,000,20,000]
", + "type": "table", + "image_path": "48863b19a866539f6560fc0e33e0d39a3c3e5fa2c77190ae47e50b65cce20fb3.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 171, + 267, + 441, + 284.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 171, + 284.6666666666667, + 441, + 302.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 171, + 302.33333333333337, + 441, + 320.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 225, + 328, + 387, + 340 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 223, + 327, + 388, + 342 + ], + "spans": [ + { + "bbox": [ + 223, + 327, + 388, + 342 + ], + "score": 1.0, + "content": "Table S3: GeppNet Training Parameters", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 103, + 341, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "Table S4 shows the training parameters for the Fixed Expansion Layer (FEL). The number of units", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 351, + 220, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 220, + 365 + ], + "score": 1.0, + "content": "in the FEL layer is given by", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 340, + 505, + 365 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 362, + 358, + 388 + ], + "lines": [ + { + "bbox": [ + 253, + 362, + 358, + 388 + ], + "spans": [ + { + "bbox": [ + 253, + 362, + 358, + 388 + ], + "score": 0.93, + "content": "\\mathrm { F E L ~ U n i t s } = { \\frac { H ^ { 2 } + H K } { K } }", + "type": "interline_equation", + "image_path": "5e73f5685a7868c3d14e025b929b9a9728b8e4a8c195152ec256afaeeefb4189.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 253, + 362, + 358, + 388 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 133, + 402 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 390, + 144, + 399 + ], + "score": 0.79, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 389, + 351, + 402 + ], + "score": 1.0, + "content": "is the number of units in the first hidden-layer and", + "type": "text" + }, + { + "bbox": [ + 351, + 390, + 361, + 399 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "is the maximum number of classes", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 400, + 500, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 500, + 414 + ], + "score": 1.0, + "content": "in the dataset. The values that are given as a range of values are the hyperparameter search spaces.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 389, + 505, + 414 + ] + }, + { + "type": "table", + "bbox": [ + 223, + 419, + 387, + 492 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 223, + 419, + 387, + 492 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 223, + 419, + 387, + 492 + ], + "spans": [ + { + "bbox": [ + 223, + 419, + 387, + 492 + ], + "score": 0.977, + "html": "
HyperparameterValues
Hidden Layer Size (H)64-1800
FEL Layer Size Number of Hidden LayersSee Equation 6
Mini-Batch Size2 8
Initial Learning Rate10-2
", + "type": "table", + "image_path": "39b7f0e874056c8831308643934f5229027de7d5c569b5e0a1430cfe73e11b2b.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 223, + 419, + 387, + 433.6 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 223, + 433.6, + 387, + 448.20000000000005 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 223, + 448.20000000000005, + 387, + 462.80000000000007 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 223, + 462.80000000000007, + 387, + 477.4000000000001 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 223, + 477.4000000000001, + 387, + 492.0000000000001 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 234, + 501, + 378, + 513 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 233, + 500, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 233, + 500, + 378, + 514 + ], + "score": 1.0, + "content": "Table S4: FEL Training Parameters", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 343, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 343, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 343, + 535 + ], + "score": 1.0, + "content": "A.2 ICARL PERFORMANCE WITH MORE EXEMPLARS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "Table S5 provides additional experimental results for when there are more exemplars per class", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "(EPC) for the iCaRL framework. Rebuffi et al. (2017) used 20 EPC in their original paper; however,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "we increased the number to 100 EPC to see if storing more training data helped iCaRL. Although", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "a higher EPC does increase iCaRL performance, it still does not outperform FearNet. Note that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "CUB-200 only has about 30 training samples per class, so iCaRL is storing the entire training set", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 598, + 340, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 340, + 609 + ], + "score": 1.0, + "content": "for 100 EPC. Our main results use the default value of 20.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 543, + 506, + 609 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 633, + 202, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 204, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 204, + 646 + ], + "score": 1.0, + "content": "A.3 BLA VARIANTS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Our BLA model is a classifier that determines whether a prediction should be made using HC (recent", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "memory) or mPFC (remote memory). An alternative approach would be to use an outlier detection", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "algorithm that determines whether the data being processed by a sub-network is an outlier for that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "sub-network and should therefore be processed by the other sub-network. To explore this alternative", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "BLA formulation, we experimented with three outlier detection algorithms: 1) one-class support", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "vector machine (SVM) (Scholkopf et al., 2001), 2) determining if the data fits into a Gaussian dis- ¨", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "tribution using a minimum covariance determinant estimation (i.e., elliptical envelope) (Rousseeuw", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 81, + 475, + 143 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 136, + 81, + 475, + 143 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 81, + 475, + 143 + ], + "spans": [ + { + "bbox": [ + 136, + 81, + 475, + 143 + ], + "score": 0.974, + "html": "
ModelCIFAR-100 Sbase Ωnew aulCUB-200 Sbase ΩnewΩallAudioSet Sbase ΩnewΩauMean Sbase
iCaRL (20 EPC)0.746 0.807 0.7490.9420.547 0.8640.740 0.4870.733Ωall [0.801
iCaRL (100 EPC)0.842 0.719 0.8220.9510.554 0.8820.820 0.4190.7710.782 0.871 0.825
FearNet0.927 0.824 0.9470.9240.598 0.8910.962 0.4550.9320.938 0.923
", + "type": "table", + "image_path": "931f35f0adc8f02ed0815e4c924a6397ad56e7fd9007ea857565546023ccfb8f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 81, + 475, + 101.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 101.66666666666667, + 475, + 122.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 122.33333333333334, + 475, + 143.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 141, + 151, + 469, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 151, + 470, + 163 + ], + "spans": [ + { + "bbox": [ + 141, + 151, + 470, + 163 + ], + "score": 1.0, + "content": "Table S5: iCaRL’s performance when the stored EPC is increased from 20 to 100.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 235, + 238, + 376, + 300 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 182, + 505, + 238 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 183, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 506, + 195 + ], + "score": 1.0, + "content": "& Driessen, 1999), and 3) the isolation forest (Liu et al., 2008). All three of these methods set a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "rejection criterion for if the test sample exists in HC; whereas the binary MLP reports a probability", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "on how likely the test sample resides in HC. Table S6 compares these individual methods. Isolation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "Forest and Elliptic Envelope seem to prefer the data in HC, one-class SVM prefers the data in mPFC,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 227, + 414, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 414, + 239 + ], + "score": 1.0, + "content": "and our binary MLP worked best at choosing the correct sub-network to use.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 235, + 238, + 376, + 300 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 235, + 238, + 376, + 300 + ], + "spans": [ + { + "bbox": [ + 235, + 238, + 376, + 300 + ], + "score": 0.968, + "html": "
BLA MethodΩbaseΩnewaul
Isolation ForestElliptic EnvelopeOne-Class SVMBinary MLP0.3280.8230.368
elope0.5180.8230.541
0.7180.4330.702
0.9270.9240.947
", + "type": "table", + "image_path": "917a5d6e0c21cc99fe67169091b535f9d81fbd545a14450e78c3cf28f186496e.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 238, + 376, + 269.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 235, + 269.0, + 376, + 300.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 205, + 308, + 405, + 319 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 205, + 307, + 406, + 320 + ], + "spans": [ + { + "bbox": [ + 205, + 307, + 406, + 320 + ], + "score": 1.0, + "content": "Table S6: Performance of different BLA variants.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 326, + 232, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 233, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 233, + 338 + ], + "score": 1.0, + "content": "A.4 FEARNET ALGORITHM", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 361 + ], + "score": 1.0, + "content": "Pseudocode for FearNet’s training and prediction algorithms are given in Algorithms 1 and 2 re-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 357, + 354, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 354, + 371 + ], + "score": 1.0, + "content": "spectively. The variables match the ones defined in the paper.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "table", + "bbox": [ + 106, + 385, + 487, + 559 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 385, + 487, + 559 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 385, + 487, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 487, + 559 + ], + "score": 0.969, + "html": "
Algorithm1: FearNet TrainingAlgorithm 2: FearNet Prediction
Data: X,y Classes/Study-Sessions: T; K: Sleep Frequency; Initialize mPFC with base-knowledge;Data: X A(X) ← PBLA (C =1|X); ← maxk PHc(C=k|X)A(X). 1-A(X)
Store μt,Σt for each class in the base-knowledge; forc←T/2toTdo StoreX,y for class c in HC;if > maxk PmPFc (C = k|X) then return arg maxk PHc (C = k|X);
if c%K==O then Fine-tune mPFC with X,y in HC and pseudo- examples generated by mPFC decoder; Update μt,∑t for all classes seen so far;else return arg maXk PmPFc (C = k/|X);
", + "type": "table", + "image_path": "bc5035b02c2dc885b6dca89cf08ce7ffb3bc1e76ab72b5cc613f9ff7f1c8f622.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 106, + 385, + 487, + 443.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 443.0, + 487, + 501.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 501.0, + 487, + 559.0 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 575, + 308, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 308, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 308, + 588 + ], + "score": 1.0, + "content": "A.5 MULTI-MODAL LEARNING EXPERIMENT", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Fig. S1 shows the plots for the multi-modal experiments in Sec. 6.2. The three base-knowledge ex-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 607, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 104, + 607, + 505, + 621 + ], + "score": 1.0, + "content": "periments were 1) CIFAR-100 is the base-knowledge and AudioSet is trained incrementally, 2) Au-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "dioSet is the base-knowledge and then AudioSet is trained incrementally, and 3) the base-knowledge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 123, + 642 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 123, + 630, + 165, + 640 + ], + "score": 0.42, + "content": "5 0 / 5 0 ~ \\mathrm { m i x }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "of the two datasets and then the remaining classes are trained incrementally. For all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "three base-knowledge experiments, we show the mean-class accuracy on the base-knowledge and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "the entire test set. FearNet works well when it adequately learns the base-knowledge (Experiment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 662, + 452, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 452, + 675 + ], + "score": 1.0, + "content": "#1 and #3); however, when FearNet learns it poorly, incremental learning deteriorates.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 107, + 688, + 336, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 337, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 337, + 701 + ], + "score": 1.0, + "content": "A.6 BASE-KNOWLEDGE EFFECT ON PERFORMANCE", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 504, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 479, + 724 + ], + "score": 1.0, + "content": "Fig. S2 shows the effect of the base-knowledge’s size on FearNet’s performance. As expected,", + "type": "text" + }, + { + "bbox": [ + 480, + 710, + 504, + 721 + ], + "score": 0.84, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 718, + 504, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 480, + 735 + ], + "score": 1.0, + "content": "increases because there are not as many sleep phases to overwrite existing base-knowledge.", + "type": "text" + }, + { + "bbox": [ + 480, + 721, + 504, + 732 + ], + "score": 0.87, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + } + ], + "index": 28 + } + ], + "index": 27.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 81, + 475, + 143 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 136, + 81, + 475, + 143 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 81, + 475, + 143 + ], + "spans": [ + { + "bbox": [ + 136, + 81, + 475, + 143 + ], + "score": 0.974, + "html": "
ModelCIFAR-100 Sbase Ωnew aulCUB-200 Sbase ΩnewΩallAudioSet Sbase ΩnewΩauMean Sbase
iCaRL (20 EPC)0.746 0.807 0.7490.9420.547 0.8640.740 0.4870.733Ωall [0.801
iCaRL (100 EPC)0.842 0.719 0.8220.9510.554 0.8820.820 0.4190.7710.782 0.871 0.825
FearNet0.927 0.824 0.9470.9240.598 0.8910.962 0.4550.9320.938 0.923
", + "type": "table", + "image_path": "931f35f0adc8f02ed0815e4c924a6397ad56e7fd9007ea857565546023ccfb8f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 81, + 475, + 101.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 101.66666666666667, + 475, + 122.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 122.33333333333334, + 475, + 143.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 141, + 151, + 469, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 151, + 470, + 163 + ], + "spans": [ + { + "bbox": [ + 141, + 151, + 470, + 163 + ], + "score": 1.0, + "content": "Table S5: iCaRL’s performance when the stored EPC is increased from 20 to 100.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 235, + 238, + 376, + 300 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 182, + 505, + 238 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 183, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 506, + 195 + ], + "score": 1.0, + "content": "& Driessen, 1999), and 3) the isolation forest (Liu et al., 2008). All three of these methods set a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "rejection criterion for if the test sample exists in HC; whereas the binary MLP reports a probability", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "on how likely the test sample resides in HC. Table S6 compares these individual methods. Isolation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "Forest and Elliptic Envelope seem to prefer the data in HC, one-class SVM prefers the data in mPFC,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 227, + 414, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 414, + 239 + ], + "score": 1.0, + "content": "and our binary MLP worked best at choosing the correct sub-network to use.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 235, + 238, + 376, + 300 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 235, + 238, + 376, + 300 + ], + "spans": [ + { + "bbox": [ + 235, + 238, + 376, + 300 + ], + "score": 0.968, + "html": "
BLA MethodΩbaseΩnewaul
Isolation ForestElliptic EnvelopeOne-Class SVMBinary MLP0.3280.8230.368
elope0.5180.8230.541
0.7180.4330.702
0.9270.9240.947
", + "type": "table", + "image_path": "917a5d6e0c21cc99fe67169091b535f9d81fbd545a14450e78c3cf28f186496e.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 238, + 376, + 269.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 235, + 269.0, + 376, + 300.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 205, + 308, + 405, + 319 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 205, + 307, + 406, + 320 + ], + "spans": [ + { + "bbox": [ + 205, + 307, + 406, + 320 + ], + "score": 1.0, + "content": "Table S6: Performance of different BLA variants.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 326, + 232, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 233, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 233, + 338 + ], + "score": 1.0, + "content": "A.4 FEARNET ALGORITHM", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 361 + ], + "score": 1.0, + "content": "Pseudocode for FearNet’s training and prediction algorithms are given in Algorithms 1 and 2 re-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 357, + 354, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 354, + 371 + ], + "score": 1.0, + "content": "spectively. The variables match the ones defined in the paper.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 346, + 505, + 371 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 385, + 487, + 559 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 385, + 487, + 559 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 385, + 487, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 487, + 559 + ], + "score": 0.969, + "html": "
Algorithm1: FearNet TrainingAlgorithm 2: FearNet Prediction
Data: X,y Classes/Study-Sessions: T; K: Sleep Frequency; Initialize mPFC with base-knowledge;Data: X A(X) ← PBLA (C =1|X); ← maxk PHc(C=k|X)A(X). 1-A(X)
Store μt,Σt for each class in the base-knowledge; forc←T/2toTdo StoreX,y for class c in HC;if > maxk PmPFc (C = k|X) then return arg maxk PHc (C = k|X);
if c%K==O then Fine-tune mPFC with X,y in HC and pseudo- examples generated by mPFC decoder; Update μt,∑t for all classes seen so far;else return arg maXk PmPFc (C = k/|X);
", + "type": "table", + "image_path": "bc5035b02c2dc885b6dca89cf08ce7ffb3bc1e76ab72b5cc613f9ff7f1c8f622.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 106, + 385, + 487, + 443.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 443.0, + 487, + 501.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 501.0, + 487, + 559.0 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 575, + 308, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 308, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 308, + 588 + ], + "score": 1.0, + "content": "A.5 MULTI-MODAL LEARNING EXPERIMENT", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Fig. S1 shows the plots for the multi-modal experiments in Sec. 6.2. The three base-knowledge ex-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 607, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 104, + 607, + 505, + 621 + ], + "score": 1.0, + "content": "periments were 1) CIFAR-100 is the base-knowledge and AudioSet is trained incrementally, 2) Au-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "dioSet is the base-knowledge and then AudioSet is trained incrementally, and 3) the base-knowledge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 123, + 642 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 123, + 630, + 165, + 640 + ], + "score": 0.42, + "content": "5 0 / 5 0 ~ \\mathrm { m i x }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "of the two datasets and then the remaining classes are trained incrementally. For all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "three base-knowledge experiments, we show the mean-class accuracy on the base-knowledge and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "the entire test set. FearNet works well when it adequately learns the base-knowledge (Experiment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 662, + 452, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 452, + 675 + ], + "score": 1.0, + "content": "#1 and #3); however, when FearNet learns it poorly, incremental learning deteriorates.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 597, + 505, + 675 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 688, + 336, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 337, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 337, + 701 + ], + "score": 1.0, + "content": "A.6 BASE-KNOWLEDGE EFFECT ON PERFORMANCE", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 504, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 479, + 724 + ], + "score": 1.0, + "content": "Fig. S2 shows the effect of the base-knowledge’s size on FearNet’s performance. As expected,", + "type": "text" + }, + { + "bbox": [ + 480, + 710, + 504, + 721 + ], + "score": 0.84, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 718, + 504, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 480, + 735 + ], + "score": 1.0, + "content": "increases because there are not as many sleep phases to overwrite existing base-knowledge.", + "type": "text" + }, + { + "bbox": [ + 480, + 721, + 504, + 732 + ], + "score": 0.87, + "content": "\\Omega _ { n e w }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "remains relatively even because the size of the base-knowledge has no effect on the HC model’s", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "score": 1.0, + "content": "ability to immediately recall new information; however, there is a very slight decrease that corre-", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 466, + 672 + ], + "score": 1.0, + "content": "sponds to the BLA model erroneously favoring mPFC in a few cases. Most importantly,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 466, + 659, + 484, + 670 + ], + "score": 0.89, + "content": "\\Omega _ { a l l }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 485, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "sees", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 271, + 683 + ], + "score": 1.0, + "content": "an increase in performance because; like", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 272, + 670, + 295, + 681 + ], + "score": 0.91, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 296, + 669, + 505, + 683 + ], + "score": 1.0, + "content": ", there are not as many sleep phases to perturb older", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 681, + 188, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 188, + 692 + ], + "score": 1.0, + "content": "memories in mPFC.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 707, + 504, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 79, + 484, + 553 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 79, + 484, + 553 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 79, + 484, + 553 + ], + "spans": [ + { + "bbox": [ + 122, + 79, + 484, + 553 + ], + "score": 0.978, + "type": "image", + "image_path": "1ee9cfc995cd1eecefdcd49c0badb1e9c5563f08021c359856092693a85d6752.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 79, + 484, + 237.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 237.0, + 484, + 395.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 395.0, + 484, + 553.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 560, + 505, + 616 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Figure S1: Detailed plots for the multi-modal experiment. The top row is when the base-knowledge", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "was CIFAR-100, the middle row is when the base-knowledge was AudioSet, and the bottom row", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 251, + 596 + ], + "score": 1.0, + "content": "is when the base-knowledge was a", + "type": "text" + }, + { + "bbox": [ + 252, + 583, + 295, + 594 + ], + "score": 0.31, + "content": "5 0 / 5 0 ~ \\mathrm { m i x }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "from the two datasets. The left column represents", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "the mean-class accuracy on the base-knowledge test set and the right column computes mean-class", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 605, + 228, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 228, + 618 + ], + "score": 1.0, + "content": "accuracy on the entire test set.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "remains relatively even because the size of the base-knowledge has no effect on the HC model’s", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "score": 1.0, + "content": "ability to immediately recall new information; however, there is a very slight decrease that corre-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 466, + 672 + ], + "score": 1.0, + "content": "sponds to the BLA model erroneously favoring mPFC in a few cases. Most importantly,", + "type": "text" + }, + { + "bbox": [ + 466, + 659, + 484, + 670 + ], + "score": 0.89, + "content": "\\Omega _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "sees", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 271, + 683 + ], + "score": 1.0, + "content": "an increase in performance because; like", + "type": "text" + }, + { + "bbox": [ + 272, + 670, + 295, + 681 + ], + "score": 0.91, + "content": "\\Omega _ { b a s e }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 669, + 505, + 683 + ], + "score": 1.0, + "content": ", there are not as many sleep phases to perturb older", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 681, + 188, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 188, + 692 + ], + "score": 1.0, + "content": "memories in mPFC.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 79, + 484, + 553 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 79, + 484, + 553 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 79, + 484, + 553 + ], + "spans": [ + { + "bbox": [ + 122, + 79, + 484, + 553 + ], + "score": 0.978, + "type": "image", + "image_path": "1ee9cfc995cd1eecefdcd49c0badb1e9c5563f08021c359856092693a85d6752.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 79, + 484, + 237.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 237.0, + 484, + 395.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 395.0, + 484, + 553.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 560, + 505, + 616 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Figure S1: Detailed plots for the multi-modal experiment. The top row is when the base-knowledge", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "was CIFAR-100, the middle row is when the base-knowledge was AudioSet, and the bottom row", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 251, + 596 + ], + "score": 1.0, + "content": "is when the base-knowledge was a", + "type": "text" + }, + { + "bbox": [ + 252, + 583, + 295, + 594 + ], + "score": 0.31, + "content": "5 0 / 5 0 ~ \\mathrm { m i x }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "from the two datasets. The left column represents", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "the mean-class accuracy on the base-knowledge test set and the right column computes mean-class", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 605, + 228, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 228, + 618 + ], + "score": 1.0, + "content": "accuracy on the entire test set.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 505, + 692 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 105, + 637, + 505, + 692 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 206, + 320, + 401, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 206, + 320, + 401, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 320, + 401, + 469 + ], + "spans": [ + { + "bbox": [ + 206, + 320, + 401, + 469 + ], + "score": 0.963, + "type": "image", + "image_path": "e8820c9ee2b911f01bafc12161f3da93d222287445581a28a5c6c5c91bd828c6.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 320, + 401, + 338.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 206, + 338.625, + 401, + 357.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 206, + 357.25, + 401, + 375.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 206, + 375.875, + 401, + 394.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 206, + 394.5, + 401, + 413.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 206, + 413.125, + 401, + 431.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 206, + 431.75, + 401, + 450.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 206, + 450.375, + 401, + 469.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 163, + 480, + 447, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 162, + 478, + 448, + 495 + ], + "spans": [ + { + "bbox": [ + 162, + 478, + 448, + 495 + ], + "score": 1.0, + "content": "Figure S2: FearNet performance as a function of base-knowledge size.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 206, + 320, + 401, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 206, + 320, + 401, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 320, + 401, + 469 + ], + "spans": [ + { + "bbox": [ + 206, + 320, + 401, + 469 + ], + "score": 0.963, + "type": "image", + "image_path": "e8820c9ee2b911f01bafc12161f3da93d222287445581a28a5c6c5c91bd828c6.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 320, + 401, + 338.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 206, + 338.625, + 401, + 357.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 206, + 357.25, + 401, + 375.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 206, + 375.875, + 401, + 394.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 206, + 394.5, + 401, + 413.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 206, + 413.125, + 401, + 431.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 206, + 431.75, + 401, + 450.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 206, + 450.375, + 401, + 469.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 163, + 480, + 447, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 162, + 478, + 448, + 495 + ], + "spans": [ + { + "bbox": [ + 162, + 478, + 448, + 495 + ], + "score": 1.0, + "content": "Figure S2: FearNet performance as a function of base-knowledge size.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file