{
"pdf_info": [
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
78,
504,
116
],
"lines": [
{
"bbox": [
106,
78,
505,
98
],
"spans": [
{
"bbox": [
106,
78,
505,
98
],
"score": 1.0,
"content": "SLASH: EMBRACING PROBABILISTIC CIRCUITS INTO",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
98,
396,
118
],
"spans": [
{
"bbox": [
105,
98,
396,
118
],
"score": 1.0,
"content": "NEURAL ANSWER SET PROGRAMMING",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
112,
135,
244,
157
],
"lines": [
{
"bbox": [
113,
135,
201,
147
],
"spans": [
{
"bbox": [
113,
135,
201,
147
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
111,
146,
245,
158
],
"spans": [
{
"bbox": [
111,
146,
245,
158
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5
},
{
"type": "title",
"bbox": [
278,
186,
333,
199
],
"lines": [
{
"bbox": [
276,
185,
336,
200
],
"spans": [
{
"bbox": [
276,
185,
336,
200
],
"score": 1.0,
"content": "ABSTRACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
142,
212,
469,
376
],
"lines": [
{
"bbox": [
141,
211,
470,
225
],
"spans": [
{
"bbox": [
141,
211,
470,
225
],
"score": 1.0,
"content": "The goal of combining the robustness of neural networks and the expressivity of",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
223,
471,
235
],
"spans": [
{
"bbox": [
141,
223,
471,
235
],
"score": 1.0,
"content": "symbolic methods has rekindled the interest in Neuro-Symbolic AI. Recent ad-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
141,
234,
469,
245
],
"spans": [
{
"bbox": [
141,
234,
469,
245
],
"score": 1.0,
"content": "vancements in Neuro-Symbolic AI often consider specifically-tailored architectures",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
141,
244,
470,
257
],
"spans": [
{
"bbox": [
141,
244,
470,
257
],
"score": 1.0,
"content": "consisting of disjoint neural and symbolic components, and thus do not exhibit",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
255,
470,
268
],
"spans": [
{
"bbox": [
141,
255,
470,
268
],
"score": 1.0,
"content": "desired gains that can be achieved by integrating them into a unifying framework.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
266,
471,
280
],
"spans": [
{
"bbox": [
141,
266,
471,
280
],
"score": 1.0,
"content": "We introduce SLASH – a novel deep probabilistic programming language (DPPL).",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
141,
277,
470,
290
],
"spans": [
{
"bbox": [
141,
277,
470,
290
],
"score": 1.0,
"content": "At its core, SLASH consists of Neural-Probabilistic Predicates (NPPs) and logical",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
140,
289,
470,
301
],
"spans": [
{
"bbox": [
140,
289,
470,
301
],
"score": 1.0,
"content": "programs which are united via answer set programming. The probability estimates",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
141,
300,
470,
312
],
"spans": [
{
"bbox": [
141,
300,
470,
312
],
"score": 1.0,
"content": "resulting from NPPs act as the binding element between the logical program and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
141,
311,
470,
322
],
"spans": [
{
"bbox": [
141,
311,
470,
322
],
"score": 1.0,
"content": "raw input data, thereby allowing SLASH to answer task-dependent logical queries.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
321,
470,
334
],
"spans": [
{
"bbox": [
141,
321,
470,
334
],
"score": 1.0,
"content": "This allows SLASH to elegantly integrate the symbolic and neural components",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
333,
470,
344
],
"spans": [
{
"bbox": [
141,
333,
470,
344
],
"score": 1.0,
"content": "in a unified framework. We evaluate SLASH on the benchmark data of MNIST",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
141,
343,
469,
356
],
"spans": [
{
"bbox": [
141,
343,
469,
356
],
"score": 1.0,
"content": "addition as well as novel tasks for DPPLs such as missing data prediction and set",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
354,
469,
367
],
"spans": [
{
"bbox": [
141,
354,
469,
367
],
"score": 1.0,
"content": "prediction with state-of-the-art performance, thereby showing the effectiveness and",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
366,
246,
377
],
"spans": [
{
"bbox": [
141,
366,
246,
377
],
"score": 1.0,
"content": "generality of our method.",
"type": "text"
}
],
"index": 19
}
],
"index": 12
},
{
"type": "title",
"bbox": [
108,
398,
206,
411
],
"lines": [
{
"bbox": [
105,
397,
208,
414
],
"spans": [
{
"bbox": [
105,
397,
208,
414
],
"score": 1.0,
"content": "1 INTRODUCTION",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
107,
424,
505,
501
],
"lines": [
{
"bbox": [
105,
423,
506,
437
],
"spans": [
{
"bbox": [
105,
423,
506,
437
],
"score": 1.0,
"content": "In recent years, Neuro-Symbolic AI approaches to learning (Hudson & Manning, 2019; d’Avila",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
435,
505,
447
],
"spans": [
{
"bbox": [
106,
435,
505,
447
],
"score": 1.0,
"content": "Garcez et al., 2019; Jiang & Ahn, 2020; d’Avila Garcez & Lamb, 2020), which integrates low-level",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
445,
505,
459
],
"spans": [
{
"bbox": [
105,
445,
505,
459
],
"score": 1.0,
"content": "perception with high-level reasoning by combining data-driven neural modules with logic-based",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
456,
506,
470
],
"spans": [
{
"bbox": [
105,
456,
506,
470
],
"score": 1.0,
"content": "symbolic modules, has gained traction. This combination of sub-symbolic and symbolic systems",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
466,
505,
481
],
"spans": [
{
"bbox": [
105,
466,
505,
481
],
"score": 1.0,
"content": "has been shown to have several advantages for various tasks such as visual question answering",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
478,
506,
492
],
"spans": [
{
"bbox": [
105,
478,
506,
492
],
"score": 1.0,
"content": "and reasoning (Yi et al., 2018), concept learning (Mao et al., 2019) and improved properties for",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
489,
428,
502
],
"spans": [
{
"bbox": [
105,
489,
428,
502
],
"score": 1.0,
"content": "explainable and revisable models (Ciravegna et al., 2020; Stammer et al., 2021).",
"type": "text"
}
],
"index": 27
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
506,
505,
605
],
"lines": [
{
"bbox": [
106,
506,
505,
518
],
"spans": [
{
"bbox": [
106,
506,
505,
518
],
"score": 1.0,
"content": "Rather than designing specifically tailored Neuro-Symbolic architectures, where often the neural and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
517,
506,
530
],
"spans": [
{
"bbox": [
105,
517,
506,
530
],
"score": 1.0,
"content": "symbolic modules are disjoint and trained independently (Yi et al., 2018; Mao et al., 2019; Stammer",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
528,
505,
541
],
"spans": [
{
"bbox": [
105,
528,
505,
541
],
"score": 1.0,
"content": "et al., 2021), deep probabilistic programming languages (DPPLs) provide an exciting alternative",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
539,
505,
552
],
"spans": [
{
"bbox": [
105,
539,
505,
552
],
"score": 1.0,
"content": "(Bingham et al., 2019; Tran et al., 2017; Manhaeve et al., 2018; Yang et al., 2020). Specifically, DPPLs",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
550,
505,
564
],
"spans": [
{
"bbox": [
105,
550,
505,
564
],
"score": 1.0,
"content": "integrate neural and symbolic modules via a unifying programming framework with probability",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
561,
505,
574
],
"spans": [
{
"bbox": [
105,
561,
505,
574
],
"score": 1.0,
"content": "estimates acting as the “glue” between separate modules allowing for reasoning over noisy, uncertain",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "data and, importantly, joint training of the modules. Additionally, prior knowledge and biases in",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
582,
505,
596
],
"spans": [
{
"bbox": [
105,
582,
505,
596
],
"score": 1.0,
"content": "the form of logical rules can easily be added with DPPLs, rather than creating implicit architectural",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
594,
444,
607
],
"spans": [
{
"bbox": [
105,
594,
444,
607
],
"score": 1.0,
"content": "biases, thereby integrating neural networks into downstream logical reasoning tasks.",
"type": "text"
}
],
"index": 36
}
],
"index": 32
},
{
"type": "text",
"bbox": [
107,
611,
505,
732
],
"lines": [
{
"bbox": [
105,
610,
506,
624
],
"spans": [
{
"bbox": [
105,
610,
506,
624
],
"score": 1.0,
"content": "Object-centric deep learning has recently brought forth several exciting avenues of research by intro-",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
622,
506,
634
],
"spans": [
{
"bbox": [
106,
622,
506,
634
],
"score": 1.0,
"content": "ducing inductive biases to neural networks to extract objects from visual scenes in an unsupervised",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
632,
506,
645
],
"spans": [
{
"bbox": [
105,
632,
506,
645
],
"score": 1.0,
"content": "manner (Zhang et al., 2019; Burgess et al., 2019; Engelcke et al., 2020; Greff et al., 2019; Lin et al.,",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
644,
506,
656
],
"spans": [
{
"bbox": [
106,
644,
506,
656
],
"score": 1.0,
"content": "2020; Locatello et al., 2020; Jiang & Ahn, 2020). We refer to Greff et al. (2020) for a detailed",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
654,
505,
667
],
"spans": [
{
"bbox": [
105,
654,
505,
667
],
"score": 1.0,
"content": "overview. A motivation for this specific line of investigation, which notably has been around for",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
666,
506,
678
],
"spans": [
{
"bbox": [
105,
666,
506,
678
],
"score": 1.0,
"content": "a longer period of time (Fodor & Pylyshyn, 1988; Marcus, 2019), is that objects occur as natural",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
677,
506,
690
],
"spans": [
{
"bbox": [
106,
677,
506,
690
],
"score": 1.0,
"content": "building blocks in human perception and possess advantageous properties for many cognitive tasks,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
688,
506,
700
],
"spans": [
{
"bbox": [
105,
688,
506,
700
],
"score": 1.0,
"content": "such as scene understanding and reasoning. With a DPPL, these advancements can be improved",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
699,
505,
711
],
"spans": [
{
"bbox": [
106,
699,
505,
711
],
"score": 1.0,
"content": "by integrating the previously mentioned components into the DPPL’s programming framework and",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
709,
506,
723
],
"spans": [
{
"bbox": [
105,
709,
506,
723
],
"score": 1.0,
"content": "further adding constraints about objects and their properties in form of logical statements e.g. about",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
720,
412,
734
],
"spans": [
{
"bbox": [
105,
720,
412,
734
],
"score": 1.0,
"content": "color singularity, rather than implicitly enforcing this via one hot encodings.",
"type": "text"
}
],
"index": 47
}
],
"index": 42
}
],
"page_idx": 0,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
760
],
"lines": [
{
"bbox": [
302,
751,
308,
761
],
"spans": [
{
"bbox": [
302,
751,
308,
761
],
"score": 1.0,
"content": "1",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
78,
504,
116
],
"lines": [
{
"bbox": [
106,
78,
505,
98
],
"spans": [
{
"bbox": [
106,
78,
505,
98
],
"score": 1.0,
"content": "SLASH: EMBRACING PROBABILISTIC CIRCUITS INTO",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
98,
396,
118
],
"spans": [
{
"bbox": [
105,
98,
396,
118
],
"score": 1.0,
"content": "NEURAL ANSWER SET PROGRAMMING",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
112,
135,
244,
157
],
"lines": [
{
"bbox": [
113,
135,
201,
147
],
"spans": [
{
"bbox": [
113,
135,
201,
147
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
111,
146,
245,
158
],
"spans": [
{
"bbox": [
111,
146,
245,
158
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5,
"bbox_fs": [
111,
135,
245,
158
]
},
{
"type": "title",
"bbox": [
278,
186,
333,
199
],
"lines": [
{
"bbox": [
276,
185,
336,
200
],
"spans": [
{
"bbox": [
276,
185,
336,
200
],
"score": 1.0,
"content": "ABSTRACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
142,
212,
469,
376
],
"lines": [
{
"bbox": [
141,
211,
470,
225
],
"spans": [
{
"bbox": [
141,
211,
470,
225
],
"score": 1.0,
"content": "The goal of combining the robustness of neural networks and the expressivity of",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
223,
471,
235
],
"spans": [
{
"bbox": [
141,
223,
471,
235
],
"score": 1.0,
"content": "symbolic methods has rekindled the interest in Neuro-Symbolic AI. Recent ad-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
141,
234,
469,
245
],
"spans": [
{
"bbox": [
141,
234,
469,
245
],
"score": 1.0,
"content": "vancements in Neuro-Symbolic AI often consider specifically-tailored architectures",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
141,
244,
470,
257
],
"spans": [
{
"bbox": [
141,
244,
470,
257
],
"score": 1.0,
"content": "consisting of disjoint neural and symbolic components, and thus do not exhibit",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
255,
470,
268
],
"spans": [
{
"bbox": [
141,
255,
470,
268
],
"score": 1.0,
"content": "desired gains that can be achieved by integrating them into a unifying framework.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
266,
471,
280
],
"spans": [
{
"bbox": [
141,
266,
471,
280
],
"score": 1.0,
"content": "We introduce SLASH – a novel deep probabilistic programming language (DPPL).",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
141,
277,
470,
290
],
"spans": [
{
"bbox": [
141,
277,
470,
290
],
"score": 1.0,
"content": "At its core, SLASH consists of Neural-Probabilistic Predicates (NPPs) and logical",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
140,
289,
470,
301
],
"spans": [
{
"bbox": [
140,
289,
470,
301
],
"score": 1.0,
"content": "programs which are united via answer set programming. The probability estimates",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
141,
300,
470,
312
],
"spans": [
{
"bbox": [
141,
300,
470,
312
],
"score": 1.0,
"content": "resulting from NPPs act as the binding element between the logical program and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
141,
311,
470,
322
],
"spans": [
{
"bbox": [
141,
311,
470,
322
],
"score": 1.0,
"content": "raw input data, thereby allowing SLASH to answer task-dependent logical queries.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
321,
470,
334
],
"spans": [
{
"bbox": [
141,
321,
470,
334
],
"score": 1.0,
"content": "This allows SLASH to elegantly integrate the symbolic and neural components",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
333,
470,
344
],
"spans": [
{
"bbox": [
141,
333,
470,
344
],
"score": 1.0,
"content": "in a unified framework. We evaluate SLASH on the benchmark data of MNIST",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
141,
343,
469,
356
],
"spans": [
{
"bbox": [
141,
343,
469,
356
],
"score": 1.0,
"content": "addition as well as novel tasks for DPPLs such as missing data prediction and set",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
354,
469,
367
],
"spans": [
{
"bbox": [
141,
354,
469,
367
],
"score": 1.0,
"content": "prediction with state-of-the-art performance, thereby showing the effectiveness and",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
366,
246,
377
],
"spans": [
{
"bbox": [
141,
366,
246,
377
],
"score": 1.0,
"content": "generality of our method.",
"type": "text"
}
],
"index": 19
}
],
"index": 12,
"bbox_fs": [
140,
211,
471,
377
]
},
{
"type": "title",
"bbox": [
108,
398,
206,
411
],
"lines": [
{
"bbox": [
105,
397,
208,
414
],
"spans": [
{
"bbox": [
105,
397,
208,
414
],
"score": 1.0,
"content": "1 INTRODUCTION",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
107,
424,
505,
501
],
"lines": [
{
"bbox": [
105,
423,
506,
437
],
"spans": [
{
"bbox": [
105,
423,
506,
437
],
"score": 1.0,
"content": "In recent years, Neuro-Symbolic AI approaches to learning (Hudson & Manning, 2019; d’Avila",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
435,
505,
447
],
"spans": [
{
"bbox": [
106,
435,
505,
447
],
"score": 1.0,
"content": "Garcez et al., 2019; Jiang & Ahn, 2020; d’Avila Garcez & Lamb, 2020), which integrates low-level",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
445,
505,
459
],
"spans": [
{
"bbox": [
105,
445,
505,
459
],
"score": 1.0,
"content": "perception with high-level reasoning by combining data-driven neural modules with logic-based",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
456,
506,
470
],
"spans": [
{
"bbox": [
105,
456,
506,
470
],
"score": 1.0,
"content": "symbolic modules, has gained traction. This combination of sub-symbolic and symbolic systems",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
466,
505,
481
],
"spans": [
{
"bbox": [
105,
466,
505,
481
],
"score": 1.0,
"content": "has been shown to have several advantages for various tasks such as visual question answering",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
478,
506,
492
],
"spans": [
{
"bbox": [
105,
478,
506,
492
],
"score": 1.0,
"content": "and reasoning (Yi et al., 2018), concept learning (Mao et al., 2019) and improved properties for",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
489,
428,
502
],
"spans": [
{
"bbox": [
105,
489,
428,
502
],
"score": 1.0,
"content": "explainable and revisable models (Ciravegna et al., 2020; Stammer et al., 2021).",
"type": "text"
}
],
"index": 27
}
],
"index": 24,
"bbox_fs": [
105,
423,
506,
502
]
},
{
"type": "text",
"bbox": [
107,
506,
505,
605
],
"lines": [
{
"bbox": [
106,
506,
505,
518
],
"spans": [
{
"bbox": [
106,
506,
505,
518
],
"score": 1.0,
"content": "Rather than designing specifically tailored Neuro-Symbolic architectures, where often the neural and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
517,
506,
530
],
"spans": [
{
"bbox": [
105,
517,
506,
530
],
"score": 1.0,
"content": "symbolic modules are disjoint and trained independently (Yi et al., 2018; Mao et al., 2019; Stammer",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
528,
505,
541
],
"spans": [
{
"bbox": [
105,
528,
505,
541
],
"score": 1.0,
"content": "et al., 2021), deep probabilistic programming languages (DPPLs) provide an exciting alternative",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
539,
505,
552
],
"spans": [
{
"bbox": [
105,
539,
505,
552
],
"score": 1.0,
"content": "(Bingham et al., 2019; Tran et al., 2017; Manhaeve et al., 2018; Yang et al., 2020). Specifically, DPPLs",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
550,
505,
564
],
"spans": [
{
"bbox": [
105,
550,
505,
564
],
"score": 1.0,
"content": "integrate neural and symbolic modules via a unifying programming framework with probability",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
561,
505,
574
],
"spans": [
{
"bbox": [
105,
561,
505,
574
],
"score": 1.0,
"content": "estimates acting as the “glue” between separate modules allowing for reasoning over noisy, uncertain",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "data and, importantly, joint training of the modules. Additionally, prior knowledge and biases in",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
582,
505,
596
],
"spans": [
{
"bbox": [
105,
582,
505,
596
],
"score": 1.0,
"content": "the form of logical rules can easily be added with DPPLs, rather than creating implicit architectural",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
594,
444,
607
],
"spans": [
{
"bbox": [
105,
594,
444,
607
],
"score": 1.0,
"content": "biases, thereby integrating neural networks into downstream logical reasoning tasks.",
"type": "text"
}
],
"index": 36
}
],
"index": 32,
"bbox_fs": [
105,
506,
506,
607
]
},
{
"type": "text",
"bbox": [
107,
611,
505,
732
],
"lines": [
{
"bbox": [
105,
610,
506,
624
],
"spans": [
{
"bbox": [
105,
610,
506,
624
],
"score": 1.0,
"content": "Object-centric deep learning has recently brought forth several exciting avenues of research by intro-",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
622,
506,
634
],
"spans": [
{
"bbox": [
106,
622,
506,
634
],
"score": 1.0,
"content": "ducing inductive biases to neural networks to extract objects from visual scenes in an unsupervised",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
632,
506,
645
],
"spans": [
{
"bbox": [
105,
632,
506,
645
],
"score": 1.0,
"content": "manner (Zhang et al., 2019; Burgess et al., 2019; Engelcke et al., 2020; Greff et al., 2019; Lin et al.,",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
644,
506,
656
],
"spans": [
{
"bbox": [
106,
644,
506,
656
],
"score": 1.0,
"content": "2020; Locatello et al., 2020; Jiang & Ahn, 2020). We refer to Greff et al. (2020) for a detailed",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
654,
505,
667
],
"spans": [
{
"bbox": [
105,
654,
505,
667
],
"score": 1.0,
"content": "overview. A motivation for this specific line of investigation, which notably has been around for",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
666,
506,
678
],
"spans": [
{
"bbox": [
105,
666,
506,
678
],
"score": 1.0,
"content": "a longer period of time (Fodor & Pylyshyn, 1988; Marcus, 2019), is that objects occur as natural",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
677,
506,
690
],
"spans": [
{
"bbox": [
106,
677,
506,
690
],
"score": 1.0,
"content": "building blocks in human perception and possess advantageous properties for many cognitive tasks,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
688,
506,
700
],
"spans": [
{
"bbox": [
105,
688,
506,
700
],
"score": 1.0,
"content": "such as scene understanding and reasoning. With a DPPL, these advancements can be improved",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
699,
505,
711
],
"spans": [
{
"bbox": [
106,
699,
505,
711
],
"score": 1.0,
"content": "by integrating the previously mentioned components into the DPPL’s programming framework and",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
709,
506,
723
],
"spans": [
{
"bbox": [
105,
709,
506,
723
],
"score": 1.0,
"content": "further adding constraints about objects and their properties in form of logical statements e.g. about",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
720,
412,
734
],
"spans": [
{
"bbox": [
105,
720,
412,
734
],
"score": 1.0,
"content": "color singularity, rather than implicitly enforcing this via one hot encodings.",
"type": "text"
}
],
"index": 47
}
],
"index": 42,
"bbox_fs": [
105,
610,
506,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
113,
83,
496,
246
],
"blocks": [
{
"type": "image_body",
"bbox": [
113,
83,
496,
246
],
"group_id": 0,
"lines": [
{
"bbox": [
113,
83,
496,
246
],
"spans": [
{
"bbox": [
113,
83,
496,
246
],
"score": 0.967,
"type": "image",
"image_path": "fbc474f94d8a2574e5362c4b2e3c3d61cc7e2bb6baf0aac66d0dc2840ab2de17.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
113,
83,
496,
137.33333333333334
],
"spans": [],
"index": 0
},
{
"bbox": [
113,
137.33333333333334,
496,
191.66666666666669
],
"spans": [],
"index": 1
},
{
"bbox": [
113,
191.66666666666669,
496,
246.00000000000003
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
258,
505,
369
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
258,
505,
271
],
"spans": [
{
"bbox": [
106,
258,
505,
271
],
"score": 1.0,
"content": "Figure 1: SLASH Attention illustrated for a visual reasoning task. SLASH with Neural-Probabilistic",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
269,
506,
282
],
"spans": [
{
"bbox": [
105,
269,
506,
282
],
"score": 1.0,
"content": "Predicates consisting of a slot attention encoder and Probabilistic Circuits (PCs) realised via EiNets.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
280,
506,
293
],
"spans": [
{
"bbox": [
105,
280,
506,
293
],
"score": 1.0,
"content": "The slot encoder is shared over all NPPs. Each triangle in the figure represents a single EiNet that",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
291,
506,
304
],
"spans": [
{
"bbox": [
105,
291,
506,
304
],
"score": 1.0,
"content": "gives us a joint distribution at the root node. Thus, each PC learns the joint distribution over slot",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
302,
506,
316
],
"spans": [
{
"bbox": [
105,
302,
153,
316
],
"score": 1.0,
"content": "encodings,",
"type": "text"
},
{
"bbox": [
153,
302,
163,
313
],
"score": 0.84,
"content": "z ^ { i }",
"type": "inline_equation"
},
{
"bbox": [
163,
302,
257,
316
],
"score": 1.0,
"content": ", and object attributes,",
"type": "text"
},
{
"bbox": [
257,
303,
266,
313
],
"score": 0.76,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
266,
302,
506,
316
],
"score": 1.0,
"content": ", of a specific category, e.g. color attributes. Via targeted",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
313,
506,
327
],
"spans": [
{
"bbox": [
105,
313,
506,
327
],
"score": 1.0,
"content": "queries to the NPPs, one can obtain task-related probabilities, e.g. conditional probabilities for the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
323,
507,
338
],
"spans": [
{
"bbox": [
105,
323,
507,
338
],
"score": 1.0,
"content": "task of set prediction. Given the probability estimates from the NPP(s) and a SLASH program,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
335,
506,
348
],
"spans": [
{
"bbox": [
105,
335,
506,
348
],
"score": 1.0,
"content": "containing a set of facts and logical statements about the world, the probability of the truth value of a",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
346,
506,
359
],
"spans": [
{
"bbox": [
106,
346,
506,
359
],
"score": 1.0,
"content": "task-related query are computed via answer set programming. The entire system, including the neural",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
357,
437,
370
],
"spans": [
{
"bbox": [
106,
357,
437,
370
],
"score": 1.0,
"content": "and probabilistic modules, are finally trained end-to-end via a single loss function.",
"type": "text"
}
],
"index": 12
}
],
"index": 7.5
}
],
"index": 4.25
},
{
"type": "text",
"bbox": [
107,
380,
505,
567
],
"lines": [
{
"bbox": [
106,
380,
505,
393
],
"spans": [
{
"bbox": [
106,
380,
505,
393
],
"score": 1.0,
"content": "We propose SLASH – a novel DPPL that, similar to the punctuation symbol, can be used to efficiently",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
390,
505,
404
],
"spans": [
{
"bbox": [
105,
390,
505,
404
],
"score": 1.0,
"content": "combine several paradigms into one. Specifically, SLASH represents a scalable programming",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
402,
505,
415
],
"spans": [
{
"bbox": [
106,
402,
505,
415
],
"score": 1.0,
"content": "language that seamlessly integrates probabilistic logical programming with neural representations",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
412,
505,
425
],
"spans": [
{
"bbox": [
106,
412,
505,
425
],
"score": 1.0,
"content": "and tractable probabilistic estimations. Fig. 1 shows an example instantiation of SLASH, termed",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
424,
506,
437
],
"spans": [
{
"bbox": [
106,
424,
506,
437
],
"score": 1.0,
"content": "SLASH Attention, for object-centric set prediction. SLASH consists of several key building blocks.",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
434,
506,
447
],
"spans": [
{
"bbox": [
106,
434,
506,
447
],
"score": 1.0,
"content": "Firstly, it makes use of Neural-Probabilistic Predicates (NPPs) for probability estimation. NPPs",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
104,
444,
506,
460
],
"spans": [
{
"bbox": [
104,
444,
506,
460
],
"score": 1.0,
"content": "consist of neural and/or probabilistic circuit (PC) modules and act as a unifying term, encompassing",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
455,
506,
470
],
"spans": [
{
"bbox": [
105,
455,
506,
470
],
"score": 1.0,
"content": "the neural predicates of DeepProbLog and NeurASP, as well as purely probabilistic predicates. In",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
467,
505,
480
],
"spans": [
{
"bbox": [
106,
467,
505,
480
],
"score": 1.0,
"content": "this work, we introduce a much more powerful “flavor” of NPPs that consist jointly of neural and PC",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
478,
506,
492
],
"spans": [
{
"bbox": [
105,
478,
506,
492
],
"score": 1.0,
"content": "modules, taking advantage of the power of neural computations together with true density estimation",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
487,
507,
504
],
"spans": [
{
"bbox": [
105,
487,
507,
504
],
"score": 1.0,
"content": "of PCs. Depending on the underlying task one can thus ask a range of queries to the NPP, e.g.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
501,
505,
513
],
"spans": [
{
"bbox": [
105,
501,
505,
513
],
"score": 1.0,
"content": "sample an unknown, desired variable, but also query for conditional class probabilities. Example",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
511,
506,
524
],
"spans": [
{
"bbox": [
105,
511,
506,
524
],
"score": 1.0,
"content": "NPPs consisting of a slot attention encoder and several PCs are depicted in Fig. 1 for the task of",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
522,
506,
535
],
"spans": [
{
"bbox": [
105,
522,
506,
535
],
"score": 1.0,
"content": "set prediction. The slot encoder is shared across all NPPs, whereas the PC of each NPP models a",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
533,
505,
547
],
"spans": [
{
"bbox": [
105,
533,
505,
547
],
"score": 1.0,
"content": "separate category of attributes. In this way, each NPP models the joint distribution over slot encodings",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
545,
505,
558
],
"spans": [
{
"bbox": [
106,
545,
505,
558
],
"score": 1.0,
"content": "and object attribute values, such as the color of an object. By querying the NPP, one can obtain",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
555,
429,
568
],
"spans": [
{
"bbox": [
106,
555,
429,
568
],
"score": 1.0,
"content": "task-related probability estimations, such as the conditional attribute probability.",
"type": "text"
}
],
"index": 29
}
],
"index": 21
},
{
"type": "text",
"bbox": [
107,
572,
505,
672
],
"lines": [
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "The second component of SLASH is the logical program, which consists of a set of facts and logical",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
583,
506,
595
],
"spans": [
{
"bbox": [
105,
583,
506,
595
],
"score": 1.0,
"content": "statements defining the state of the world of the underlying task. For example, one can define the rules",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
594,
506,
607
],
"spans": [
{
"bbox": [
105,
594,
506,
607
],
"score": 1.0,
"content": "for when an object possesses a specific set of attributes (cf. Fig. 1). Thirdly, an ASP module is used to",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
605,
505,
618
],
"spans": [
{
"bbox": [
105,
605,
505,
618
],
"score": 1.0,
"content": "combine the first two components. Given a logical query about the input data, the logical program and",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
615,
505,
629
],
"spans": [
{
"bbox": [
105,
615,
505,
629
],
"score": 1.0,
"content": "the probability estimates obtained from the NPP(s), the ASP module produces a probability estimate",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
625,
505,
641
],
"spans": [
{
"bbox": [
105,
625,
505,
641
],
"score": 1.0,
"content": "about the truth value of the query, stating, e.g., how likely it is for a specific object in an image to be",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
638,
506,
651
],
"spans": [
{
"bbox": [
105,
638,
506,
651
],
"score": 1.0,
"content": "a large, dark red triangle. In contrast to query evaluation in Prolog (Colmerauer & Roussel, 1993;",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
648,
506,
662
],
"spans": [
{
"bbox": [
105,
648,
506,
662
],
"score": 1.0,
"content": "Clocksin & Mellish, 1981) which may lead to an infinite loop, many modern answer set solvers use",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
659,
433,
673
],
"spans": [
{
"bbox": [
105,
659,
433,
673
],
"score": 1.0,
"content": "Conflict-Driven-Clause-Learning (CDPL) which, in principle, always terminates.",
"type": "text"
}
],
"index": 38
}
],
"index": 34
},
{
"type": "text",
"bbox": [
107,
677,
505,
732
],
"lines": [
{
"bbox": [
105,
676,
505,
690
],
"spans": [
{
"bbox": [
105,
676,
505,
690
],
"score": 1.0,
"content": "Training in SLASH is performed efficiently in a batch-wise and end-to-end fashion, by integrating",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "the parameters of all modules, neural and probabilistic, into a single loss term. SLASH thus allows a",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
698,
506,
712
],
"spans": [
{
"bbox": [
105,
698,
506,
712
],
"score": 1.0,
"content": "simple, fast and effective integration of sub-symbolic and symbolic computations. In our experiments,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
506,
722
],
"score": 1.0,
"content": "we investigate the advantages of SLASH in comparison to SOTA DPPLs on the benchmark task of",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
720,
505,
734
],
"spans": [
{
"bbox": [
105,
720,
505,
734
],
"score": 1.0,
"content": "MNIST-Addition (Manhaeve et al., 2018). We hereby show SLASH’s increased scalability regarding",
"type": "text"
}
],
"index": 43
}
],
"index": 41
}
],
"page_idx": 1,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
39
],
"spans": [
{
"bbox": [
106,
25,
309,
39
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
301,
750,
310,
763
],
"spans": [
{
"bbox": [
301,
750,
310,
763
],
"score": 1.0,
"content": "2",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
113,
83,
496,
246
],
"blocks": [
{
"type": "image_body",
"bbox": [
113,
83,
496,
246
],
"group_id": 0,
"lines": [
{
"bbox": [
113,
83,
496,
246
],
"spans": [
{
"bbox": [
113,
83,
496,
246
],
"score": 0.967,
"type": "image",
"image_path": "fbc474f94d8a2574e5362c4b2e3c3d61cc7e2bb6baf0aac66d0dc2840ab2de17.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
113,
83,
496,
137.33333333333334
],
"spans": [],
"index": 0
},
{
"bbox": [
113,
137.33333333333334,
496,
191.66666666666669
],
"spans": [],
"index": 1
},
{
"bbox": [
113,
191.66666666666669,
496,
246.00000000000003
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
258,
505,
369
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
258,
505,
271
],
"spans": [
{
"bbox": [
106,
258,
505,
271
],
"score": 1.0,
"content": "Figure 1: SLASH Attention illustrated for a visual reasoning task. SLASH with Neural-Probabilistic",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
269,
506,
282
],
"spans": [
{
"bbox": [
105,
269,
506,
282
],
"score": 1.0,
"content": "Predicates consisting of a slot attention encoder and Probabilistic Circuits (PCs) realised via EiNets.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
280,
506,
293
],
"spans": [
{
"bbox": [
105,
280,
506,
293
],
"score": 1.0,
"content": "The slot encoder is shared over all NPPs. Each triangle in the figure represents a single EiNet that",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
291,
506,
304
],
"spans": [
{
"bbox": [
105,
291,
506,
304
],
"score": 1.0,
"content": "gives us a joint distribution at the root node. Thus, each PC learns the joint distribution over slot",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
302,
506,
316
],
"spans": [
{
"bbox": [
105,
302,
153,
316
],
"score": 1.0,
"content": "encodings,",
"type": "text"
},
{
"bbox": [
153,
302,
163,
313
],
"score": 0.84,
"content": "z ^ { i }",
"type": "inline_equation"
},
{
"bbox": [
163,
302,
257,
316
],
"score": 1.0,
"content": ", and object attributes,",
"type": "text"
},
{
"bbox": [
257,
303,
266,
313
],
"score": 0.76,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
266,
302,
506,
316
],
"score": 1.0,
"content": ", of a specific category, e.g. color attributes. Via targeted",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
313,
506,
327
],
"spans": [
{
"bbox": [
105,
313,
506,
327
],
"score": 1.0,
"content": "queries to the NPPs, one can obtain task-related probabilities, e.g. conditional probabilities for the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
323,
507,
338
],
"spans": [
{
"bbox": [
105,
323,
507,
338
],
"score": 1.0,
"content": "task of set prediction. Given the probability estimates from the NPP(s) and a SLASH program,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
335,
506,
348
],
"spans": [
{
"bbox": [
105,
335,
506,
348
],
"score": 1.0,
"content": "containing a set of facts and logical statements about the world, the probability of the truth value of a",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
346,
506,
359
],
"spans": [
{
"bbox": [
106,
346,
506,
359
],
"score": 1.0,
"content": "task-related query are computed via answer set programming. The entire system, including the neural",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
357,
437,
370
],
"spans": [
{
"bbox": [
106,
357,
437,
370
],
"score": 1.0,
"content": "and probabilistic modules, are finally trained end-to-end via a single loss function.",
"type": "text"
}
],
"index": 12
}
],
"index": 7.5
}
],
"index": 4.25
},
{
"type": "text",
"bbox": [
107,
380,
505,
567
],
"lines": [
{
"bbox": [
106,
380,
505,
393
],
"spans": [
{
"bbox": [
106,
380,
505,
393
],
"score": 1.0,
"content": "We propose SLASH – a novel DPPL that, similar to the punctuation symbol, can be used to efficiently",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
390,
505,
404
],
"spans": [
{
"bbox": [
105,
390,
505,
404
],
"score": 1.0,
"content": "combine several paradigms into one. Specifically, SLASH represents a scalable programming",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
402,
505,
415
],
"spans": [
{
"bbox": [
106,
402,
505,
415
],
"score": 1.0,
"content": "language that seamlessly integrates probabilistic logical programming with neural representations",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
412,
505,
425
],
"spans": [
{
"bbox": [
106,
412,
505,
425
],
"score": 1.0,
"content": "and tractable probabilistic estimations. Fig. 1 shows an example instantiation of SLASH, termed",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
424,
506,
437
],
"spans": [
{
"bbox": [
106,
424,
506,
437
],
"score": 1.0,
"content": "SLASH Attention, for object-centric set prediction. SLASH consists of several key building blocks.",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
434,
506,
447
],
"spans": [
{
"bbox": [
106,
434,
506,
447
],
"score": 1.0,
"content": "Firstly, it makes use of Neural-Probabilistic Predicates (NPPs) for probability estimation. NPPs",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
104,
444,
506,
460
],
"spans": [
{
"bbox": [
104,
444,
506,
460
],
"score": 1.0,
"content": "consist of neural and/or probabilistic circuit (PC) modules and act as a unifying term, encompassing",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
455,
506,
470
],
"spans": [
{
"bbox": [
105,
455,
506,
470
],
"score": 1.0,
"content": "the neural predicates of DeepProbLog and NeurASP, as well as purely probabilistic predicates. In",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
467,
505,
480
],
"spans": [
{
"bbox": [
106,
467,
505,
480
],
"score": 1.0,
"content": "this work, we introduce a much more powerful “flavor” of NPPs that consist jointly of neural and PC",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
478,
506,
492
],
"spans": [
{
"bbox": [
105,
478,
506,
492
],
"score": 1.0,
"content": "modules, taking advantage of the power of neural computations together with true density estimation",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
487,
507,
504
],
"spans": [
{
"bbox": [
105,
487,
507,
504
],
"score": 1.0,
"content": "of PCs. Depending on the underlying task one can thus ask a range of queries to the NPP, e.g.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
501,
505,
513
],
"spans": [
{
"bbox": [
105,
501,
505,
513
],
"score": 1.0,
"content": "sample an unknown, desired variable, but also query for conditional class probabilities. Example",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
511,
506,
524
],
"spans": [
{
"bbox": [
105,
511,
506,
524
],
"score": 1.0,
"content": "NPPs consisting of a slot attention encoder and several PCs are depicted in Fig. 1 for the task of",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
522,
506,
535
],
"spans": [
{
"bbox": [
105,
522,
506,
535
],
"score": 1.0,
"content": "set prediction. The slot encoder is shared across all NPPs, whereas the PC of each NPP models a",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
533,
505,
547
],
"spans": [
{
"bbox": [
105,
533,
505,
547
],
"score": 1.0,
"content": "separate category of attributes. In this way, each NPP models the joint distribution over slot encodings",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
545,
505,
558
],
"spans": [
{
"bbox": [
106,
545,
505,
558
],
"score": 1.0,
"content": "and object attribute values, such as the color of an object. By querying the NPP, one can obtain",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
555,
429,
568
],
"spans": [
{
"bbox": [
106,
555,
429,
568
],
"score": 1.0,
"content": "task-related probability estimations, such as the conditional attribute probability.",
"type": "text"
}
],
"index": 29
}
],
"index": 21,
"bbox_fs": [
104,
380,
507,
568
]
},
{
"type": "text",
"bbox": [
107,
572,
505,
672
],
"lines": [
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "The second component of SLASH is the logical program, which consists of a set of facts and logical",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
583,
506,
595
],
"spans": [
{
"bbox": [
105,
583,
506,
595
],
"score": 1.0,
"content": "statements defining the state of the world of the underlying task. For example, one can define the rules",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
594,
506,
607
],
"spans": [
{
"bbox": [
105,
594,
506,
607
],
"score": 1.0,
"content": "for when an object possesses a specific set of attributes (cf. Fig. 1). Thirdly, an ASP module is used to",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
605,
505,
618
],
"spans": [
{
"bbox": [
105,
605,
505,
618
],
"score": 1.0,
"content": "combine the first two components. Given a logical query about the input data, the logical program and",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
615,
505,
629
],
"spans": [
{
"bbox": [
105,
615,
505,
629
],
"score": 1.0,
"content": "the probability estimates obtained from the NPP(s), the ASP module produces a probability estimate",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
625,
505,
641
],
"spans": [
{
"bbox": [
105,
625,
505,
641
],
"score": 1.0,
"content": "about the truth value of the query, stating, e.g., how likely it is for a specific object in an image to be",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
638,
506,
651
],
"spans": [
{
"bbox": [
105,
638,
506,
651
],
"score": 1.0,
"content": "a large, dark red triangle. In contrast to query evaluation in Prolog (Colmerauer & Roussel, 1993;",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
648,
506,
662
],
"spans": [
{
"bbox": [
105,
648,
506,
662
],
"score": 1.0,
"content": "Clocksin & Mellish, 1981) which may lead to an infinite loop, many modern answer set solvers use",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
659,
433,
673
],
"spans": [
{
"bbox": [
105,
659,
433,
673
],
"score": 1.0,
"content": "Conflict-Driven-Clause-Learning (CDPL) which, in principle, always terminates.",
"type": "text"
}
],
"index": 38
}
],
"index": 34,
"bbox_fs": [
105,
572,
506,
673
]
},
{
"type": "text",
"bbox": [
107,
677,
505,
732
],
"lines": [
{
"bbox": [
105,
676,
505,
690
],
"spans": [
{
"bbox": [
105,
676,
505,
690
],
"score": 1.0,
"content": "Training in SLASH is performed efficiently in a batch-wise and end-to-end fashion, by integrating",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "the parameters of all modules, neural and probabilistic, into a single loss term. SLASH thus allows a",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
698,
506,
712
],
"spans": [
{
"bbox": [
105,
698,
506,
712
],
"score": 1.0,
"content": "simple, fast and effective integration of sub-symbolic and symbolic computations. In our experiments,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
506,
722
],
"score": 1.0,
"content": "we investigate the advantages of SLASH in comparison to SOTA DPPLs on the benchmark task of",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
720,
505,
734
],
"spans": [
{
"bbox": [
105,
720,
505,
734
],
"score": 1.0,
"content": "MNIST-Addition (Manhaeve et al., 2018). We hereby show SLASH’s increased scalability regarding",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
82,
505,
95
],
"spans": [
{
"bbox": [
105,
82,
505,
95
],
"score": 1.0,
"content": "computation time, as well as SLASH’s ability to handle incomplete data via true probabilistic density",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
106,
94,
505,
106
],
"spans": [
{
"bbox": [
106,
94,
505,
106
],
"score": 1.0,
"content": "modelling. Next, we show that SLASH Attention provides superior results for set prediction in terms",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
105,
105,
505,
117
],
"spans": [
{
"bbox": [
105,
105,
505,
117
],
"score": 1.0,
"content": "of accuracy and generalization abilities compared to a baseline slot attention encoder. With our",
"type": "text",
"cross_page": true
}
],
"index": 2
},
{
"bbox": [
105,
114,
506,
129
],
"spans": [
{
"bbox": [
105,
114,
506,
129
],
"score": 1.0,
"content": "experiments, we thus show that SLASH is a realization of “one system – two approaches” (Bengio,",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
105,
126,
489,
140
],
"spans": [
{
"bbox": [
105,
126,
489,
140
],
"score": 1.0,
"content": "2019), that can successfully be used for performing various tasks and on a variety of data types.",
"type": "text",
"cross_page": true
}
],
"index": 4
}
],
"index": 41,
"bbox_fs": [
105,
676,
506,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
138
],
"lines": [
{
"bbox": [
105,
82,
505,
95
],
"spans": [
{
"bbox": [
105,
82,
505,
95
],
"score": 1.0,
"content": "computation time, as well as SLASH’s ability to handle incomplete data via true probabilistic density",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
94,
505,
106
],
"spans": [
{
"bbox": [
106,
94,
505,
106
],
"score": 1.0,
"content": "modelling. Next, we show that SLASH Attention provides superior results for set prediction in terms",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
105,
505,
117
],
"spans": [
{
"bbox": [
105,
105,
505,
117
],
"score": 1.0,
"content": "of accuracy and generalization abilities compared to a baseline slot attention encoder. With our",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
114,
506,
129
],
"spans": [
{
"bbox": [
105,
114,
506,
129
],
"score": 1.0,
"content": "experiments, we thus show that SLASH is a realization of “one system – two approaches” (Bengio,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
126,
489,
140
],
"spans": [
{
"bbox": [
105,
126,
489,
140
],
"score": 1.0,
"content": "2019), that can successfully be used for performing various tasks and on a variety of data types.",
"type": "text"
}
],
"index": 4
}
],
"index": 2
},
{
"type": "text",
"bbox": [
107,
143,
505,
198
],
"lines": [
{
"bbox": [
106,
144,
505,
155
],
"spans": [
{
"bbox": [
106,
144,
505,
155
],
"score": 1.0,
"content": "We make the following contributions: (1) We introduce neural-probabilistic predicates, efficiently",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
155,
505,
167
],
"spans": [
{
"bbox": [
106,
155,
505,
167
],
"score": 1.0,
"content": "integrating answer set programming with probabilistic inference via our novel DPPL, SLASH. (2) We",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
165,
506,
178
],
"spans": [
{
"bbox": [
105,
165,
506,
178
],
"score": 1.0,
"content": "successfully train neural, probabilistic and logic modules within SLASH for complex data structures",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
175,
505,
190
],
"spans": [
{
"bbox": [
105,
175,
505,
190
],
"score": 1.0,
"content": "end-to-end via a simple, single loss term. (3) We show that SLASH provides various advantages",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
480,
200
],
"spans": [
{
"bbox": [
105,
187,
480,
200
],
"score": 1.0,
"content": "across a variety of tasks and data sets compared to state-of-the-art DPPLs and neural models.",
"type": "text"
}
],
"index": 9
}
],
"index": 7
},
{
"type": "title",
"bbox": [
107,
214,
344,
227
],
"lines": [
{
"bbox": [
104,
213,
345,
229
],
"spans": [
{
"bbox": [
104,
213,
345,
229
],
"score": 1.0,
"content": "2 NEURO-SYMBOLIC LOGIC PROGRAMMING",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
239,
506,
273
],
"lines": [
{
"bbox": [
105,
237,
506,
252
],
"spans": [
{
"bbox": [
105,
237,
506,
252
],
"score": 1.0,
"content": "Neuro-Symbolic AI can be divided into two lines of research, depending on the starting point. Both,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
250,
505,
263
],
"spans": [
{
"bbox": [
105,
250,
505,
263
],
"score": 1.0,
"content": "however, have the same final goal: to combine low-level perception with logical constraints and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
260,
150,
275
],
"spans": [
{
"bbox": [
105,
260,
150,
275
],
"score": 1.0,
"content": "reasoning.",
"type": "text"
}
],
"index": 13
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
277,
505,
399
],
"lines": [
{
"bbox": [
106,
278,
506,
290
],
"spans": [
{
"bbox": [
106,
278,
506,
290
],
"score": 1.0,
"content": "A key motivation of Neuro-Symbolic AI (d’Avila Garcez et al., 2009; Mao et al., 2019; Hudson &",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
289,
505,
302
],
"spans": [
{
"bbox": [
106,
289,
505,
302
],
"score": 1.0,
"content": "Manning, 2019; d’Avila Garcez et al., 2019; Jiang & Ahn, 2020; d’Avila Garcez & Lamb, 2020) is",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
300,
505,
313
],
"spans": [
{
"bbox": [
106,
300,
505,
313
],
"score": 1.0,
"content": "to combine the advantages of symbolic and neural representations into a joint system. This is often",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
311,
506,
324
],
"spans": [
{
"bbox": [
106,
311,
506,
324
],
"score": 1.0,
"content": "done in a hybrid approach where a neural network acts as a perception module that interfaces with a",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
322,
506,
334
],
"spans": [
{
"bbox": [
105,
322,
506,
334
],
"score": 1.0,
"content": "symbolic reasoning system, e.g. (Mao et al., 2019; Yi et al., 2018). The goal of such an approach is",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
332,
506,
346
],
"spans": [
{
"bbox": [
105,
332,
506,
346
],
"score": 1.0,
"content": "to mitigate the issues of one type of representation by the other, e.g. using the power of symbolic",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
344,
506,
356
],
"spans": [
{
"bbox": [
105,
344,
506,
356
],
"score": 1.0,
"content": "reasoning systems to handle the generalizability issues of neural networks and on the other hand",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
354,
506,
368
],
"spans": [
{
"bbox": [
105,
354,
506,
368
],
"score": 1.0,
"content": "handle the difficulty of noisy data for symbolic systems via neural networks. Recent work has also",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
365,
505,
379
],
"spans": [
{
"bbox": [
105,
365,
505,
379
],
"score": 1.0,
"content": "shown the advantage of Neuro-Symbolic approaches for explaining and revising incorrect decisions",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
376,
506,
390
],
"spans": [
{
"bbox": [
105,
376,
506,
390
],
"score": 1.0,
"content": "(Ciravegna et al., 2020; Stammer et al., 2021). Many of these previous works, however, train the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
388,
299,
400
],
"spans": [
{
"bbox": [
106,
388,
299,
400
],
"score": 1.0,
"content": "sub-symbolic and symbolic modules separately.",
"type": "text"
}
],
"index": 24
}
],
"index": 19
},
{
"type": "text",
"bbox": [
107,
405,
505,
493
],
"lines": [
{
"bbox": [
105,
403,
506,
419
],
"spans": [
{
"bbox": [
105,
403,
506,
419
],
"score": 1.0,
"content": "Deep Probabilistic Programming Languages (DPPLs) are programming languages that combine deep",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
416,
506,
428
],
"spans": [
{
"bbox": [
105,
416,
506,
428
],
"score": 1.0,
"content": "neural networks with probabilistic models and allow a user to express a probabilistic model via a",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
426,
506,
439
],
"spans": [
{
"bbox": [
105,
426,
506,
439
],
"score": 1.0,
"content": "logical program. Similar to Neuro-Symbolic architectures, DPPLs thereby unite the advantages of",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
437,
506,
450
],
"spans": [
{
"bbox": [
105,
437,
506,
450
],
"score": 1.0,
"content": "different paradigms. DPPLs are related to earlier works such as Markov Logic Networks (MLNs)",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
448,
506,
461
],
"spans": [
{
"bbox": [
105,
448,
506,
461
],
"score": 1.0,
"content": "(Richardson & Domingos, 2006). Thereby, the binding link is the Weighted Model Counting (WMC)",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
456,
507,
474
],
"spans": [
{
"bbox": [
104,
456,
160,
474
],
"score": 1.0,
"content": "introduced in",
"type": "text"
},
{
"bbox": [
160,
459,
189,
470
],
"score": 0.41,
"content": "\\mathrm { L P ^ { M L N } }",
"type": "inline_equation"
},
{
"bbox": [
190,
456,
507,
474
],
"score": 1.0,
"content": "(Lee & Wang, 2016). Several DPPLs have been proposed by now, among which",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
470,
507,
484
],
"spans": [
{
"bbox": [
105,
470,
507,
484
],
"score": 1.0,
"content": "are Pyro (Bingham et al., 2019), Edward (Tran et al., 2017), DeepProbLog (Manhaeve et al., 2018),",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
482,
243,
494
],
"spans": [
{
"bbox": [
105,
482,
243,
494
],
"score": 1.0,
"content": "and NeurASP (Yang et al., 2020).",
"type": "text"
}
],
"index": 32
}
],
"index": 28.5
},
{
"type": "text",
"bbox": [
107,
498,
506,
597
],
"lines": [
{
"bbox": [
106,
498,
506,
511
],
"spans": [
{
"bbox": [
106,
498,
506,
511
],
"score": 1.0,
"content": "To resolve the scalability issues of DeepProbLog, which use Sentential Decision Diagrams (SDDs)",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
509,
506,
521
],
"spans": [
{
"bbox": [
106,
509,
506,
521
],
"score": 1.0,
"content": "(Darwiche, 2011) as the underlying data structure to evaluate queries, NeurASP (Yang et al., 2020),",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
520,
506,
533
],
"spans": [
{
"bbox": [
105,
520,
506,
533
],
"score": 1.0,
"content": "offers a solution by utilizing Answer Set Programming (ASP) (Dimopoulos et al., 1997; Soininen &",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
529,
506,
545
],
"spans": [
{
"bbox": [
105,
529,
506,
545
],
"score": 1.0,
"content": "Niemelä, 1999; Marek & Truszczynski, 1999; Calimeri et al., 2020). In this way, NeurASP changes",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
541,
506,
556
],
"spans": [
{
"bbox": [
105,
541,
506,
556
],
"score": 1.0,
"content": "the paradigm from query evaluation to model generation, i.e. instead of constructing an SDD or",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
554,
505,
565
],
"spans": [
{
"bbox": [
106,
554,
505,
565
],
"score": 1.0,
"content": "similar knowledge representation system, NeurASP generates a set of all possible solutions (one",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
564,
505,
576
],
"spans": [
{
"bbox": [
106,
564,
505,
576
],
"score": 1.0,
"content": "model per solution) and estimates the probability for the truth value of each of these solutions. Of",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
575,
506,
588
],
"spans": [
{
"bbox": [
106,
575,
506,
588
],
"score": 1.0,
"content": "those DPPLs that handle learning in a relational, probabilistic setting and in an end-to-end fashion,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
585,
396,
599
],
"spans": [
{
"bbox": [
105,
585,
396,
599
],
"score": 1.0,
"content": "all of these are limited to estimating only conditional class probabilities.",
"type": "text"
}
],
"index": 41
}
],
"index": 37
},
{
"type": "title",
"bbox": [
108,
613,
263,
626
],
"lines": [
{
"bbox": [
105,
612,
265,
628
],
"spans": [
{
"bbox": [
105,
612,
265,
628
],
"score": 1.0,
"content": "3 THE SLASH FRAMEWORK",
"type": "text"
}
],
"index": 42
}
],
"index": 42
},
{
"type": "text",
"bbox": [
108,
637,
504,
671
],
"lines": [
{
"bbox": [
105,
638,
505,
650
],
"spans": [
{
"bbox": [
105,
638,
505,
650
],
"score": 1.0,
"content": "In this section, we introduce our novel DPPL, SLASH. Before we dive into the details of this, it is",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
649,
506,
661
],
"spans": [
{
"bbox": [
105,
649,
506,
661
],
"score": 1.0,
"content": "necessary to first introduce Neural-Probabilistic Predicates, for which we require an understanding of",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
660,
430,
672
],
"spans": [
{
"bbox": [
105,
660,
430,
672
],
"score": 1.0,
"content": "Probabilistic Circuits. Finally, we will present the learning paradigm of SLASH.",
"type": "text"
}
],
"index": 45
}
],
"index": 44
},
{
"type": "text",
"bbox": [
107,
676,
504,
731
],
"lines": [
{
"bbox": [
106,
677,
506,
689
],
"spans": [
{
"bbox": [
106,
677,
506,
689
],
"score": 1.0,
"content": "The term probabilistic circuit (PC) (Choi et al., 2020) represents a unifying framework that en-",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "compasses all computational graphs which encode probability distributions and guarantee tractable",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
698,
506,
712
],
"spans": [
{
"bbox": [
105,
698,
506,
712
],
"score": 1.0,
"content": "probabilistic modelling. These include Sum-Product Networks (SPNs) (Poon & Domingos, 2011)",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
709,
505,
723
],
"spans": [
{
"bbox": [
105,
709,
505,
723
],
"score": 1.0,
"content": "which are deep mixture models represented via a rooted directed acyclic graphs with a recursively",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
720,
178,
733
],
"spans": [
{
"bbox": [
105,
720,
178,
733
],
"score": 1.0,
"content": "defined structure.",
"type": "text"
}
],
"index": 50
}
],
"index": 48
}
],
"page_idx": 2,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
303,
751,
309,
759
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "3",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
138
],
"lines": [],
"index": 2,
"bbox_fs": [
105,
82,
506,
140
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
143,
505,
198
],
"lines": [
{
"bbox": [
106,
144,
505,
155
],
"spans": [
{
"bbox": [
106,
144,
505,
155
],
"score": 1.0,
"content": "We make the following contributions: (1) We introduce neural-probabilistic predicates, efficiently",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
155,
505,
167
],
"spans": [
{
"bbox": [
106,
155,
505,
167
],
"score": 1.0,
"content": "integrating answer set programming with probabilistic inference via our novel DPPL, SLASH. (2) We",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
165,
506,
178
],
"spans": [
{
"bbox": [
105,
165,
506,
178
],
"score": 1.0,
"content": "successfully train neural, probabilistic and logic modules within SLASH for complex data structures",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
175,
505,
190
],
"spans": [
{
"bbox": [
105,
175,
505,
190
],
"score": 1.0,
"content": "end-to-end via a simple, single loss term. (3) We show that SLASH provides various advantages",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
480,
200
],
"spans": [
{
"bbox": [
105,
187,
480,
200
],
"score": 1.0,
"content": "across a variety of tasks and data sets compared to state-of-the-art DPPLs and neural models.",
"type": "text"
}
],
"index": 9
}
],
"index": 7,
"bbox_fs": [
105,
144,
506,
200
]
},
{
"type": "title",
"bbox": [
107,
214,
344,
227
],
"lines": [
{
"bbox": [
104,
213,
345,
229
],
"spans": [
{
"bbox": [
104,
213,
345,
229
],
"score": 1.0,
"content": "2 NEURO-SYMBOLIC LOGIC PROGRAMMING",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
239,
506,
273
],
"lines": [
{
"bbox": [
105,
237,
506,
252
],
"spans": [
{
"bbox": [
105,
237,
506,
252
],
"score": 1.0,
"content": "Neuro-Symbolic AI can be divided into two lines of research, depending on the starting point. Both,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
250,
505,
263
],
"spans": [
{
"bbox": [
105,
250,
505,
263
],
"score": 1.0,
"content": "however, have the same final goal: to combine low-level perception with logical constraints and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
260,
150,
275
],
"spans": [
{
"bbox": [
105,
260,
150,
275
],
"score": 1.0,
"content": "reasoning.",
"type": "text"
}
],
"index": 13
}
],
"index": 12,
"bbox_fs": [
105,
237,
506,
275
]
},
{
"type": "text",
"bbox": [
107,
277,
505,
399
],
"lines": [
{
"bbox": [
106,
278,
506,
290
],
"spans": [
{
"bbox": [
106,
278,
506,
290
],
"score": 1.0,
"content": "A key motivation of Neuro-Symbolic AI (d’Avila Garcez et al., 2009; Mao et al., 2019; Hudson &",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
289,
505,
302
],
"spans": [
{
"bbox": [
106,
289,
505,
302
],
"score": 1.0,
"content": "Manning, 2019; d’Avila Garcez et al., 2019; Jiang & Ahn, 2020; d’Avila Garcez & Lamb, 2020) is",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
300,
505,
313
],
"spans": [
{
"bbox": [
106,
300,
505,
313
],
"score": 1.0,
"content": "to combine the advantages of symbolic and neural representations into a joint system. This is often",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
311,
506,
324
],
"spans": [
{
"bbox": [
106,
311,
506,
324
],
"score": 1.0,
"content": "done in a hybrid approach where a neural network acts as a perception module that interfaces with a",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
322,
506,
334
],
"spans": [
{
"bbox": [
105,
322,
506,
334
],
"score": 1.0,
"content": "symbolic reasoning system, e.g. (Mao et al., 2019; Yi et al., 2018). The goal of such an approach is",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
332,
506,
346
],
"spans": [
{
"bbox": [
105,
332,
506,
346
],
"score": 1.0,
"content": "to mitigate the issues of one type of representation by the other, e.g. using the power of symbolic",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
344,
506,
356
],
"spans": [
{
"bbox": [
105,
344,
506,
356
],
"score": 1.0,
"content": "reasoning systems to handle the generalizability issues of neural networks and on the other hand",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
354,
506,
368
],
"spans": [
{
"bbox": [
105,
354,
506,
368
],
"score": 1.0,
"content": "handle the difficulty of noisy data for symbolic systems via neural networks. Recent work has also",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
365,
505,
379
],
"spans": [
{
"bbox": [
105,
365,
505,
379
],
"score": 1.0,
"content": "shown the advantage of Neuro-Symbolic approaches for explaining and revising incorrect decisions",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
376,
506,
390
],
"spans": [
{
"bbox": [
105,
376,
506,
390
],
"score": 1.0,
"content": "(Ciravegna et al., 2020; Stammer et al., 2021). Many of these previous works, however, train the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
388,
299,
400
],
"spans": [
{
"bbox": [
106,
388,
299,
400
],
"score": 1.0,
"content": "sub-symbolic and symbolic modules separately.",
"type": "text"
}
],
"index": 24
}
],
"index": 19,
"bbox_fs": [
105,
278,
506,
400
]
},
{
"type": "text",
"bbox": [
107,
405,
505,
493
],
"lines": [
{
"bbox": [
105,
403,
506,
419
],
"spans": [
{
"bbox": [
105,
403,
506,
419
],
"score": 1.0,
"content": "Deep Probabilistic Programming Languages (DPPLs) are programming languages that combine deep",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
416,
506,
428
],
"spans": [
{
"bbox": [
105,
416,
506,
428
],
"score": 1.0,
"content": "neural networks with probabilistic models and allow a user to express a probabilistic model via a",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
426,
506,
439
],
"spans": [
{
"bbox": [
105,
426,
506,
439
],
"score": 1.0,
"content": "logical program. Similar to Neuro-Symbolic architectures, DPPLs thereby unite the advantages of",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
437,
506,
450
],
"spans": [
{
"bbox": [
105,
437,
506,
450
],
"score": 1.0,
"content": "different paradigms. DPPLs are related to earlier works such as Markov Logic Networks (MLNs)",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
448,
506,
461
],
"spans": [
{
"bbox": [
105,
448,
506,
461
],
"score": 1.0,
"content": "(Richardson & Domingos, 2006). Thereby, the binding link is the Weighted Model Counting (WMC)",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
456,
507,
474
],
"spans": [
{
"bbox": [
104,
456,
160,
474
],
"score": 1.0,
"content": "introduced in",
"type": "text"
},
{
"bbox": [
160,
459,
189,
470
],
"score": 0.41,
"content": "\\mathrm { L P ^ { M L N } }",
"type": "inline_equation"
},
{
"bbox": [
190,
456,
507,
474
],
"score": 1.0,
"content": "(Lee & Wang, 2016). Several DPPLs have been proposed by now, among which",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
470,
507,
484
],
"spans": [
{
"bbox": [
105,
470,
507,
484
],
"score": 1.0,
"content": "are Pyro (Bingham et al., 2019), Edward (Tran et al., 2017), DeepProbLog (Manhaeve et al., 2018),",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
482,
243,
494
],
"spans": [
{
"bbox": [
105,
482,
243,
494
],
"score": 1.0,
"content": "and NeurASP (Yang et al., 2020).",
"type": "text"
}
],
"index": 32
}
],
"index": 28.5,
"bbox_fs": [
104,
403,
507,
494
]
},
{
"type": "text",
"bbox": [
107,
498,
506,
597
],
"lines": [
{
"bbox": [
106,
498,
506,
511
],
"spans": [
{
"bbox": [
106,
498,
506,
511
],
"score": 1.0,
"content": "To resolve the scalability issues of DeepProbLog, which use Sentential Decision Diagrams (SDDs)",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
509,
506,
521
],
"spans": [
{
"bbox": [
106,
509,
506,
521
],
"score": 1.0,
"content": "(Darwiche, 2011) as the underlying data structure to evaluate queries, NeurASP (Yang et al., 2020),",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
520,
506,
533
],
"spans": [
{
"bbox": [
105,
520,
506,
533
],
"score": 1.0,
"content": "offers a solution by utilizing Answer Set Programming (ASP) (Dimopoulos et al., 1997; Soininen &",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
529,
506,
545
],
"spans": [
{
"bbox": [
105,
529,
506,
545
],
"score": 1.0,
"content": "Niemelä, 1999; Marek & Truszczynski, 1999; Calimeri et al., 2020). In this way, NeurASP changes",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
541,
506,
556
],
"spans": [
{
"bbox": [
105,
541,
506,
556
],
"score": 1.0,
"content": "the paradigm from query evaluation to model generation, i.e. instead of constructing an SDD or",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
554,
505,
565
],
"spans": [
{
"bbox": [
106,
554,
505,
565
],
"score": 1.0,
"content": "similar knowledge representation system, NeurASP generates a set of all possible solutions (one",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
564,
505,
576
],
"spans": [
{
"bbox": [
106,
564,
505,
576
],
"score": 1.0,
"content": "model per solution) and estimates the probability for the truth value of each of these solutions. Of",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
575,
506,
588
],
"spans": [
{
"bbox": [
106,
575,
506,
588
],
"score": 1.0,
"content": "those DPPLs that handle learning in a relational, probabilistic setting and in an end-to-end fashion,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
585,
396,
599
],
"spans": [
{
"bbox": [
105,
585,
396,
599
],
"score": 1.0,
"content": "all of these are limited to estimating only conditional class probabilities.",
"type": "text"
}
],
"index": 41
}
],
"index": 37,
"bbox_fs": [
105,
498,
506,
599
]
},
{
"type": "title",
"bbox": [
108,
613,
263,
626
],
"lines": [
{
"bbox": [
105,
612,
265,
628
],
"spans": [
{
"bbox": [
105,
612,
265,
628
],
"score": 1.0,
"content": "3 THE SLASH FRAMEWORK",
"type": "text"
}
],
"index": 42
}
],
"index": 42
},
{
"type": "text",
"bbox": [
108,
637,
504,
671
],
"lines": [
{
"bbox": [
105,
638,
505,
650
],
"spans": [
{
"bbox": [
105,
638,
505,
650
],
"score": 1.0,
"content": "In this section, we introduce our novel DPPL, SLASH. Before we dive into the details of this, it is",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
649,
506,
661
],
"spans": [
{
"bbox": [
105,
649,
506,
661
],
"score": 1.0,
"content": "necessary to first introduce Neural-Probabilistic Predicates, for which we require an understanding of",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
660,
430,
672
],
"spans": [
{
"bbox": [
105,
660,
430,
672
],
"score": 1.0,
"content": "Probabilistic Circuits. Finally, we will present the learning paradigm of SLASH.",
"type": "text"
}
],
"index": 45
}
],
"index": 44,
"bbox_fs": [
105,
638,
506,
672
]
},
{
"type": "text",
"bbox": [
107,
676,
504,
731
],
"lines": [
{
"bbox": [
106,
677,
506,
689
],
"spans": [
{
"bbox": [
106,
677,
506,
689
],
"score": 1.0,
"content": "The term probabilistic circuit (PC) (Choi et al., 2020) represents a unifying framework that en-",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "compasses all computational graphs which encode probability distributions and guarantee tractable",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
698,
506,
712
],
"spans": [
{
"bbox": [
105,
698,
506,
712
],
"score": 1.0,
"content": "probabilistic modelling. These include Sum-Product Networks (SPNs) (Poon & Domingos, 2011)",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
709,
505,
723
],
"spans": [
{
"bbox": [
105,
709,
505,
723
],
"score": 1.0,
"content": "which are deep mixture models represented via a rooted directed acyclic graphs with a recursively",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
720,
178,
733
],
"spans": [
{
"bbox": [
105,
720,
178,
733
],
"score": 1.0,
"content": "defined structure.",
"type": "text"
}
],
"index": 50
}
],
"index": 48,
"bbox_fs": [
105,
677,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
138,
82,
299,
190
],
"blocks": [
{
"type": "image_body",
"bbox": [
138,
82,
299,
190
],
"group_id": 1,
"lines": [
{
"bbox": [
138,
82,
299,
190
],
"spans": [
{
"bbox": [
138,
82,
299,
190
],
"score": 0.968,
"type": "image",
"image_path": "fe12c632ab68913a19fb5c0639cac79765cabf16e331224edc0cd2c84bb8eebe.jpg"
}
]
}
],
"index": 3.5,
"virtual_lines": [
{
"bbox": [
138,
82,
299,
95.5
],
"spans": [],
"index": 0
},
{
"bbox": [
138,
95.5,
299,
109.0
],
"spans": [],
"index": 1
},
{
"bbox": [
138,
109.0,
299,
122.5
],
"spans": [],
"index": 2
},
{
"bbox": [
138,
122.5,
299,
136.0
],
"spans": [],
"index": 3
},
{
"bbox": [
138,
136.0,
299,
149.5
],
"spans": [],
"index": 4
},
{
"bbox": [
138,
149.5,
299,
163.0
],
"spans": [],
"index": 5
},
{
"bbox": [
138,
163.0,
299,
176.5
],
"spans": [],
"index": 6
},
{
"bbox": [
138,
176.5,
299,
190.0
],
"spans": [],
"index": 7
}
]
}
],
"index": 3.5
},
{
"type": "text",
"bbox": [
128,
195,
309,
215
],
"lines": [
{
"bbox": [
128,
194,
309,
207
],
"spans": [
{
"bbox": [
128,
194,
309,
207
],
"score": 1.0,
"content": "(a) NPPs come in various flavors depending on the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
128,
205,
221,
215
],
"spans": [
{
"bbox": [
128,
205,
221,
215
],
"score": 1.0,
"content": "data and underlying task.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
329,
80,
474,
189
],
"blocks": [
{
"type": "image_body",
"bbox": [
329,
80,
474,
189
],
"group_id": 0,
"lines": [
{
"bbox": [
329,
80,
474,
189
],
"spans": [
{
"bbox": [
329,
80,
474,
189
],
"score": 0.954,
"type": "image",
"image_path": "deaa2f06b12c44d64f5ac8f226a3fdb61c0d8bba50b15128cddc5ccafbad3205.jpg"
}
]
}
],
"index": 9.0,
"virtual_lines": [
{
"bbox": [
329,
80,
474,
134.5
],
"spans": [],
"index": 8
},
{
"bbox": [
329,
134.5,
474,
189.0
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
321,
195,
482,
214
],
"group_id": 0,
"lines": [
{
"bbox": [
321,
194,
483,
206
],
"spans": [
{
"bbox": [
321,
194,
483,
206
],
"score": 1.0,
"content": "(b) Minimal SLASH program and query for",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
321,
205,
375,
216
],
"spans": [
{
"bbox": [
321,
205,
375,
216
],
"score": 1.0,
"content": "set prediction.",
"type": "text"
}
],
"index": 13
}
],
"index": 12.5
},
{
"type": "image_caption",
"bbox": [
106,
225,
506,
280
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
224,
507,
237
],
"spans": [
{
"bbox": [
105,
224,
507,
237
],
"score": 1.0,
"content": "Figure 2: (a) Depending on the data set and underlying task, SLASH requires a suitable Neural-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
235,
505,
248
],
"spans": [
{
"bbox": [
105,
235,
505,
248
],
"score": 1.0,
"content": "Probabilistic Predicate (NPP) that computes query-dependent probability estimates. An NPP can be",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
246,
507,
260
],
"spans": [
{
"bbox": [
105,
246,
507,
260
],
"score": 1.0,
"content": "composed of neural and probabilistic modules, or (depicted via slash symbol) only one of these two.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
257,
506,
270
],
"spans": [
{
"bbox": [
105,
257,
506,
270
],
"score": 1.0,
"content": "(b) A minimal SLASH program and query for the set prediction task, here only showing the NPP that",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
268,
447,
281
],
"spans": [
{
"bbox": [
105,
268,
447,
281
],
"score": 1.0,
"content": "models the color category per object. For the full program, we refer to the Appendix.",
"type": "text"
}
],
"index": 18
}
],
"index": 16
}
],
"index": 12.5
},
{
"type": "title",
"bbox": [
108,
302,
294,
313
],
"lines": [
{
"bbox": [
106,
302,
295,
315
],
"spans": [
{
"bbox": [
106,
302,
295,
315
],
"score": 1.0,
"content": "3.1 NEURAL-PROBABILISTIC PREDICATES",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
322,
505,
378
],
"lines": [
{
"bbox": [
105,
322,
505,
336
],
"spans": [
{
"bbox": [
105,
322,
505,
336
],
"score": 1.0,
"content": "Previous DPPLs, DeepProbLog (Manhaeve et al., 2018) and NeurASP (Yang et al., 2020), introduced",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
333,
505,
347
],
"spans": [
{
"bbox": [
105,
333,
505,
347
],
"score": 1.0,
"content": "the Neural Predicate as an annotated-disjunction or as a propositional atom, respectively, to acquire",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
343,
506,
359
],
"spans": [
{
"bbox": [
105,
343,
237,
359
],
"score": 1.0,
"content": "conditional class probabilities,",
"type": "text"
},
{
"bbox": [
237,
345,
273,
357
],
"score": 0.93,
"content": "P ( C | X )",
"type": "inline_equation"
},
{
"bbox": [
273,
343,
506,
359
],
"score": 1.0,
"content": ", via the softmax function at the output of an arbitrary",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
355,
505,
369
],
"spans": [
{
"bbox": [
105,
355,
505,
369
],
"score": 1.0,
"content": "DNN. As mentioned in the introduction, this approach has certain limitations concerning inference",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
366,
463,
380
],
"spans": [
{
"bbox": [
106,
366,
463,
380
],
"score": 1.0,
"content": "capabilities. To resolve this issue, we introduce Neural-Probabilisitic Predicates (NPPs).",
"type": "text"
}
],
"index": 24
}
],
"index": 22
},
{
"type": "text",
"bbox": [
107,
384,
210,
395
],
"lines": [
{
"bbox": [
106,
383,
210,
396
],
"spans": [
{
"bbox": [
106,
383,
210,
396
],
"score": 1.0,
"content": "Formally, we denote with",
"type": "text"
}
],
"index": 25
}
],
"index": 25
},
{
"type": "interline_equation",
"bbox": [
256,
395,
354,
408
],
"lines": [
{
"bbox": [
256,
395,
354,
408
],
"spans": [
{
"bbox": [
256,
395,
354,
408
],
"score": 0.91,
"content": "n p p \\left( h ( x ) , [ v _ { 1 } , . . . , v _ { n } ] \\right)",
"type": "interline_equation",
"image_path": "187dcc297a855ad44e5c575f0d91e7583f0c038c9c3016485793aabe684f440a.jpg"
}
]
}
],
"index": 26,
"virtual_lines": [
{
"bbox": [
256,
395,
354,
408
],
"spans": [],
"index": 26
}
]
},
{
"type": "text",
"bbox": [
107,
411,
505,
466
],
"lines": [
{
"bbox": [
105,
411,
506,
425
],
"spans": [
{
"bbox": [
105,
411,
240,
425
],
"score": 1.0,
"content": "a Neural-Probabilistic Predicate",
"type": "text"
},
{
"bbox": [
241,
412,
248,
422
],
"score": 0.8,
"content": "h",
"type": "inline_equation"
},
{
"bbox": [
248,
411,
489,
425
],
"score": 1.0,
"content": ". Thereby, (i) npp is a reserved word to label an NPP, (ii)",
"type": "text"
},
{
"bbox": [
489,
412,
497,
422
],
"score": 0.74,
"content": "h",
"type": "inline_equation"
},
{
"bbox": [
497,
411,
506,
425
],
"score": 1.0,
"content": "a",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
422,
505,
435
],
"spans": [
{
"bbox": [
105,
422,
505,
435
],
"score": 1.0,
"content": "symbolic name of either a PC, NN or a joint of a PC and NN (cf. Fig. 2a), e.g., color_attr is the name",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
433,
506,
447
],
"spans": [
{
"bbox": [
105,
433,
267,
447
],
"score": 1.0,
"content": "of an NPP of Fig. 2b. Additionally, (iii)",
"type": "text"
},
{
"bbox": [
267,
435,
275,
444
],
"score": 0.79,
"content": "x",
"type": "inline_equation"
},
{
"bbox": [
275,
433,
379,
447
],
"score": 1.0,
"content": "denotes a “term” and (iv)",
"type": "text"
},
{
"bbox": [
379,
435,
422,
445
],
"score": 0.9,
"content": "v _ { 1 } , \\ldots , v _ { n }",
"type": "inline_equation"
},
{
"bbox": [
422,
433,
506,
447
],
"score": 1.0,
"content": "are placeholders for",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
444,
505,
457
],
"spans": [
{
"bbox": [
105,
444,
154,
457
],
"score": 1.0,
"content": "each of the",
"type": "text"
},
{
"bbox": [
155,
446,
162,
455
],
"score": 0.71,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
163,
444,
253,
457
],
"score": 1.0,
"content": "possible outcomes of",
"type": "text"
},
{
"bbox": [
253,
445,
260,
455
],
"score": 0.82,
"content": "h",
"type": "inline_equation"
},
{
"bbox": [
260,
444,
505,
457
],
"score": 1.0,
"content": ". For example, the placeholders for color_attr are the color",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
455,
294,
468
],
"spans": [
{
"bbox": [
106,
455,
294,
468
],
"score": 1.0,
"content": "attributes of an object (Red, Blue, Green, etc.).",
"type": "text"
}
],
"index": 31
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
467,
505,
511
],
"lines": [
{
"bbox": [
105,
466,
506,
479
],
"spans": [
{
"bbox": [
105,
466,
321,
479
],
"score": 1.0,
"content": "An NPP abbreviates an arithmetic literal of the form",
"type": "text"
},
{
"bbox": [
321,
468,
345,
477
],
"score": 0.88,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
346,
466,
367,
479
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
367,
466,
414,
479
],
"score": 0.92,
"content": "c \\in \\{ h ( x ) \\}",
"type": "inline_equation"
},
{
"bbox": [
414,
466,
433,
479
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
433,
466,
503,
479
],
"score": 0.92,
"content": "v \\in \\{ v _ { 1 } , \\ldots , v _ { n } \\}",
"type": "inline_equation"
},
{
"bbox": [
503,
466,
506,
479
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
477,
506,
489
],
"spans": [
{
"bbox": [
105,
477,
224,
489
],
"score": 1.0,
"content": "Furthermore, we denote with",
"type": "text"
},
{
"bbox": [
224,
478,
246,
488
],
"score": 0.88,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
246,
477,
433,
489
],
"score": 1.0,
"content": "a set of NPPs of the form stated in (Eq. 1) and",
"type": "text"
},
{
"bbox": [
433,
479,
452,
487
],
"score": 0.87,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
452,
477,
506,
489
],
"score": 1.0,
"content": "the set of all",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
104,
486,
507,
502
],
"spans": [
{
"bbox": [
104,
486,
128,
502
],
"score": 1.0,
"content": "rules",
"type": "text"
},
{
"bbox": [
129,
490,
153,
498
],
"score": 0.86,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
153,
486,
463,
502
],
"score": 1.0,
"content": "of one NPP, which denotes the possible outcomes, obtained from an NPP in",
"type": "text"
},
{
"bbox": [
464,
489,
485,
499
],
"score": 0.88,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
486,
486,
507,
502
],
"score": 1.0,
"content": ", e.g.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
496,
460,
514
],
"spans": [
{
"bbox": [
106,
499,
313,
512
],
"score": 0.88,
"content": "r ^ { c o l o r \\_ a t t r } = \\{ c = R e d , c = B l u e , c = G r \\bar { e } e n , . . . \\}",
"type": "inline_equation"
},
{
"bbox": [
314,
496,
460,
514
],
"score": 1.0,
"content": "for the example depicted in Fig. 2b.",
"type": "text"
}
],
"index": 35
}
],
"index": 33.5
},
{
"type": "text",
"bbox": [
106,
515,
505,
561
],
"lines": [
{
"bbox": [
105,
515,
506,
529
],
"spans": [
{
"bbox": [
105,
515,
181,
529
],
"score": 1.0,
"content": "Rules of the form",
"type": "text"
},
{
"bbox": [
182,
516,
317,
528
],
"score": 0.86,
"content": "n p p ( h ( x ) , [ v _ { 1 } , \\ldots , v _ { n } ] ) B o d y",
"type": "inline_equation"
},
{
"bbox": [
317,
515,
506,
529
],
"score": 1.0,
"content": "are used as an abbreviation for application to",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
528,
505,
540
],
"spans": [
{
"bbox": [
106,
528,
505,
540
],
"score": 1.0,
"content": "multiple entities, e.g. multiple slots for the task of set prediction (cf. Fig. 2b). Hereby, Body of the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
538,
505,
551
],
"spans": [
{
"bbox": [
105,
538,
188,
551
],
"score": 1.0,
"content": "rule is identified by",
"type": "text"
},
{
"bbox": [
189,
539,
198,
548
],
"score": 0.73,
"content": "\\top",
"type": "inline_equation"
},
{
"bbox": [
198,
538,
278,
551
],
"score": 1.0,
"content": "(tautology, true) or",
"type": "text"
},
{
"bbox": [
278,
539,
288,
549
],
"score": 0.74,
"content": "\\perp",
"type": "inline_equation"
},
{
"bbox": [
288,
538,
505,
551
],
"score": 1.0,
"content": "(contradiction, false) during grounding. Rules of the",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
547,
403,
562
],
"spans": [
{
"bbox": [
104,
547,
128,
562
],
"score": 1.0,
"content": "form",
"type": "text"
},
{
"bbox": [
128,
549,
186,
560
],
"score": 0.81,
"content": "H e a d \\gets B o d y",
"type": "inline_equation"
},
{
"bbox": [
187,
547,
208,
562
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
208,
550,
227,
559
],
"score": 0.87,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
228,
547,
378,
562
],
"score": 1.0,
"content": "appearing in Head are prohibited for",
"type": "text"
},
{
"bbox": [
378,
549,
399,
559
],
"score": 0.88,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
400,
547,
403,
562
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 39
}
],
"index": 37.5
},
{
"type": "text",
"bbox": [
108,
565,
504,
600
],
"lines": [
{
"bbox": [
105,
565,
506,
579
],
"spans": [
{
"bbox": [
105,
565,
506,
579
],
"score": 1.0,
"content": "In this work, we largely make use of NPPs that contain probabilistic circuits (specifically SPNs)",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
577,
506,
590
],
"spans": [
{
"bbox": [
106,
577,
506,
590
],
"score": 1.0,
"content": "which allow for tractable density estimation and modelling of joint probabilities. In this way, it is",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
588,
502,
601
],
"spans": [
{
"bbox": [
106,
588,
365,
601
],
"score": 1.0,
"content": "possible to answer a much richer set of probabilistic queries, i.e.",
"type": "text"
},
{
"bbox": [
366,
588,
403,
600
],
"score": 0.85,
"content": "P ( X , C )",
"type": "inline_equation"
},
{
"bbox": [
403,
588,
407,
601
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
407,
588,
443,
600
],
"score": 0.87,
"content": "P ( X | C )",
"type": "inline_equation"
},
{
"bbox": [
443,
588,
461,
601
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
461,
588,
497,
600
],
"score": 0.92,
"content": "P ( \\bar { C } | X )",
"type": "inline_equation"
},
{
"bbox": [
498,
588,
502,
601
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 42
}
],
"index": 41
},
{
"type": "text",
"bbox": [
107,
604,
505,
671
],
"lines": [
{
"bbox": [
105,
605,
505,
618
],
"spans": [
{
"bbox": [
105,
605,
505,
618
],
"score": 1.0,
"content": "In addition to this, we introduce the arguably more interesting type of NPP that combines a neural",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
616,
505,
628
],
"spans": [
{
"bbox": [
105,
616,
505,
628
],
"score": 1.0,
"content": "module with a PC. Hereby, the neural module learns to map the raw input data into an optimal",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
627,
505,
639
],
"spans": [
{
"bbox": [
105,
627,
505,
639
],
"score": 1.0,
"content": "latent representation, e.g. object-based slot representations. The PC, in turn, learns to model the",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
638,
506,
651
],
"spans": [
{
"bbox": [
105,
638,
506,
651
],
"score": 1.0,
"content": "joint distribution of these latent variables and produces the final probability estimates. This type of",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
648,
505,
661
],
"spans": [
{
"bbox": [
105,
648,
505,
661
],
"score": 1.0,
"content": "NPP nicely combines the representational power of neural networks with the advantages of PCs in",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
660,
282,
673
],
"spans": [
{
"bbox": [
105,
660,
282,
673
],
"score": 1.0,
"content": "probability estimation and query flexibility.",
"type": "text"
}
],
"index": 48
}
],
"index": 45.5
},
{
"type": "text",
"bbox": [
107,
676,
505,
732
],
"lines": [
{
"bbox": [
105,
676,
505,
690
],
"spans": [
{
"bbox": [
105,
676,
505,
690
],
"score": 1.0,
"content": "For making the different probabilistic queries distinguishable in a SLASH program, we introduce the",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
318,
700
],
"score": 1.0,
"content": "following notation. We denote a given variable with",
"type": "text"
},
{
"bbox": [
319,
689,
328,
698
],
"score": 0.69,
"content": "^ +",
"type": "inline_equation"
},
{
"bbox": [
329,
687,
505,
700
],
"score": 1.0,
"content": "and the query variable with −. E.g., within",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
697,
506,
713
],
"spans": [
{
"bbox": [
105,
697,
262,
713
],
"score": 1.0,
"content": "the running example of set prediction",
"type": "text"
},
{
"bbox": [
263,
699,
273,
710
],
"score": 0.31,
"content": "\\cdot e f .",
"type": "inline_equation"
},
{
"bbox": [
273,
697,
441,
713
],
"score": 1.0,
"content": "Fig. 1 and 2b), with the query color_attr",
"type": "text"
},
{
"bbox": [
442,
699,
486,
711
],
"score": 0.87,
"content": "\\cdot ( + X , - C )",
"type": "inline_equation"
},
{
"bbox": [
487,
697,
506,
713
],
"score": 1.0,
"content": "one",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
708,
507,
723
],
"spans": [
{
"bbox": [
105,
708,
158,
723
],
"score": 1.0,
"content": "is asking for",
"type": "text"
},
{
"bbox": [
158,
710,
194,
721
],
"score": 0.91,
"content": "P ( C | X )",
"type": "inline_equation"
},
{
"bbox": [
195,
708,
305,
723
],
"score": 1.0,
"content": ". Similarly, with color_attr",
"type": "text"
},
{
"bbox": [
305,
711,
349,
722
],
"score": 0.86,
"content": "( - X , + C )",
"type": "inline_equation"
},
{
"bbox": [
349,
708,
420,
723
],
"score": 1.0,
"content": "one is asking for",
"type": "text"
},
{
"bbox": [
420,
710,
456,
722
],
"score": 0.93,
"content": "P ( X | C )",
"type": "inline_equation"
},
{
"bbox": [
456,
708,
507,
723
],
"score": 1.0,
"content": "and, finally,",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
105,
719,
272,
734
],
"spans": [
{
"bbox": [
105,
719,
171,
734
],
"score": 1.0,
"content": "with color_attr",
"type": "text"
},
{
"bbox": [
171,
721,
215,
732
],
"score": 0.82,
"content": "( - X , - C )",
"type": "inline_equation"
},
{
"bbox": [
216,
719,
231,
734
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
232,
722,
268,
732
],
"score": 0.93,
"content": "P ( X , C )",
"type": "inline_equation"
},
{
"bbox": [
268,
719,
272,
734
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 53
}
],
"index": 51
}
],
"page_idx": 3,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
39
],
"spans": [
{
"bbox": [
106,
25,
309,
39
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
752,
308,
759
],
"lines": []
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
138,
82,
299,
190
],
"blocks": [
{
"type": "image_body",
"bbox": [
138,
82,
299,
190
],
"group_id": 1,
"lines": [
{
"bbox": [
138,
82,
299,
190
],
"spans": [
{
"bbox": [
138,
82,
299,
190
],
"score": 0.968,
"type": "image",
"image_path": "fe12c632ab68913a19fb5c0639cac79765cabf16e331224edc0cd2c84bb8eebe.jpg"
}
]
}
],
"index": 3.5,
"virtual_lines": [
{
"bbox": [
138,
82,
299,
95.5
],
"spans": [],
"index": 0
},
{
"bbox": [
138,
95.5,
299,
109.0
],
"spans": [],
"index": 1
},
{
"bbox": [
138,
109.0,
299,
122.5
],
"spans": [],
"index": 2
},
{
"bbox": [
138,
122.5,
299,
136.0
],
"spans": [],
"index": 3
},
{
"bbox": [
138,
136.0,
299,
149.5
],
"spans": [],
"index": 4
},
{
"bbox": [
138,
149.5,
299,
163.0
],
"spans": [],
"index": 5
},
{
"bbox": [
138,
163.0,
299,
176.5
],
"spans": [],
"index": 6
},
{
"bbox": [
138,
176.5,
299,
190.0
],
"spans": [],
"index": 7
}
]
}
],
"index": 3.5
},
{
"type": "text",
"bbox": [
128,
195,
309,
215
],
"lines": [
{
"bbox": [
128,
194,
309,
207
],
"spans": [
{
"bbox": [
128,
194,
309,
207
],
"score": 1.0,
"content": "(a) NPPs come in various flavors depending on the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
128,
205,
221,
215
],
"spans": [
{
"bbox": [
128,
205,
221,
215
],
"score": 1.0,
"content": "data and underlying task.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.0,
"bbox_fs": [
128,
194,
309,
215
]
},
{
"type": "image",
"bbox": [
329,
80,
474,
189
],
"blocks": [
{
"type": "image_body",
"bbox": [
329,
80,
474,
189
],
"group_id": 0,
"lines": [
{
"bbox": [
329,
80,
474,
189
],
"spans": [
{
"bbox": [
329,
80,
474,
189
],
"score": 0.954,
"type": "image",
"image_path": "deaa2f06b12c44d64f5ac8f226a3fdb61c0d8bba50b15128cddc5ccafbad3205.jpg"
}
]
}
],
"index": 9.0,
"virtual_lines": [
{
"bbox": [
329,
80,
474,
134.5
],
"spans": [],
"index": 8
},
{
"bbox": [
329,
134.5,
474,
189.0
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
321,
195,
482,
214
],
"group_id": 0,
"lines": [
{
"bbox": [
321,
194,
483,
206
],
"spans": [
{
"bbox": [
321,
194,
483,
206
],
"score": 1.0,
"content": "(b) Minimal SLASH program and query for",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
321,
205,
375,
216
],
"spans": [
{
"bbox": [
321,
205,
375,
216
],
"score": 1.0,
"content": "set prediction.",
"type": "text"
}
],
"index": 13
}
],
"index": 12.5
},
{
"type": "image_caption",
"bbox": [
106,
225,
506,
280
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
224,
507,
237
],
"spans": [
{
"bbox": [
105,
224,
507,
237
],
"score": 1.0,
"content": "Figure 2: (a) Depending on the data set and underlying task, SLASH requires a suitable Neural-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
235,
505,
248
],
"spans": [
{
"bbox": [
105,
235,
505,
248
],
"score": 1.0,
"content": "Probabilistic Predicate (NPP) that computes query-dependent probability estimates. An NPP can be",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
246,
507,
260
],
"spans": [
{
"bbox": [
105,
246,
507,
260
],
"score": 1.0,
"content": "composed of neural and probabilistic modules, or (depicted via slash symbol) only one of these two.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
257,
506,
270
],
"spans": [
{
"bbox": [
105,
257,
506,
270
],
"score": 1.0,
"content": "(b) A minimal SLASH program and query for the set prediction task, here only showing the NPP that",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
268,
447,
281
],
"spans": [
{
"bbox": [
105,
268,
447,
281
],
"score": 1.0,
"content": "models the color category per object. For the full program, we refer to the Appendix.",
"type": "text"
}
],
"index": 18
}
],
"index": 16
}
],
"index": 12.5
},
{
"type": "title",
"bbox": [
108,
302,
294,
313
],
"lines": [
{
"bbox": [
106,
302,
295,
315
],
"spans": [
{
"bbox": [
106,
302,
295,
315
],
"score": 1.0,
"content": "3.1 NEURAL-PROBABILISTIC PREDICATES",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
322,
505,
378
],
"lines": [
{
"bbox": [
105,
322,
505,
336
],
"spans": [
{
"bbox": [
105,
322,
505,
336
],
"score": 1.0,
"content": "Previous DPPLs, DeepProbLog (Manhaeve et al., 2018) and NeurASP (Yang et al., 2020), introduced",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
333,
505,
347
],
"spans": [
{
"bbox": [
105,
333,
505,
347
],
"score": 1.0,
"content": "the Neural Predicate as an annotated-disjunction or as a propositional atom, respectively, to acquire",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
343,
506,
359
],
"spans": [
{
"bbox": [
105,
343,
237,
359
],
"score": 1.0,
"content": "conditional class probabilities,",
"type": "text"
},
{
"bbox": [
237,
345,
273,
357
],
"score": 0.93,
"content": "P ( C | X )",
"type": "inline_equation"
},
{
"bbox": [
273,
343,
506,
359
],
"score": 1.0,
"content": ", via the softmax function at the output of an arbitrary",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
355,
505,
369
],
"spans": [
{
"bbox": [
105,
355,
505,
369
],
"score": 1.0,
"content": "DNN. As mentioned in the introduction, this approach has certain limitations concerning inference",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
366,
463,
380
],
"spans": [
{
"bbox": [
106,
366,
463,
380
],
"score": 1.0,
"content": "capabilities. To resolve this issue, we introduce Neural-Probabilisitic Predicates (NPPs).",
"type": "text"
}
],
"index": 24
}
],
"index": 22,
"bbox_fs": [
105,
322,
506,
380
]
},
{
"type": "text",
"bbox": [
107,
384,
210,
395
],
"lines": [
{
"bbox": [
106,
383,
210,
396
],
"spans": [
{
"bbox": [
106,
383,
210,
396
],
"score": 1.0,
"content": "Formally, we denote with",
"type": "text"
}
],
"index": 25
}
],
"index": 25,
"bbox_fs": [
106,
383,
210,
396
]
},
{
"type": "interline_equation",
"bbox": [
256,
395,
354,
408
],
"lines": [
{
"bbox": [
256,
395,
354,
408
],
"spans": [
{
"bbox": [
256,
395,
354,
408
],
"score": 0.91,
"content": "n p p \\left( h ( x ) , [ v _ { 1 } , . . . , v _ { n } ] \\right)",
"type": "interline_equation",
"image_path": "187dcc297a855ad44e5c575f0d91e7583f0c038c9c3016485793aabe684f440a.jpg"
}
]
}
],
"index": 26,
"virtual_lines": [
{
"bbox": [
256,
395,
354,
408
],
"spans": [],
"index": 26
}
]
},
{
"type": "text",
"bbox": [
107,
411,
505,
466
],
"lines": [
{
"bbox": [
105,
411,
506,
425
],
"spans": [
{
"bbox": [
105,
411,
240,
425
],
"score": 1.0,
"content": "a Neural-Probabilistic Predicate",
"type": "text"
},
{
"bbox": [
241,
412,
248,
422
],
"score": 0.8,
"content": "h",
"type": "inline_equation"
},
{
"bbox": [
248,
411,
489,
425
],
"score": 1.0,
"content": ". Thereby, (i) npp is a reserved word to label an NPP, (ii)",
"type": "text"
},
{
"bbox": [
489,
412,
497,
422
],
"score": 0.74,
"content": "h",
"type": "inline_equation"
},
{
"bbox": [
497,
411,
506,
425
],
"score": 1.0,
"content": "a",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
422,
505,
435
],
"spans": [
{
"bbox": [
105,
422,
505,
435
],
"score": 1.0,
"content": "symbolic name of either a PC, NN or a joint of a PC and NN (cf. Fig. 2a), e.g., color_attr is the name",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
433,
506,
447
],
"spans": [
{
"bbox": [
105,
433,
267,
447
],
"score": 1.0,
"content": "of an NPP of Fig. 2b. Additionally, (iii)",
"type": "text"
},
{
"bbox": [
267,
435,
275,
444
],
"score": 0.79,
"content": "x",
"type": "inline_equation"
},
{
"bbox": [
275,
433,
379,
447
],
"score": 1.0,
"content": "denotes a “term” and (iv)",
"type": "text"
},
{
"bbox": [
379,
435,
422,
445
],
"score": 0.9,
"content": "v _ { 1 } , \\ldots , v _ { n }",
"type": "inline_equation"
},
{
"bbox": [
422,
433,
506,
447
],
"score": 1.0,
"content": "are placeholders for",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
444,
505,
457
],
"spans": [
{
"bbox": [
105,
444,
154,
457
],
"score": 1.0,
"content": "each of the",
"type": "text"
},
{
"bbox": [
155,
446,
162,
455
],
"score": 0.71,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
163,
444,
253,
457
],
"score": 1.0,
"content": "possible outcomes of",
"type": "text"
},
{
"bbox": [
253,
445,
260,
455
],
"score": 0.82,
"content": "h",
"type": "inline_equation"
},
{
"bbox": [
260,
444,
505,
457
],
"score": 1.0,
"content": ". For example, the placeholders for color_attr are the color",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
455,
294,
468
],
"spans": [
{
"bbox": [
106,
455,
294,
468
],
"score": 1.0,
"content": "attributes of an object (Red, Blue, Green, etc.).",
"type": "text"
}
],
"index": 31
}
],
"index": 29,
"bbox_fs": [
105,
411,
506,
468
]
},
{
"type": "text",
"bbox": [
107,
467,
505,
511
],
"lines": [
{
"bbox": [
105,
466,
506,
479
],
"spans": [
{
"bbox": [
105,
466,
321,
479
],
"score": 1.0,
"content": "An NPP abbreviates an arithmetic literal of the form",
"type": "text"
},
{
"bbox": [
321,
468,
345,
477
],
"score": 0.88,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
346,
466,
367,
479
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
367,
466,
414,
479
],
"score": 0.92,
"content": "c \\in \\{ h ( x ) \\}",
"type": "inline_equation"
},
{
"bbox": [
414,
466,
433,
479
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
433,
466,
503,
479
],
"score": 0.92,
"content": "v \\in \\{ v _ { 1 } , \\ldots , v _ { n } \\}",
"type": "inline_equation"
},
{
"bbox": [
503,
466,
506,
479
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
477,
506,
489
],
"spans": [
{
"bbox": [
105,
477,
224,
489
],
"score": 1.0,
"content": "Furthermore, we denote with",
"type": "text"
},
{
"bbox": [
224,
478,
246,
488
],
"score": 0.88,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
246,
477,
433,
489
],
"score": 1.0,
"content": "a set of NPPs of the form stated in (Eq. 1) and",
"type": "text"
},
{
"bbox": [
433,
479,
452,
487
],
"score": 0.87,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
452,
477,
506,
489
],
"score": 1.0,
"content": "the set of all",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
104,
486,
507,
502
],
"spans": [
{
"bbox": [
104,
486,
128,
502
],
"score": 1.0,
"content": "rules",
"type": "text"
},
{
"bbox": [
129,
490,
153,
498
],
"score": 0.86,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
153,
486,
463,
502
],
"score": 1.0,
"content": "of one NPP, which denotes the possible outcomes, obtained from an NPP in",
"type": "text"
},
{
"bbox": [
464,
489,
485,
499
],
"score": 0.88,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
486,
486,
507,
502
],
"score": 1.0,
"content": ", e.g.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
496,
460,
514
],
"spans": [
{
"bbox": [
106,
499,
313,
512
],
"score": 0.88,
"content": "r ^ { c o l o r \\_ a t t r } = \\{ c = R e d , c = B l u e , c = G r \\bar { e } e n , . . . \\}",
"type": "inline_equation"
},
{
"bbox": [
314,
496,
460,
514
],
"score": 1.0,
"content": "for the example depicted in Fig. 2b.",
"type": "text"
}
],
"index": 35
}
],
"index": 33.5,
"bbox_fs": [
104,
466,
507,
514
]
},
{
"type": "text",
"bbox": [
106,
515,
505,
561
],
"lines": [
{
"bbox": [
105,
515,
506,
529
],
"spans": [
{
"bbox": [
105,
515,
181,
529
],
"score": 1.0,
"content": "Rules of the form",
"type": "text"
},
{
"bbox": [
182,
516,
317,
528
],
"score": 0.86,
"content": "n p p ( h ( x ) , [ v _ { 1 } , \\ldots , v _ { n } ] ) B o d y",
"type": "inline_equation"
},
{
"bbox": [
317,
515,
506,
529
],
"score": 1.0,
"content": "are used as an abbreviation for application to",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
528,
505,
540
],
"spans": [
{
"bbox": [
106,
528,
505,
540
],
"score": 1.0,
"content": "multiple entities, e.g. multiple slots for the task of set prediction (cf. Fig. 2b). Hereby, Body of the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
538,
505,
551
],
"spans": [
{
"bbox": [
105,
538,
188,
551
],
"score": 1.0,
"content": "rule is identified by",
"type": "text"
},
{
"bbox": [
189,
539,
198,
548
],
"score": 0.73,
"content": "\\top",
"type": "inline_equation"
},
{
"bbox": [
198,
538,
278,
551
],
"score": 1.0,
"content": "(tautology, true) or",
"type": "text"
},
{
"bbox": [
278,
539,
288,
549
],
"score": 0.74,
"content": "\\perp",
"type": "inline_equation"
},
{
"bbox": [
288,
538,
505,
551
],
"score": 1.0,
"content": "(contradiction, false) during grounding. Rules of the",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
547,
403,
562
],
"spans": [
{
"bbox": [
104,
547,
128,
562
],
"score": 1.0,
"content": "form",
"type": "text"
},
{
"bbox": [
128,
549,
186,
560
],
"score": 0.81,
"content": "H e a d \\gets B o d y",
"type": "inline_equation"
},
{
"bbox": [
187,
547,
208,
562
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
208,
550,
227,
559
],
"score": 0.87,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
228,
547,
378,
562
],
"score": 1.0,
"content": "appearing in Head are prohibited for",
"type": "text"
},
{
"bbox": [
378,
549,
399,
559
],
"score": 0.88,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
400,
547,
403,
562
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 39
}
],
"index": 37.5,
"bbox_fs": [
104,
515,
506,
562
]
},
{
"type": "text",
"bbox": [
108,
565,
504,
600
],
"lines": [
{
"bbox": [
105,
565,
506,
579
],
"spans": [
{
"bbox": [
105,
565,
506,
579
],
"score": 1.0,
"content": "In this work, we largely make use of NPPs that contain probabilistic circuits (specifically SPNs)",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
577,
506,
590
],
"spans": [
{
"bbox": [
106,
577,
506,
590
],
"score": 1.0,
"content": "which allow for tractable density estimation and modelling of joint probabilities. In this way, it is",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
588,
502,
601
],
"spans": [
{
"bbox": [
106,
588,
365,
601
],
"score": 1.0,
"content": "possible to answer a much richer set of probabilistic queries, i.e.",
"type": "text"
},
{
"bbox": [
366,
588,
403,
600
],
"score": 0.85,
"content": "P ( X , C )",
"type": "inline_equation"
},
{
"bbox": [
403,
588,
407,
601
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
407,
588,
443,
600
],
"score": 0.87,
"content": "P ( X | C )",
"type": "inline_equation"
},
{
"bbox": [
443,
588,
461,
601
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
461,
588,
497,
600
],
"score": 0.92,
"content": "P ( \\bar { C } | X )",
"type": "inline_equation"
},
{
"bbox": [
498,
588,
502,
601
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 42
}
],
"index": 41,
"bbox_fs": [
105,
565,
506,
601
]
},
{
"type": "text",
"bbox": [
107,
604,
505,
671
],
"lines": [
{
"bbox": [
105,
605,
505,
618
],
"spans": [
{
"bbox": [
105,
605,
505,
618
],
"score": 1.0,
"content": "In addition to this, we introduce the arguably more interesting type of NPP that combines a neural",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
616,
505,
628
],
"spans": [
{
"bbox": [
105,
616,
505,
628
],
"score": 1.0,
"content": "module with a PC. Hereby, the neural module learns to map the raw input data into an optimal",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
627,
505,
639
],
"spans": [
{
"bbox": [
105,
627,
505,
639
],
"score": 1.0,
"content": "latent representation, e.g. object-based slot representations. The PC, in turn, learns to model the",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
638,
506,
651
],
"spans": [
{
"bbox": [
105,
638,
506,
651
],
"score": 1.0,
"content": "joint distribution of these latent variables and produces the final probability estimates. This type of",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
648,
505,
661
],
"spans": [
{
"bbox": [
105,
648,
505,
661
],
"score": 1.0,
"content": "NPP nicely combines the representational power of neural networks with the advantages of PCs in",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
660,
282,
673
],
"spans": [
{
"bbox": [
105,
660,
282,
673
],
"score": 1.0,
"content": "probability estimation and query flexibility.",
"type": "text"
}
],
"index": 48
}
],
"index": 45.5,
"bbox_fs": [
105,
605,
506,
673
]
},
{
"type": "text",
"bbox": [
107,
676,
505,
732
],
"lines": [
{
"bbox": [
105,
676,
505,
690
],
"spans": [
{
"bbox": [
105,
676,
505,
690
],
"score": 1.0,
"content": "For making the different probabilistic queries distinguishable in a SLASH program, we introduce the",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
318,
700
],
"score": 1.0,
"content": "following notation. We denote a given variable with",
"type": "text"
},
{
"bbox": [
319,
689,
328,
698
],
"score": 0.69,
"content": "^ +",
"type": "inline_equation"
},
{
"bbox": [
329,
687,
505,
700
],
"score": 1.0,
"content": "and the query variable with −. E.g., within",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
697,
506,
713
],
"spans": [
{
"bbox": [
105,
697,
262,
713
],
"score": 1.0,
"content": "the running example of set prediction",
"type": "text"
},
{
"bbox": [
263,
699,
273,
710
],
"score": 0.31,
"content": "\\cdot e f .",
"type": "inline_equation"
},
{
"bbox": [
273,
697,
441,
713
],
"score": 1.0,
"content": "Fig. 1 and 2b), with the query color_attr",
"type": "text"
},
{
"bbox": [
442,
699,
486,
711
],
"score": 0.87,
"content": "\\cdot ( + X , - C )",
"type": "inline_equation"
},
{
"bbox": [
487,
697,
506,
713
],
"score": 1.0,
"content": "one",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
708,
507,
723
],
"spans": [
{
"bbox": [
105,
708,
158,
723
],
"score": 1.0,
"content": "is asking for",
"type": "text"
},
{
"bbox": [
158,
710,
194,
721
],
"score": 0.91,
"content": "P ( C | X )",
"type": "inline_equation"
},
{
"bbox": [
195,
708,
305,
723
],
"score": 1.0,
"content": ". Similarly, with color_attr",
"type": "text"
},
{
"bbox": [
305,
711,
349,
722
],
"score": 0.86,
"content": "( - X , + C )",
"type": "inline_equation"
},
{
"bbox": [
349,
708,
420,
723
],
"score": 1.0,
"content": "one is asking for",
"type": "text"
},
{
"bbox": [
420,
710,
456,
722
],
"score": 0.93,
"content": "P ( X | C )",
"type": "inline_equation"
},
{
"bbox": [
456,
708,
507,
723
],
"score": 1.0,
"content": "and, finally,",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
105,
719,
272,
734
],
"spans": [
{
"bbox": [
105,
719,
171,
734
],
"score": 1.0,
"content": "with color_attr",
"type": "text"
},
{
"bbox": [
171,
721,
215,
732
],
"score": 0.82,
"content": "( - X , - C )",
"type": "inline_equation"
},
{
"bbox": [
216,
719,
231,
734
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
232,
722,
268,
732
],
"score": 0.93,
"content": "P ( X , C )",
"type": "inline_equation"
},
{
"bbox": [
268,
719,
272,
734
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 53
}
],
"index": 51,
"bbox_fs": [
105,
676,
507,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
81,
505,
149
],
"lines": [
{
"bbox": [
105,
81,
506,
96
],
"spans": [
{
"bbox": [
105,
81,
506,
96
],
"score": 1.0,
"content": "To summarize, an NPP can consist of neural and/or probabilistic modules and produces query-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
505,
106
],
"spans": [
{
"bbox": [
105,
93,
505,
106
],
"score": 1.0,
"content": "dependent probability estimates. Due to the flexibility of its definition, the term NPP contains the",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
104,
104,
506,
118
],
"spans": [
{
"bbox": [
104,
104,
506,
118
],
"score": 1.0,
"content": "predicates of previous works (Manhaeve et al., 2018; Yang et al., 2020), but also more interesting",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
104,
115,
506,
128
],
"spans": [
{
"bbox": [
104,
115,
506,
128
],
"score": 1.0,
"content": "predicates discussed above. The specific “flavor” of an NPP should be chosen depending on what",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
126,
506,
139
],
"spans": [
{
"bbox": [
105,
126,
506,
139
],
"score": 1.0,
"content": "type of probability estimation is required (cf. Fig 2a). Lastly, NPPs have the unified loss function of",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
137,
218,
150
],
"spans": [
{
"bbox": [
106,
137,
218,
150
],
"score": 1.0,
"content": "the negative log-likelihood:",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5
},
{
"type": "interline_equation",
"bbox": [
111,
154,
504,
187
],
"lines": [
{
"bbox": [
111,
154,
504,
187
],
"spans": [
{
"bbox": [
111,
154,
504,
187
],
"score": 0.93,
"content": "L _ { N P P } : = - \\log L H ( x , \\hat { x } ) = \\sum _ { i = 1 } ^ { n } L H ( x _ { i } , \\hat { x } _ { i } ) = - \\sum _ { i = 1 } ^ { n } x _ { i } \\cdot \\log ( P _ { \\xi } ^ { ( X , C ) } ( x _ { i } ) ) = - \\sum _ { i = 1 } ^ { n } \\log ( P _ { \\xi } ^ { ( X , C ) } )",
"type": "interline_equation",
"image_path": "2df2b15aa55bfa392ca1eea024da8df01763335460315fe384a335f6d03243b3.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
111,
154,
504,
165.0
],
"spans": [],
"index": 6
},
{
"bbox": [
111,
165.0,
504,
176.0
],
"spans": [],
"index": 7
},
{
"bbox": [
111,
176.0,
504,
187.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "text",
"bbox": [
107,
195,
506,
222
],
"lines": [
{
"bbox": [
105,
194,
505,
208
],
"spans": [
{
"bbox": [
105,
194,
347,
208
],
"score": 1.0,
"content": "whereby we are assuming the data to be i.i.d., ground truth",
"type": "text"
},
{
"bbox": [
347,
197,
357,
207
],
"score": 0.85,
"content": "x _ { i }",
"type": "inline_equation"
},
{
"bbox": [
357,
194,
460,
208
],
"score": 1.0,
"content": "to be the all-ones vector,",
"type": "text"
},
{
"bbox": [
460,
196,
467,
207
],
"score": 0.85,
"content": "\\xi",
"type": "inline_equation"
},
{
"bbox": [
467,
194,
505,
208
],
"score": 1.0,
"content": "to be the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
101,
205,
506,
228
],
"spans": [
{
"bbox": [
101,
205,
216,
228
],
"score": 1.0,
"content": "parameters of the NPP and",
"type": "text"
},
{
"bbox": [
216,
207,
246,
223
],
"score": 0.94,
"content": "P _ { \\xi } ^ { ( X , C ) }",
"type": "inline_equation"
},
{
"bbox": [
247,
205,
324,
228
],
"score": 1.0,
"content": "are the predictions",
"type": "text"
},
{
"bbox": [
324,
210,
334,
220
],
"score": 0.88,
"content": "{ \\hat { x } } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
335,
205,
506,
228
],
"score": 1.0,
"content": "obtained from the PC encoded in the NPP.",
"type": "text"
}
],
"index": 10
}
],
"index": 9.5
},
{
"type": "title",
"bbox": [
107,
235,
313,
247
],
"lines": [
{
"bbox": [
106,
235,
313,
248
],
"spans": [
{
"bbox": [
106,
235,
313,
248
],
"score": 1.0,
"content": "3.2 THE SLASH LANGUAGE AND SEMANTICS",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "text",
"bbox": [
106,
255,
505,
301
],
"lines": [
{
"bbox": [
105,
256,
505,
269
],
"spans": [
{
"bbox": [
105,
256,
505,
269
],
"score": 1.0,
"content": "Fig. 1 presents an illustration of SLASH, exemplified for the task of set prediction, with all of its key",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
267,
505,
280
],
"spans": [
{
"bbox": [
105,
267,
505,
280
],
"score": 1.0,
"content": "components. Having introduced the NPPs previously, which produce probability estimates, we now",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
279,
505,
292
],
"spans": [
{
"bbox": [
105,
279,
505,
292
],
"score": 1.0,
"content": "continue in the pipeline on how to use these probability estimates for answering logical queries. We",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
289,
295,
302
],
"spans": [
{
"bbox": [
105,
289,
295,
302
],
"score": 1.0,
"content": "begin by formally defining a SLASH program.",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
106,
304,
505,
338
],
"lines": [
{
"bbox": [
105,
302,
507,
317
],
"spans": [
{
"bbox": [
105,
302,
324,
317
],
"score": 1.0,
"content": "Definition 1. A SLASH program Π is the union of",
"type": "text"
},
{
"bbox": [
324,
304,
345,
315
],
"score": 0.8,
"content": "\\Pi ^ { a s p }",
"type": "inline_equation"
},
{
"bbox": [
345,
302,
350,
317
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
351,
304,
372,
315
],
"score": 0.59,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
373,
302,
428,
317
],
"score": 1.0,
"content": ". Therewith,",
"type": "text"
},
{
"bbox": [
429,
304,
450,
315
],
"score": 0.85,
"content": "\\Pi ^ { a s p }",
"type": "inline_equation"
},
{
"bbox": [
451,
302,
507,
317
],
"score": 1.0,
"content": "is the set of",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
104,
313,
506,
328
],
"spans": [
{
"bbox": [
104,
313,
440,
328
],
"score": 1.0,
"content": "propositional rules (standard rules from ASP-Core-2 (Calimeri et al., 2020)), and",
"type": "text"
},
{
"bbox": [
441,
315,
462,
325
],
"score": 0.8,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
463,
313,
506,
328
],
"score": 1.0,
"content": "is a set of",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
326,
344,
339
],
"spans": [
{
"bbox": [
104,
326,
344,
339
],
"score": 1.0,
"content": "Neural-Probabilistic Predicates of the form stated in Eq. 1.",
"type": "text"
}
],
"index": 18
}
],
"index": 17
},
{
"type": "text",
"bbox": [
106,
346,
505,
414
],
"lines": [
{
"bbox": [
105,
347,
505,
360
],
"spans": [
{
"bbox": [
105,
347,
505,
360
],
"score": 1.0,
"content": "Fig. 2b depicts a minimal SLASH program for the task of set prediction, exemplifying a set of",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
358,
505,
371
],
"spans": [
{
"bbox": [
105,
358,
505,
371
],
"score": 1.0,
"content": "propositional rules and neural predicates. Similar to NeurASP, SLASH requires ASP and as such",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
370,
505,
381
],
"spans": [
{
"bbox": [
106,
370,
505,
381
],
"score": 1.0,
"content": "adopts its syntax to most part. We therefore now address integrating our NPPs into an ASP compatible",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
379,
505,
393
],
"spans": [
{
"bbox": [
105,
379,
505,
393
],
"score": 1.0,
"content": "form to obtain the success probability for the logical query given all possible solutions. Thus, we",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
391,
505,
404
],
"spans": [
{
"bbox": [
105,
391,
369,
404
],
"score": 1.0,
"content": "define SLASH’s semantics. For SLASH to translate the program",
"type": "text"
},
{
"bbox": [
369,
391,
378,
401
],
"score": 0.63,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
379,
391,
505,
404
],
"score": 1.0,
"content": "to the ASP-solver’s compatible",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
403,
339,
414
],
"spans": [
{
"bbox": [
106,
403,
339,
414
],
"score": 1.0,
"content": "form, the rules (Eq. 1) will be rewritten to the set of rules:",
"type": "text"
}
],
"index": 24
}
],
"index": 21.5
},
{
"type": "interline_equation",
"bbox": [
241,
419,
370,
433
],
"lines": [
{
"bbox": [
241,
419,
370,
433
],
"spans": [
{
"bbox": [
241,
419,
370,
433
],
"score": 0.93,
"content": "1 \\{ h ( x ) = v _ { 1 } ; \\ldots ; h ( x ) = v _ { n } \\} 1",
"type": "interline_equation",
"image_path": "e4c76dbbc6e4f0607f36cfe65c5c74b6f382d229457f6b2b9ff4996044984268.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
241,
419,
370,
433
],
"spans": [],
"index": 25
}
]
},
{
"type": "text",
"bbox": [
106,
438,
505,
516
],
"lines": [
{
"bbox": [
105,
438,
505,
450
],
"spans": [
{
"bbox": [
105,
438,
505,
450
],
"score": 1.0,
"content": "The ASP-solver should understand this as “Pick exactly one rule from the set”. After the translation",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
450,
505,
461
],
"spans": [
{
"bbox": [
106,
450,
326,
461
],
"score": 1.0,
"content": "is done, we can ask an ASP-solver for the solutions for",
"type": "text"
},
{
"bbox": [
326,
450,
335,
459
],
"score": 0.52,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
335,
450,
505,
461
],
"score": 1.0,
"content": ". We denote a set of ASP constraints in the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
107,
460,
505,
473
],
"spans": [
{
"bbox": [
107,
461,
163,
471
],
"score": 0.79,
"content": "\\mathrm { f o r m } \\gets B o d y",
"type": "inline_equation"
},
{
"bbox": [
163,
460,
209,
473
],
"score": 1.0,
"content": ", as queries",
"type": "text"
},
{
"bbox": [
210,
461,
219,
472
],
"score": 0.84,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
219,
460,
438,
473
],
"score": 1.0,
"content": "(annotation). and each of the solutions with respect to",
"type": "text"
},
{
"bbox": [
439,
461,
448,
472
],
"score": 0.85,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
448,
460,
505,
473
],
"score": 1.0,
"content": "as a potential",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
471,
505,
484
],
"spans": [
{
"bbox": [
105,
471,
143,
484
],
"score": 1.0,
"content": "solution,",
"type": "text"
},
{
"bbox": [
144,
472,
150,
481
],
"score": 0.73,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
150,
471,
324,
484
],
"score": 1.0,
"content": ", (referred to as stable model in ASP). With",
"type": "text"
},
{
"bbox": [
324,
471,
349,
483
],
"score": 0.92,
"content": "I | _ { r ^ { n _ { P } p } }",
"type": "inline_equation"
},
{
"bbox": [
349,
471,
477,
484
],
"score": 1.0,
"content": "we denote the projection of the",
"type": "text"
},
{
"bbox": [
477,
472,
483,
481
],
"score": 0.8,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
484,
471,
505,
484
],
"score": 1.0,
"content": "onto",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
107,
480,
503,
496
],
"spans": [
{
"bbox": [
107,
482,
126,
492
],
"score": 0.72,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
126,
480,
130,
496
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
131,
482,
204,
494
],
"score": 0.77,
"content": "N u m ( I | _ { r ^ { n p p } } , \\Pi ) \\ -",
"type": "inline_equation"
},
{
"bbox": [
204,
480,
412,
496
],
"score": 1.0,
"content": "– the number of the possible solutions of the program",
"type": "text"
},
{
"bbox": [
412,
483,
421,
492
],
"score": 0.67,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
421,
480,
478,
496
],
"score": 1.0,
"content": "agreeing with",
"type": "text"
},
{
"bbox": [
478,
483,
503,
494
],
"score": 0.91,
"content": "I | _ { r ^ { n p p } }",
"type": "inline_equation"
}
],
"index": 30
},
{
"bbox": [
105,
492,
505,
506
],
"spans": [
{
"bbox": [
105,
492,
119,
506
],
"score": 1.0,
"content": "on",
"type": "text"
},
{
"bbox": [
120,
494,
139,
504
],
"score": 0.88,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
140,
492,
419,
506
],
"score": 1.0,
"content": ". Because we aim to calculate the success probability of the query",
"type": "text"
},
{
"bbox": [
419,
494,
428,
505
],
"score": 0.84,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
428,
492,
505,
506
],
"score": 1.0,
"content": ", we formalize the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
504,
300,
516
],
"spans": [
{
"bbox": [
105,
504,
242,
516
],
"score": 1.0,
"content": "probability of a potential solution",
"type": "text"
},
{
"bbox": [
243,
505,
249,
514
],
"score": 0.78,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
249,
504,
300,
516
],
"score": 1.0,
"content": "beforehand.",
"type": "text"
}
],
"index": 32
}
],
"index": 29
},
{
"type": "text",
"bbox": [
106,
519,
504,
552
],
"lines": [
{
"bbox": [
105,
519,
505,
532
],
"spans": [
{
"bbox": [
105,
519,
363,
532
],
"score": 1.0,
"content": "Definition 2. We specify the probability of the potential solution,",
"type": "text"
},
{
"bbox": [
364,
520,
370,
529
],
"score": 0.4,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
370,
519,
437,
532
],
"score": 1.0,
"content": ", for the program",
"type": "text"
},
{
"bbox": [
438,
519,
446,
529
],
"score": 0.31,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
447,
519,
505,
532
],
"score": 1.0,
"content": "as the product",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
530,
504,
543
],
"spans": [
{
"bbox": [
105,
530,
237,
543
],
"score": 1.0,
"content": "of the probabilities of all atoms",
"type": "text"
},
{
"bbox": [
238,
532,
262,
540
],
"score": 0.86,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
263,
530,
274,
543
],
"score": 1.0,
"content": "in",
"type": "text"
},
{
"bbox": [
275,
531,
299,
542
],
"score": 0.92,
"content": "I | _ { r ^ { n p p } }",
"type": "inline_equation"
},
{
"bbox": [
300,
530,
495,
543
],
"score": 1.0,
"content": "divided by the number of potential solutions of",
"type": "text"
},
{
"bbox": [
495,
532,
504,
540
],
"score": 0.27,
"content": "\\Pi",
"type": "inline_equation"
}
],
"index": 34
},
{
"bbox": [
105,
540,
228,
554
],
"spans": [
{
"bbox": [
105,
540,
228,
554
],
"score": 1.0,
"content": "agreeing with I|rnpp on rnpp:",
"type": "text"
}
],
"index": 35
}
],
"index": 34
},
{
"type": "interline_equation",
"bbox": [
171,
558,
439,
592
],
"lines": [
{
"bbox": [
171,
558,
439,
592
],
"spans": [
{
"bbox": [
171,
558,
439,
592
],
"score": 0.92,
"content": "P _ { \\Pi } ( I ) = \\left\\{ \\begin{array} { l l } { \\frac { \\prod _ { c = v \\in { I \\vert _ { r } n p p } } P _ { \\Pi } ( c = v ) } { N u m ( I \\vert _ { r ^ { n p p } } , \\Pi ) } , } & { i f I i s a p o t e n t i a l s o l u t i o n o f \\Pi , } \\\\ { 0 , } & { o t h e r w i s e . } \\end{array} \\right.",
"type": "interline_equation",
"image_path": "8412f1d11746ec4046c38cac73f234346a59ec206ee4309a3774af105b0f2f9a.jpg"
}
]
}
],
"index": 37,
"virtual_lines": [
{
"bbox": [
171,
558,
439,
569.3333333333334
],
"spans": [],
"index": 36
},
{
"bbox": [
171,
569.3333333333334,
439,
580.6666666666667
],
"spans": [],
"index": 37
},
{
"bbox": [
171,
580.6666666666667,
439,
592.0000000000001
],
"spans": [],
"index": 38
}
]
},
{
"type": "text",
"bbox": [
106,
611,
361,
624
],
"lines": [
{
"bbox": [
106,
612,
360,
625
],
"spans": [
{
"bbox": [
106,
612,
360,
625
],
"score": 1.0,
"content": "Therefore, the probability of a query can be defined as follows.",
"type": "text"
}
],
"index": 39
}
],
"index": 39
},
{
"type": "text",
"bbox": [
105,
627,
484,
639
],
"lines": [
{
"bbox": [
105,
626,
484,
641
],
"spans": [
{
"bbox": [
105,
626,
277,
641
],
"score": 1.0,
"content": "Definition 3. The probability of the query",
"type": "text"
},
{
"bbox": [
277,
628,
287,
639
],
"score": 0.83,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
287,
626,
424,
641
],
"score": 1.0,
"content": "given the set of possible solutions",
"type": "text"
},
{
"bbox": [
424,
628,
430,
637
],
"score": 0.7,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
431,
626,
484,
641
],
"score": 1.0,
"content": "is defined as",
"type": "text"
}
],
"index": 40
}
],
"index": 40
},
{
"type": "interline_equation",
"bbox": [
258,
644,
352,
673
],
"lines": [
{
"bbox": [
258,
644,
352,
673
],
"spans": [
{
"bbox": [
258,
644,
352,
673
],
"score": 0.94,
"content": "P _ { \\Pi } ( Q ) : = \\sum _ { I \\ v { = } Q } P _ { \\Pi } ( I ) .",
"type": "interline_equation",
"image_path": "85f8048e7bb05e818b56885781d4a00f2c446c1eab1ad62d131afc2c89182051.jpg"
}
]
}
],
"index": 41.5,
"virtual_lines": [
{
"bbox": [
258,
644,
352,
658.5
],
"spans": [],
"index": 41
},
{
"bbox": [
258,
658.5,
352,
673.0
],
"spans": [],
"index": 42
}
]
},
{
"type": "text",
"bbox": [
106,
678,
506,
702
],
"lines": [
{
"bbox": [
105,
677,
506,
692
],
"spans": [
{
"bbox": [
105,
677,
142,
692
],
"score": 1.0,
"content": "Thereby,",
"type": "text"
},
{
"bbox": [
142,
679,
171,
690
],
"score": 0.91,
"content": "I \\models Q",
"type": "inline_equation"
},
{
"bbox": [
171,
677,
208,
692
],
"score": 1.0,
"content": "reads as",
"type": "text"
},
{
"bbox": [
209,
679,
219,
689
],
"score": 0.44,
"content": "^ { * } I",
"type": "inline_equation"
},
{
"bbox": [
220,
677,
254,
692
],
"score": 1.0,
"content": "satisfies",
"type": "text"
},
{
"bbox": [
254,
679,
264,
690
],
"score": 0.75,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
264,
677,
416,
692
],
"score": 1.0,
"content": "”. The probability of the set of queries",
"type": "text"
},
{
"bbox": [
416,
678,
495,
691
],
"score": 0.93,
"content": "\\mathbf { Q } = \\{ Q _ { 1 } , \\ldots , Q _ { l } \\}",
"type": "inline_equation"
},
{
"bbox": [
495,
677,
506,
692
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
690,
317,
703
],
"spans": [
{
"bbox": [
106,
690,
317,
703
],
"score": 1.0,
"content": "defined as the product of the probability of each. I.e.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5
},
{
"type": "interline_equation",
"bbox": [
211,
706,
400,
735
],
"lines": [
{
"bbox": [
211,
706,
400,
735
],
"spans": [
{
"bbox": [
211,
706,
400,
735
],
"score": 0.93,
"content": "P _ { \\Pi } \\left( \\mathbf { Q } \\right) : = \\prod _ { Q _ { i } \\in \\mathbf { Q } } P _ { \\Pi } ( Q _ { i } ) = \\prod _ { Q _ { i } \\in \\mathbf { Q } } \\sum _ { I \\ v { | } = Q } P _ { \\Pi } ( I ) .",
"type": "interline_equation",
"image_path": "8fbd6787ba9865c4142b45115e9ee1a16382065a884dc2a918eaf718c77fae02.jpg"
}
]
}
],
"index": 45.5,
"virtual_lines": [
{
"bbox": [
211,
706,
400,
720.5
],
"spans": [],
"index": 45
},
{
"bbox": [
211,
720.5,
400,
735.0
],
"spans": [],
"index": 46
}
]
}
],
"page_idx": 4,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "5",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
81,
505,
149
],
"lines": [
{
"bbox": [
105,
81,
506,
96
],
"spans": [
{
"bbox": [
105,
81,
506,
96
],
"score": 1.0,
"content": "To summarize, an NPP can consist of neural and/or probabilistic modules and produces query-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
505,
106
],
"spans": [
{
"bbox": [
105,
93,
505,
106
],
"score": 1.0,
"content": "dependent probability estimates. Due to the flexibility of its definition, the term NPP contains the",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
104,
104,
506,
118
],
"spans": [
{
"bbox": [
104,
104,
506,
118
],
"score": 1.0,
"content": "predicates of previous works (Manhaeve et al., 2018; Yang et al., 2020), but also more interesting",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
104,
115,
506,
128
],
"spans": [
{
"bbox": [
104,
115,
506,
128
],
"score": 1.0,
"content": "predicates discussed above. The specific “flavor” of an NPP should be chosen depending on what",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
126,
506,
139
],
"spans": [
{
"bbox": [
105,
126,
506,
139
],
"score": 1.0,
"content": "type of probability estimation is required (cf. Fig 2a). Lastly, NPPs have the unified loss function of",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
137,
218,
150
],
"spans": [
{
"bbox": [
106,
137,
218,
150
],
"score": 1.0,
"content": "the negative log-likelihood:",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5,
"bbox_fs": [
104,
81,
506,
150
]
},
{
"type": "interline_equation",
"bbox": [
111,
154,
504,
187
],
"lines": [
{
"bbox": [
111,
154,
504,
187
],
"spans": [
{
"bbox": [
111,
154,
504,
187
],
"score": 0.93,
"content": "L _ { N P P } : = - \\log L H ( x , \\hat { x } ) = \\sum _ { i = 1 } ^ { n } L H ( x _ { i } , \\hat { x } _ { i } ) = - \\sum _ { i = 1 } ^ { n } x _ { i } \\cdot \\log ( P _ { \\xi } ^ { ( X , C ) } ( x _ { i } ) ) = - \\sum _ { i = 1 } ^ { n } \\log ( P _ { \\xi } ^ { ( X , C ) } )",
"type": "interline_equation",
"image_path": "2df2b15aa55bfa392ca1eea024da8df01763335460315fe384a335f6d03243b3.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
111,
154,
504,
165.0
],
"spans": [],
"index": 6
},
{
"bbox": [
111,
165.0,
504,
176.0
],
"spans": [],
"index": 7
},
{
"bbox": [
111,
176.0,
504,
187.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "text",
"bbox": [
107,
195,
506,
222
],
"lines": [
{
"bbox": [
105,
194,
505,
208
],
"spans": [
{
"bbox": [
105,
194,
347,
208
],
"score": 1.0,
"content": "whereby we are assuming the data to be i.i.d., ground truth",
"type": "text"
},
{
"bbox": [
347,
197,
357,
207
],
"score": 0.85,
"content": "x _ { i }",
"type": "inline_equation"
},
{
"bbox": [
357,
194,
460,
208
],
"score": 1.0,
"content": "to be the all-ones vector,",
"type": "text"
},
{
"bbox": [
460,
196,
467,
207
],
"score": 0.85,
"content": "\\xi",
"type": "inline_equation"
},
{
"bbox": [
467,
194,
505,
208
],
"score": 1.0,
"content": "to be the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
101,
205,
506,
228
],
"spans": [
{
"bbox": [
101,
205,
216,
228
],
"score": 1.0,
"content": "parameters of the NPP and",
"type": "text"
},
{
"bbox": [
216,
207,
246,
223
],
"score": 0.94,
"content": "P _ { \\xi } ^ { ( X , C ) }",
"type": "inline_equation"
},
{
"bbox": [
247,
205,
324,
228
],
"score": 1.0,
"content": "are the predictions",
"type": "text"
},
{
"bbox": [
324,
210,
334,
220
],
"score": 0.88,
"content": "{ \\hat { x } } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
335,
205,
506,
228
],
"score": 1.0,
"content": "obtained from the PC encoded in the NPP.",
"type": "text"
}
],
"index": 10
}
],
"index": 9.5,
"bbox_fs": [
101,
194,
506,
228
]
},
{
"type": "title",
"bbox": [
107,
235,
313,
247
],
"lines": [
{
"bbox": [
106,
235,
313,
248
],
"spans": [
{
"bbox": [
106,
235,
313,
248
],
"score": 1.0,
"content": "3.2 THE SLASH LANGUAGE AND SEMANTICS",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "text",
"bbox": [
106,
255,
505,
301
],
"lines": [
{
"bbox": [
105,
256,
505,
269
],
"spans": [
{
"bbox": [
105,
256,
505,
269
],
"score": 1.0,
"content": "Fig. 1 presents an illustration of SLASH, exemplified for the task of set prediction, with all of its key",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
267,
505,
280
],
"spans": [
{
"bbox": [
105,
267,
505,
280
],
"score": 1.0,
"content": "components. Having introduced the NPPs previously, which produce probability estimates, we now",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
279,
505,
292
],
"spans": [
{
"bbox": [
105,
279,
505,
292
],
"score": 1.0,
"content": "continue in the pipeline on how to use these probability estimates for answering logical queries. We",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
289,
295,
302
],
"spans": [
{
"bbox": [
105,
289,
295,
302
],
"score": 1.0,
"content": "begin by formally defining a SLASH program.",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5,
"bbox_fs": [
105,
256,
505,
302
]
},
{
"type": "text",
"bbox": [
106,
304,
505,
338
],
"lines": [
{
"bbox": [
105,
302,
507,
317
],
"spans": [
{
"bbox": [
105,
302,
324,
317
],
"score": 1.0,
"content": "Definition 1. A SLASH program Π is the union of",
"type": "text"
},
{
"bbox": [
324,
304,
345,
315
],
"score": 0.8,
"content": "\\Pi ^ { a s p }",
"type": "inline_equation"
},
{
"bbox": [
345,
302,
350,
317
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
351,
304,
372,
315
],
"score": 0.59,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
373,
302,
428,
317
],
"score": 1.0,
"content": ". Therewith,",
"type": "text"
},
{
"bbox": [
429,
304,
450,
315
],
"score": 0.85,
"content": "\\Pi ^ { a s p }",
"type": "inline_equation"
},
{
"bbox": [
451,
302,
507,
317
],
"score": 1.0,
"content": "is the set of",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
104,
313,
506,
328
],
"spans": [
{
"bbox": [
104,
313,
440,
328
],
"score": 1.0,
"content": "propositional rules (standard rules from ASP-Core-2 (Calimeri et al., 2020)), and",
"type": "text"
},
{
"bbox": [
441,
315,
462,
325
],
"score": 0.8,
"content": "\\Pi ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
463,
313,
506,
328
],
"score": 1.0,
"content": "is a set of",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
326,
344,
339
],
"spans": [
{
"bbox": [
104,
326,
344,
339
],
"score": 1.0,
"content": "Neural-Probabilistic Predicates of the form stated in Eq. 1.",
"type": "text"
}
],
"index": 18
}
],
"index": 17,
"bbox_fs": [
104,
302,
507,
339
]
},
{
"type": "text",
"bbox": [
106,
346,
505,
414
],
"lines": [
{
"bbox": [
105,
347,
505,
360
],
"spans": [
{
"bbox": [
105,
347,
505,
360
],
"score": 1.0,
"content": "Fig. 2b depicts a minimal SLASH program for the task of set prediction, exemplifying a set of",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
358,
505,
371
],
"spans": [
{
"bbox": [
105,
358,
505,
371
],
"score": 1.0,
"content": "propositional rules and neural predicates. Similar to NeurASP, SLASH requires ASP and as such",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
370,
505,
381
],
"spans": [
{
"bbox": [
106,
370,
505,
381
],
"score": 1.0,
"content": "adopts its syntax to most part. We therefore now address integrating our NPPs into an ASP compatible",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
379,
505,
393
],
"spans": [
{
"bbox": [
105,
379,
505,
393
],
"score": 1.0,
"content": "form to obtain the success probability for the logical query given all possible solutions. Thus, we",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
391,
505,
404
],
"spans": [
{
"bbox": [
105,
391,
369,
404
],
"score": 1.0,
"content": "define SLASH’s semantics. For SLASH to translate the program",
"type": "text"
},
{
"bbox": [
369,
391,
378,
401
],
"score": 0.63,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
379,
391,
505,
404
],
"score": 1.0,
"content": "to the ASP-solver’s compatible",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
403,
339,
414
],
"spans": [
{
"bbox": [
106,
403,
339,
414
],
"score": 1.0,
"content": "form, the rules (Eq. 1) will be rewritten to the set of rules:",
"type": "text"
}
],
"index": 24
}
],
"index": 21.5,
"bbox_fs": [
105,
347,
505,
414
]
},
{
"type": "interline_equation",
"bbox": [
241,
419,
370,
433
],
"lines": [
{
"bbox": [
241,
419,
370,
433
],
"spans": [
{
"bbox": [
241,
419,
370,
433
],
"score": 0.93,
"content": "1 \\{ h ( x ) = v _ { 1 } ; \\ldots ; h ( x ) = v _ { n } \\} 1",
"type": "interline_equation",
"image_path": "e4c76dbbc6e4f0607f36cfe65c5c74b6f382d229457f6b2b9ff4996044984268.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
241,
419,
370,
433
],
"spans": [],
"index": 25
}
]
},
{
"type": "text",
"bbox": [
106,
438,
505,
516
],
"lines": [
{
"bbox": [
105,
438,
505,
450
],
"spans": [
{
"bbox": [
105,
438,
505,
450
],
"score": 1.0,
"content": "The ASP-solver should understand this as “Pick exactly one rule from the set”. After the translation",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
450,
505,
461
],
"spans": [
{
"bbox": [
106,
450,
326,
461
],
"score": 1.0,
"content": "is done, we can ask an ASP-solver for the solutions for",
"type": "text"
},
{
"bbox": [
326,
450,
335,
459
],
"score": 0.52,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
335,
450,
505,
461
],
"score": 1.0,
"content": ". We denote a set of ASP constraints in the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
107,
460,
505,
473
],
"spans": [
{
"bbox": [
107,
461,
163,
471
],
"score": 0.79,
"content": "\\mathrm { f o r m } \\gets B o d y",
"type": "inline_equation"
},
{
"bbox": [
163,
460,
209,
473
],
"score": 1.0,
"content": ", as queries",
"type": "text"
},
{
"bbox": [
210,
461,
219,
472
],
"score": 0.84,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
219,
460,
438,
473
],
"score": 1.0,
"content": "(annotation). and each of the solutions with respect to",
"type": "text"
},
{
"bbox": [
439,
461,
448,
472
],
"score": 0.85,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
448,
460,
505,
473
],
"score": 1.0,
"content": "as a potential",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
471,
505,
484
],
"spans": [
{
"bbox": [
105,
471,
143,
484
],
"score": 1.0,
"content": "solution,",
"type": "text"
},
{
"bbox": [
144,
472,
150,
481
],
"score": 0.73,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
150,
471,
324,
484
],
"score": 1.0,
"content": ", (referred to as stable model in ASP). With",
"type": "text"
},
{
"bbox": [
324,
471,
349,
483
],
"score": 0.92,
"content": "I | _ { r ^ { n _ { P } p } }",
"type": "inline_equation"
},
{
"bbox": [
349,
471,
477,
484
],
"score": 1.0,
"content": "we denote the projection of the",
"type": "text"
},
{
"bbox": [
477,
472,
483,
481
],
"score": 0.8,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
484,
471,
505,
484
],
"score": 1.0,
"content": "onto",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
107,
480,
503,
496
],
"spans": [
{
"bbox": [
107,
482,
126,
492
],
"score": 0.72,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
126,
480,
130,
496
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
131,
482,
204,
494
],
"score": 0.77,
"content": "N u m ( I | _ { r ^ { n p p } } , \\Pi ) \\ -",
"type": "inline_equation"
},
{
"bbox": [
204,
480,
412,
496
],
"score": 1.0,
"content": "– the number of the possible solutions of the program",
"type": "text"
},
{
"bbox": [
412,
483,
421,
492
],
"score": 0.67,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
421,
480,
478,
496
],
"score": 1.0,
"content": "agreeing with",
"type": "text"
},
{
"bbox": [
478,
483,
503,
494
],
"score": 0.91,
"content": "I | _ { r ^ { n p p } }",
"type": "inline_equation"
}
],
"index": 30
},
{
"bbox": [
105,
492,
505,
506
],
"spans": [
{
"bbox": [
105,
492,
119,
506
],
"score": 1.0,
"content": "on",
"type": "text"
},
{
"bbox": [
120,
494,
139,
504
],
"score": 0.88,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
140,
492,
419,
506
],
"score": 1.0,
"content": ". Because we aim to calculate the success probability of the query",
"type": "text"
},
{
"bbox": [
419,
494,
428,
505
],
"score": 0.84,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
428,
492,
505,
506
],
"score": 1.0,
"content": ", we formalize the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
504,
300,
516
],
"spans": [
{
"bbox": [
105,
504,
242,
516
],
"score": 1.0,
"content": "probability of a potential solution",
"type": "text"
},
{
"bbox": [
243,
505,
249,
514
],
"score": 0.78,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
249,
504,
300,
516
],
"score": 1.0,
"content": "beforehand.",
"type": "text"
}
],
"index": 32
}
],
"index": 29,
"bbox_fs": [
105,
438,
505,
516
]
},
{
"type": "text",
"bbox": [
106,
519,
504,
552
],
"lines": [
{
"bbox": [
105,
519,
505,
532
],
"spans": [
{
"bbox": [
105,
519,
363,
532
],
"score": 1.0,
"content": "Definition 2. We specify the probability of the potential solution,",
"type": "text"
},
{
"bbox": [
364,
520,
370,
529
],
"score": 0.4,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
370,
519,
437,
532
],
"score": 1.0,
"content": ", for the program",
"type": "text"
},
{
"bbox": [
438,
519,
446,
529
],
"score": 0.31,
"content": "\\Pi",
"type": "inline_equation"
},
{
"bbox": [
447,
519,
505,
532
],
"score": 1.0,
"content": "as the product",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
530,
504,
543
],
"spans": [
{
"bbox": [
105,
530,
237,
543
],
"score": 1.0,
"content": "of the probabilities of all atoms",
"type": "text"
},
{
"bbox": [
238,
532,
262,
540
],
"score": 0.86,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
263,
530,
274,
543
],
"score": 1.0,
"content": "in",
"type": "text"
},
{
"bbox": [
275,
531,
299,
542
],
"score": 0.92,
"content": "I | _ { r ^ { n p p } }",
"type": "inline_equation"
},
{
"bbox": [
300,
530,
495,
543
],
"score": 1.0,
"content": "divided by the number of potential solutions of",
"type": "text"
},
{
"bbox": [
495,
532,
504,
540
],
"score": 0.27,
"content": "\\Pi",
"type": "inline_equation"
}
],
"index": 34
},
{
"bbox": [
105,
540,
228,
554
],
"spans": [
{
"bbox": [
105,
540,
228,
554
],
"score": 1.0,
"content": "agreeing with I|rnpp on rnpp:",
"type": "text"
}
],
"index": 35
}
],
"index": 34,
"bbox_fs": [
105,
519,
505,
554
]
},
{
"type": "interline_equation",
"bbox": [
171,
558,
439,
592
],
"lines": [
{
"bbox": [
171,
558,
439,
592
],
"spans": [
{
"bbox": [
171,
558,
439,
592
],
"score": 0.92,
"content": "P _ { \\Pi } ( I ) = \\left\\{ \\begin{array} { l l } { \\frac { \\prod _ { c = v \\in { I \\vert _ { r } n p p } } P _ { \\Pi } ( c = v ) } { N u m ( I \\vert _ { r ^ { n p p } } , \\Pi ) } , } & { i f I i s a p o t e n t i a l s o l u t i o n o f \\Pi , } \\\\ { 0 , } & { o t h e r w i s e . } \\end{array} \\right.",
"type": "interline_equation",
"image_path": "8412f1d11746ec4046c38cac73f234346a59ec206ee4309a3774af105b0f2f9a.jpg"
}
]
}
],
"index": 37,
"virtual_lines": [
{
"bbox": [
171,
558,
439,
569.3333333333334
],
"spans": [],
"index": 36
},
{
"bbox": [
171,
569.3333333333334,
439,
580.6666666666667
],
"spans": [],
"index": 37
},
{
"bbox": [
171,
580.6666666666667,
439,
592.0000000000001
],
"spans": [],
"index": 38
}
]
},
{
"type": "text",
"bbox": [
106,
611,
361,
624
],
"lines": [
{
"bbox": [
106,
612,
360,
625
],
"spans": [
{
"bbox": [
106,
612,
360,
625
],
"score": 1.0,
"content": "Therefore, the probability of a query can be defined as follows.",
"type": "text"
}
],
"index": 39
}
],
"index": 39,
"bbox_fs": [
106,
612,
360,
625
]
},
{
"type": "text",
"bbox": [
105,
627,
484,
639
],
"lines": [
{
"bbox": [
105,
626,
484,
641
],
"spans": [
{
"bbox": [
105,
626,
277,
641
],
"score": 1.0,
"content": "Definition 3. The probability of the query",
"type": "text"
},
{
"bbox": [
277,
628,
287,
639
],
"score": 0.83,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
287,
626,
424,
641
],
"score": 1.0,
"content": "given the set of possible solutions",
"type": "text"
},
{
"bbox": [
424,
628,
430,
637
],
"score": 0.7,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
431,
626,
484,
641
],
"score": 1.0,
"content": "is defined as",
"type": "text"
}
],
"index": 40
}
],
"index": 40,
"bbox_fs": [
105,
626,
484,
641
]
},
{
"type": "interline_equation",
"bbox": [
258,
644,
352,
673
],
"lines": [
{
"bbox": [
258,
644,
352,
673
],
"spans": [
{
"bbox": [
258,
644,
352,
673
],
"score": 0.94,
"content": "P _ { \\Pi } ( Q ) : = \\sum _ { I \\ v { = } Q } P _ { \\Pi } ( I ) .",
"type": "interline_equation",
"image_path": "85f8048e7bb05e818b56885781d4a00f2c446c1eab1ad62d131afc2c89182051.jpg"
}
]
}
],
"index": 41.5,
"virtual_lines": [
{
"bbox": [
258,
644,
352,
658.5
],
"spans": [],
"index": 41
},
{
"bbox": [
258,
658.5,
352,
673.0
],
"spans": [],
"index": 42
}
]
},
{
"type": "text",
"bbox": [
106,
678,
506,
702
],
"lines": [
{
"bbox": [
105,
677,
506,
692
],
"spans": [
{
"bbox": [
105,
677,
142,
692
],
"score": 1.0,
"content": "Thereby,",
"type": "text"
},
{
"bbox": [
142,
679,
171,
690
],
"score": 0.91,
"content": "I \\models Q",
"type": "inline_equation"
},
{
"bbox": [
171,
677,
208,
692
],
"score": 1.0,
"content": "reads as",
"type": "text"
},
{
"bbox": [
209,
679,
219,
689
],
"score": 0.44,
"content": "^ { * } I",
"type": "inline_equation"
},
{
"bbox": [
220,
677,
254,
692
],
"score": 1.0,
"content": "satisfies",
"type": "text"
},
{
"bbox": [
254,
679,
264,
690
],
"score": 0.75,
"content": "Q",
"type": "inline_equation"
},
{
"bbox": [
264,
677,
416,
692
],
"score": 1.0,
"content": "”. The probability of the set of queries",
"type": "text"
},
{
"bbox": [
416,
678,
495,
691
],
"score": 0.93,
"content": "\\mathbf { Q } = \\{ Q _ { 1 } , \\ldots , Q _ { l } \\}",
"type": "inline_equation"
},
{
"bbox": [
495,
677,
506,
692
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
690,
317,
703
],
"spans": [
{
"bbox": [
106,
690,
317,
703
],
"score": 1.0,
"content": "defined as the product of the probability of each. I.e.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5,
"bbox_fs": [
105,
677,
506,
703
]
},
{
"type": "interline_equation",
"bbox": [
211,
706,
400,
735
],
"lines": [
{
"bbox": [
211,
706,
400,
735
],
"spans": [
{
"bbox": [
211,
706,
400,
735
],
"score": 0.93,
"content": "P _ { \\Pi } \\left( \\mathbf { Q } \\right) : = \\prod _ { Q _ { i } \\in \\mathbf { Q } } P _ { \\Pi } ( Q _ { i } ) = \\prod _ { Q _ { i } \\in \\mathbf { Q } } \\sum _ { I \\ v { | } = Q } P _ { \\Pi } ( I ) .",
"type": "interline_equation",
"image_path": "8fbd6787ba9865c4142b45115e9ee1a16382065a884dc2a918eaf718c77fae02.jpg"
}
]
}
],
"index": 45.5,
"virtual_lines": [
{
"bbox": [
211,
706,
400,
720.5
],
"spans": [],
"index": 45
},
{
"bbox": [
211,
720.5,
400,
735.0
],
"spans": [],
"index": 46
}
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
82,
282,
94
],
"lines": [
{
"bbox": [
106,
82,
283,
95
],
"spans": [
{
"bbox": [
106,
82,
283,
95
],
"score": 1.0,
"content": "3.3 PARAMETER LEARNING IN SLASH",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
102,
505,
170
],
"lines": [
{
"bbox": [
105,
102,
505,
116
],
"spans": [
{
"bbox": [
105,
102,
169,
116
],
"score": 1.0,
"content": "We denote with",
"type": "text"
},
{
"bbox": [
169,
103,
191,
115
],
"score": 0.86,
"content": "\\Pi ( \\pmb \\theta )",
"type": "inline_equation"
},
{
"bbox": [
192,
102,
390,
116
],
"score": 1.0,
"content": "the SLASH program under consideration, thereby",
"type": "text"
},
{
"bbox": [
390,
104,
397,
113
],
"score": 0.77,
"content": "\\pmb \\theta",
"type": "inline_equation"
},
{
"bbox": [
398,
102,
505,
116
],
"score": 1.0,
"content": "is the set of the parameters",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
114,
505,
127
],
"spans": [
{
"bbox": [
105,
114,
403,
127
],
"score": 1.0,
"content": "associated with Π. Further, making the i.i.d. assumption of the query set",
"type": "text"
},
{
"bbox": [
403,
114,
414,
126
],
"score": 0.49,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
414,
114,
505,
127
],
"score": 1.0,
"content": ", we follow Manhaeve",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
125,
505,
138
],
"spans": [
{
"bbox": [
106,
125,
505,
138
],
"score": 1.0,
"content": "et al. (2018) and Skryagin et al. (2020), and use the learning from entailment setting. That is, the",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
135,
505,
150
],
"spans": [
{
"bbox": [
105,
135,
460,
150
],
"score": 1.0,
"content": "training examples are logical queries that are known to be true in the SLASH program",
"type": "text"
},
{
"bbox": [
460,
136,
482,
148
],
"score": 0.74,
"content": "\\Pi ( \\theta )",
"type": "inline_equation"
},
{
"bbox": [
482,
135,
505,
150
],
"score": 1.0,
"content": ". The",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
146,
505,
161
],
"spans": [
{
"bbox": [
105,
146,
249,
161
],
"score": 1.0,
"content": "goal is now to learn the parameters",
"type": "text"
},
{
"bbox": [
249,
147,
256,
157
],
"score": 0.78,
"content": "\\pmb \\theta",
"type": "inline_equation"
},
{
"bbox": [
257,
146,
353,
161
],
"score": 1.0,
"content": "of the SLASH program",
"type": "text"
},
{
"bbox": [
353,
147,
375,
159
],
"score": 0.85,
"content": "\\Pi ( \\pmb \\theta )",
"type": "inline_equation"
},
{
"bbox": [
376,
146,
505,
161
],
"score": 1.0,
"content": "so that the observed queries are",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
156,
155,
172
],
"spans": [
{
"bbox": [
105,
156,
155,
172
],
"score": 1.0,
"content": "most likely.",
"type": "text"
}
],
"index": 6
}
],
"index": 3.5
},
{
"type": "text",
"bbox": [
107,
174,
505,
210
],
"lines": [
{
"bbox": [
105,
173,
505,
188
],
"spans": [
{
"bbox": [
105,
173,
505,
188
],
"score": 1.0,
"content": "To this end, we employ the negative log-likelihood and the cross-entropy of the observed queries",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
185,
507,
200
],
"spans": [
{
"bbox": [
106,
186,
150,
199
],
"score": 0.88,
"content": "P _ { \\mathrm { { I I } } ( \\theta ) } ( Q _ { i } )",
"type": "inline_equation"
},
{
"bbox": [
150,
185,
312,
200
],
"score": 1.0,
"content": "and their predicted probability value",
"type": "text"
},
{
"bbox": [
312,
186,
371,
199
],
"score": 0.93,
"content": "P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } )",
"type": "inline_equation"
},
{
"bbox": [
372,
185,
507,
200
],
"score": 1.0,
"content": ", assuming the NPPs are fixed:",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
200,
149,
210
],
"spans": [
{
"bbox": [
106,
200,
149,
210
],
"score": 0.78,
"content": "L _ { E N T } : =",
"type": "inline_equation"
}
],
"index": 9
}
],
"index": 8
},
{
"type": "interline_equation",
"bbox": [
109,
214,
490,
248
],
"lines": [
{
"bbox": [
109,
214,
490,
248
],
"spans": [
{
"bbox": [
109,
214,
490,
248
],
"score": 0.94,
"content": "- \\log L H \\left( \\log ( P _ { \\Pi ( \\theta ) } ( \\mathbf { Q } ) ) , P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { \\mathbf { Q } } ) \\right) = - \\sum _ { j = 1 } ^ { m } \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) .",
"type": "interline_equation",
"image_path": "73189bdbb324a7a5de4d157ea631cf184596aa216a9062f2a7d12c533ba1b87c.jpg"
}
]
}
],
"index": 11,
"virtual_lines": [
{
"bbox": [
109,
214,
490,
225.33333333333334
],
"spans": [],
"index": 10
},
{
"bbox": [
109,
225.33333333333334,
490,
236.66666666666669
],
"spans": [],
"index": 11
},
{
"bbox": [
109,
236.66666666666669,
490,
248.00000000000003
],
"spans": [],
"index": 12
}
]
},
{
"type": "text",
"bbox": [
106,
253,
506,
299
],
"lines": [
{
"bbox": [
106,
253,
505,
265
],
"spans": [
{
"bbox": [
106,
253,
505,
265
],
"score": 1.0,
"content": "This loss function aims at maximizing the estimated success probability. We remark that the defined",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
263,
506,
277
],
"spans": [
{
"bbox": [
105,
263,
506,
277
],
"score": 1.0,
"content": "loss function is true regardless of the NPP’s form (NN with Softmax, PC or PC jointly with NN). The",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
274,
506,
289
],
"spans": [
{
"bbox": [
105,
274,
282,
289
],
"score": 1.0,
"content": "only difference will be the second term, i.e.",
"type": "text"
},
{
"bbox": [
282,
276,
339,
289
],
"score": 0.92,
"content": "P ^ { ( C | X _ { \\mathbf { Q } } ) } ( x _ { \\mathbf { Q } } )",
"type": "inline_equation"
},
{
"bbox": [
339,
274,
351,
289
],
"score": 1.0,
"content": "or",
"type": "text"
},
{
"bbox": [
352,
275,
411,
289
],
"score": 0.91,
"content": "P ^ { ( X _ { \\mathbf { Q } } | C ) } ( x _ { \\mathbf { Q } } ) )",
"type": "inline_equation"
},
{
"bbox": [
411,
274,
506,
289
],
"score": 1.0,
"content": "depending on the NPP",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
288,
383,
299
],
"spans": [
{
"bbox": [
106,
288,
347,
299
],
"score": 1.0,
"content": "and task. Furthermore, we assume that for the set of queries",
"type": "text"
},
{
"bbox": [
348,
288,
357,
299
],
"score": 0.8,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
358,
288,
383,
299
],
"score": 1.0,
"content": "holds",
"type": "text"
}
],
"index": 16
}
],
"index": 14.5
},
{
"type": "interline_equation",
"bbox": [
251,
304,
360,
319
],
"lines": [
{
"bbox": [
251,
304,
360,
319
],
"spans": [
{
"bbox": [
251,
304,
360,
319
],
"score": 0.92,
"content": "P _ { \\Pi ( \\pmb \\theta ) } ( Q ) > 0 \\quad \\forall Q \\in { \\bf Q } .",
"type": "interline_equation",
"image_path": "38f649600323f2cd8117cf3828720d0ed67716c406e417f4cd2dd9149045732f.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
251,
304,
360,
319
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
107,
323,
505,
357
],
"lines": [
{
"bbox": [
105,
321,
506,
338
],
"spans": [
{
"bbox": [
105,
321,
398,
338
],
"score": 1.0,
"content": "In accordance with the semantics, we seek to reward the right solutions",
"type": "text"
},
{
"bbox": [
398,
326,
422,
334
],
"score": 0.86,
"content": "v = c",
"type": "inline_equation"
},
{
"bbox": [
423,
321,
506,
338
],
"score": 1.0,
"content": "and penalize wrong",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
334,
506,
348
],
"spans": [
{
"bbox": [
105,
334,
127,
348
],
"score": 1.0,
"content": "ones",
"type": "text"
},
{
"bbox": [
127,
335,
151,
347
],
"score": 0.9,
"content": "v \\neq c",
"type": "inline_equation"
},
{
"bbox": [
151,
334,
282,
348
],
"score": 1.0,
"content": ". Referring to the probabilities in",
"type": "text"
},
{
"bbox": [
282,
335,
302,
345
],
"score": 0.86,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
302,
334,
506,
348
],
"score": 1.0,
"content": "(the set of logical rules denoting NPPs, see Def. 2)",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
346,
384,
359
],
"spans": [
{
"bbox": [
105,
346,
117,
359
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
117,
348,
125,
357
],
"score": 0.69,
"content": "\\mathbf { p }",
"type": "inline_equation"
},
{
"bbox": [
125,
346,
281,
359
],
"score": 1.0,
"content": ", one can compute their gradients w.r.t.",
"type": "text"
},
{
"bbox": [
281,
347,
288,
356
],
"score": 0.8,
"content": "\\pmb { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
288,
346,
384,
359
],
"score": 1.0,
"content": "via backpropagation as",
"type": "text"
}
],
"index": 20
}
],
"index": 19
},
{
"type": "interline_equation",
"bbox": [
189,
362,
424,
397
],
"lines": [
{
"bbox": [
189,
362,
424,
397
],
"spans": [
{
"bbox": [
189,
362,
424,
397
],
"score": 0.94,
"content": "\\sum _ { Q \\in { \\bf Q } } \\frac { \\partial \\log \\left( P _ { \\Pi ( \\pmb \\theta ) } ( Q ) \\right) } { \\partial \\pmb \\theta } = \\sum _ { Q \\in { \\bf Q } } \\frac { \\partial \\log \\left( P _ { \\Pi ( \\pmb \\theta ) } ( Q ) \\right) } { \\partial { \\bf p } } \\times \\frac { \\partial { \\bf p } } { \\partial \\pmb \\theta } .",
"type": "interline_equation",
"image_path": "c71b1b57c842f235633b5397660cc117df3a103d2b93a4c5089f527d193524eb.jpg"
}
]
}
],
"index": 21.5,
"virtual_lines": [
{
"bbox": [
189,
362,
424,
379.5
],
"spans": [],
"index": 21
},
{
"bbox": [
189,
379.5,
424,
397.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "text",
"bbox": [
108,
403,
504,
445
],
"lines": [
{
"bbox": [
106,
400,
506,
419
],
"spans": [
{
"bbox": [
106,
400,
145,
419
],
"score": 1.0,
"content": "The term",
"type": "text"
},
{
"bbox": [
145,
402,
158,
417
],
"score": 0.9,
"content": "\\textstyle { \\frac { \\partial \\mathbf { p } } { \\partial \\theta } }",
"type": "inline_equation"
},
{
"bbox": [
158,
400,
506,
419
],
"score": 1.0,
"content": "can now be computed as usual via backward propagation through the NPPs (see Eq. 13",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
413,
504,
427
],
"spans": [
{
"bbox": [
105,
413,
263,
427
],
"score": 1.0,
"content": "in the appendix for details). By letting",
"type": "text"
},
{
"bbox": [
263,
417,
270,
426
],
"score": 0.79,
"content": "p",
"type": "inline_equation"
},
{
"bbox": [
270,
413,
448,
427
],
"score": 1.0,
"content": "to be the label of the probability of an atom",
"type": "text"
},
{
"bbox": [
449,
416,
473,
425
],
"score": 0.88,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
473,
413,
484,
427
],
"score": 1.0,
"content": "in",
"type": "text"
},
{
"bbox": [
484,
415,
504,
425
],
"score": 0.87,
"content": "r ^ { n p p }",
"type": "inline_equation"
}
],
"index": 24
},
{
"bbox": [
106,
426,
499,
446
],
"spans": [
{
"bbox": [
106,
431,
160,
444
],
"score": 1.0,
"content": "and denoting",
"type": "text"
},
{
"bbox": [
161,
430,
216,
444
],
"score": 0.91,
"content": "P _ { \\Pi ( \\pmb { \\theta } ) } ( c = v )",
"type": "inline_equation"
},
{
"bbox": [
216,
431,
255,
444
],
"score": 1.0,
"content": ", the term",
"type": "text"
},
{
"bbox": [
256,
426,
315,
446
],
"score": 0.94,
"content": "\\frac { \\partial \\log \\left( P _ { \\mathrm { I I } ( \\pmb { \\theta } ) } ( Q ) \\right) } { \\partial \\mathbf { p } }",
"type": "inline_equation"
},
{
"bbox": [
313,
429,
499,
444
],
"score": 1.0,
"content": "follows from NeurASP (Yang et al., 2020) as",
"type": "text"
}
],
"index": 25
}
],
"index": 24
},
{
"type": "interline_equation",
"bbox": [
171,
450,
440,
509
],
"lines": [
{
"bbox": [
171,
450,
440,
509
],
"spans": [
{
"bbox": [
171,
450,
440,
509
],
"score": 0.95,
"content": "\\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q ) \\right) } { \\partial \\mathbf { p } } = \\frac { I \\underset { I : I \\mid = Q } { \\sum } \\frac { P _ { \\Pi ( \\theta ) } ( I ) } { P _ { \\Pi ( \\theta ) } ( c = v ) } - \\underset { I : I , v ^ { \\prime } \\mid = Q } { \\sum } \\frac { P _ { \\Pi ( \\theta ) } ( I ) } { P _ { \\Pi ( \\theta ) } ( c = v ^ { \\prime } ) } } { \\underset { I : I \\mid = Q } { \\sum } P _ { \\Pi ( \\theta ) } ( I ) } .",
"type": "interline_equation",
"image_path": "16ead2545e3a3b0901ab36b2cc5263e0781b7c70ac9da887d966ab918f15d447.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
171,
450,
440,
469.6666666666667
],
"spans": [],
"index": 26
},
{
"bbox": [
171,
469.6666666666667,
440,
489.33333333333337
],
"spans": [],
"index": 27
},
{
"bbox": [
171,
489.33333333333337,
440,
509.00000000000006
],
"spans": [],
"index": 28
}
]
},
{
"type": "text",
"bbox": [
105,
520,
505,
542
],
"lines": [
{
"bbox": [
105,
519,
507,
533
],
"spans": [
{
"bbox": [
105,
519,
507,
533
],
"score": 1.0,
"content": "This is sensible. For instance, if a query to be true is not likely to be entailed, the gradient is positive.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
530,
318,
543
],
"spans": [
{
"bbox": [
106,
530,
318,
543
],
"score": 1.0,
"content": "Putting everything together, the final loss function is",
"type": "text"
}
],
"index": 30
}
],
"index": 29.5
},
{
"type": "interline_equation",
"bbox": [
246,
549,
364,
561
],
"lines": [
{
"bbox": [
246,
549,
364,
561
],
"spans": [
{
"bbox": [
246,
549,
364,
561
],
"score": 0.9,
"content": "{ \\cal L } _ { S L A S H } = { \\cal L } _ { N P P } + { \\cal L } _ { E N T }",
"type": "interline_equation",
"image_path": "f145bb6a4a4cfb00ec9261d4033d6ebfe1f151e47554c538aff178796d7c3820.jpg"
}
]
}
],
"index": 31,
"virtual_lines": [
{
"bbox": [
246,
549,
364,
561
],
"spans": [],
"index": 31
}
]
},
{
"type": "text",
"bbox": [
105,
568,
505,
590
],
"lines": [
{
"bbox": [
106,
568,
505,
580
],
"spans": [
{
"bbox": [
106,
568,
505,
580
],
"score": 1.0,
"content": "and we perform training using coordinate descent, i.e., we train the NPPs, the train the program with",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
578,
355,
591
],
"spans": [
{
"bbox": [
105,
578,
355,
591
],
"score": 1.0,
"content": "fixed NPPs, train the NPPs with the program fixed, and so on.",
"type": "text"
}
],
"index": 33
}
],
"index": 32.5
},
{
"type": "text",
"bbox": [
106,
595,
505,
640
],
"lines": [
{
"bbox": [
105,
595,
505,
607
],
"spans": [
{
"bbox": [
105,
595,
505,
607
],
"score": 1.0,
"content": "In hindsight, rather than requiring a novel loss function for each individual task and data set, with",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
606,
505,
619
],
"spans": [
{
"bbox": [
106,
606,
505,
619
],
"score": 1.0,
"content": "SLASH, it is possible to simply incorporate the specific requirements into the logic program. The",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
617,
505,
631
],
"spans": [
{
"bbox": [
105,
617,
505,
631
],
"score": 1.0,
"content": "training loss, however, remains the same. We refer to the Appendix A for further details, including",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
629,
264,
641
],
"spans": [
{
"bbox": [
106,
629,
264,
641
],
"score": 1.0,
"content": "the derivation of the total loss gradient.",
"type": "text"
}
],
"index": 37
}
],
"index": 35.5
},
{
"type": "title",
"bbox": [
108,
656,
234,
669
],
"lines": [
{
"bbox": [
105,
655,
235,
671
],
"spans": [
{
"bbox": [
105,
655,
235,
671
],
"score": 1.0,
"content": "4 EMPIRICAL RESULTS",
"type": "text"
}
],
"index": 38
}
],
"index": 38
},
{
"type": "text",
"bbox": [
106,
681,
503,
704
],
"lines": [
{
"bbox": [
105,
680,
505,
694
],
"spans": [
{
"bbox": [
105,
680,
505,
694
],
"score": 1.0,
"content": "The advantage of SLASH lies in the efficient integration of neural, probabilistic and symbolic",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
693,
443,
705
],
"spans": [
{
"bbox": [
106,
693,
443,
705
],
"score": 1.0,
"content": "computations. To emphasize this, we conduct a variety of experimental evaluations.",
"type": "text"
}
],
"index": 40
}
],
"index": 39.5
},
{
"type": "text",
"bbox": [
106,
709,
503,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Experimental Details. We use two benchmark data sets, namely MNIST (LeCun et al., 1998b) for",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
720,
505,
733
],
"spans": [
{
"bbox": [
106,
720,
505,
733
],
"score": 1.0,
"content": "the task of MNIST-Addition and a variant of the ShapeWorld data set (Kuhnle & Copestake, 2017) for",
"type": "text"
}
],
"index": 42
}
],
"index": 41.5
}
],
"page_idx": 5,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
26,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
308,
38
],
"spans": [
{
"bbox": [
106,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
302,
750,
310,
761
],
"spans": [
{
"bbox": [
302,
750,
310,
761
],
"score": 1.0,
"content": "6",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
82,
282,
94
],
"lines": [
{
"bbox": [
106,
82,
283,
95
],
"spans": [
{
"bbox": [
106,
82,
283,
95
],
"score": 1.0,
"content": "3.3 PARAMETER LEARNING IN SLASH",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
102,
505,
170
],
"lines": [
{
"bbox": [
105,
102,
505,
116
],
"spans": [
{
"bbox": [
105,
102,
169,
116
],
"score": 1.0,
"content": "We denote with",
"type": "text"
},
{
"bbox": [
169,
103,
191,
115
],
"score": 0.86,
"content": "\\Pi ( \\pmb \\theta )",
"type": "inline_equation"
},
{
"bbox": [
192,
102,
390,
116
],
"score": 1.0,
"content": "the SLASH program under consideration, thereby",
"type": "text"
},
{
"bbox": [
390,
104,
397,
113
],
"score": 0.77,
"content": "\\pmb \\theta",
"type": "inline_equation"
},
{
"bbox": [
398,
102,
505,
116
],
"score": 1.0,
"content": "is the set of the parameters",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
114,
505,
127
],
"spans": [
{
"bbox": [
105,
114,
403,
127
],
"score": 1.0,
"content": "associated with Π. Further, making the i.i.d. assumption of the query set",
"type": "text"
},
{
"bbox": [
403,
114,
414,
126
],
"score": 0.49,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
414,
114,
505,
127
],
"score": 1.0,
"content": ", we follow Manhaeve",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
125,
505,
138
],
"spans": [
{
"bbox": [
106,
125,
505,
138
],
"score": 1.0,
"content": "et al. (2018) and Skryagin et al. (2020), and use the learning from entailment setting. That is, the",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
135,
505,
150
],
"spans": [
{
"bbox": [
105,
135,
460,
150
],
"score": 1.0,
"content": "training examples are logical queries that are known to be true in the SLASH program",
"type": "text"
},
{
"bbox": [
460,
136,
482,
148
],
"score": 0.74,
"content": "\\Pi ( \\theta )",
"type": "inline_equation"
},
{
"bbox": [
482,
135,
505,
150
],
"score": 1.0,
"content": ". The",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
146,
505,
161
],
"spans": [
{
"bbox": [
105,
146,
249,
161
],
"score": 1.0,
"content": "goal is now to learn the parameters",
"type": "text"
},
{
"bbox": [
249,
147,
256,
157
],
"score": 0.78,
"content": "\\pmb \\theta",
"type": "inline_equation"
},
{
"bbox": [
257,
146,
353,
161
],
"score": 1.0,
"content": "of the SLASH program",
"type": "text"
},
{
"bbox": [
353,
147,
375,
159
],
"score": 0.85,
"content": "\\Pi ( \\pmb \\theta )",
"type": "inline_equation"
},
{
"bbox": [
376,
146,
505,
161
],
"score": 1.0,
"content": "so that the observed queries are",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
156,
155,
172
],
"spans": [
{
"bbox": [
105,
156,
155,
172
],
"score": 1.0,
"content": "most likely.",
"type": "text"
}
],
"index": 6
}
],
"index": 3.5,
"bbox_fs": [
105,
102,
505,
172
]
},
{
"type": "text",
"bbox": [
107,
174,
505,
210
],
"lines": [
{
"bbox": [
105,
173,
505,
188
],
"spans": [
{
"bbox": [
105,
173,
505,
188
],
"score": 1.0,
"content": "To this end, we employ the negative log-likelihood and the cross-entropy of the observed queries",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
185,
507,
200
],
"spans": [
{
"bbox": [
106,
186,
150,
199
],
"score": 0.88,
"content": "P _ { \\mathrm { { I I } } ( \\theta ) } ( Q _ { i } )",
"type": "inline_equation"
},
{
"bbox": [
150,
185,
312,
200
],
"score": 1.0,
"content": "and their predicted probability value",
"type": "text"
},
{
"bbox": [
312,
186,
371,
199
],
"score": 0.93,
"content": "P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } )",
"type": "inline_equation"
},
{
"bbox": [
372,
185,
507,
200
],
"score": 1.0,
"content": ", assuming the NPPs are fixed:",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
200,
149,
210
],
"spans": [
{
"bbox": [
106,
200,
149,
210
],
"score": 0.78,
"content": "L _ { E N T } : =",
"type": "inline_equation"
}
],
"index": 9
}
],
"index": 8,
"bbox_fs": [
105,
173,
507,
210
]
},
{
"type": "interline_equation",
"bbox": [
109,
214,
490,
248
],
"lines": [
{
"bbox": [
109,
214,
490,
248
],
"spans": [
{
"bbox": [
109,
214,
490,
248
],
"score": 0.94,
"content": "- \\log L H \\left( \\log ( P _ { \\Pi ( \\theta ) } ( \\mathbf { Q } ) ) , P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { \\mathbf { Q } } ) \\right) = - \\sum _ { j = 1 } ^ { m } \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) .",
"type": "interline_equation",
"image_path": "73189bdbb324a7a5de4d157ea631cf184596aa216a9062f2a7d12c533ba1b87c.jpg"
}
]
}
],
"index": 11,
"virtual_lines": [
{
"bbox": [
109,
214,
490,
225.33333333333334
],
"spans": [],
"index": 10
},
{
"bbox": [
109,
225.33333333333334,
490,
236.66666666666669
],
"spans": [],
"index": 11
},
{
"bbox": [
109,
236.66666666666669,
490,
248.00000000000003
],
"spans": [],
"index": 12
}
]
},
{
"type": "text",
"bbox": [
106,
253,
506,
299
],
"lines": [
{
"bbox": [
106,
253,
505,
265
],
"spans": [
{
"bbox": [
106,
253,
505,
265
],
"score": 1.0,
"content": "This loss function aims at maximizing the estimated success probability. We remark that the defined",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
263,
506,
277
],
"spans": [
{
"bbox": [
105,
263,
506,
277
],
"score": 1.0,
"content": "loss function is true regardless of the NPP’s form (NN with Softmax, PC or PC jointly with NN). The",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
274,
506,
289
],
"spans": [
{
"bbox": [
105,
274,
282,
289
],
"score": 1.0,
"content": "only difference will be the second term, i.e.",
"type": "text"
},
{
"bbox": [
282,
276,
339,
289
],
"score": 0.92,
"content": "P ^ { ( C | X _ { \\mathbf { Q } } ) } ( x _ { \\mathbf { Q } } )",
"type": "inline_equation"
},
{
"bbox": [
339,
274,
351,
289
],
"score": 1.0,
"content": "or",
"type": "text"
},
{
"bbox": [
352,
275,
411,
289
],
"score": 0.91,
"content": "P ^ { ( X _ { \\mathbf { Q } } | C ) } ( x _ { \\mathbf { Q } } ) )",
"type": "inline_equation"
},
{
"bbox": [
411,
274,
506,
289
],
"score": 1.0,
"content": "depending on the NPP",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
288,
383,
299
],
"spans": [
{
"bbox": [
106,
288,
347,
299
],
"score": 1.0,
"content": "and task. Furthermore, we assume that for the set of queries",
"type": "text"
},
{
"bbox": [
348,
288,
357,
299
],
"score": 0.8,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
358,
288,
383,
299
],
"score": 1.0,
"content": "holds",
"type": "text"
}
],
"index": 16
}
],
"index": 14.5,
"bbox_fs": [
105,
253,
506,
299
]
},
{
"type": "interline_equation",
"bbox": [
251,
304,
360,
319
],
"lines": [
{
"bbox": [
251,
304,
360,
319
],
"spans": [
{
"bbox": [
251,
304,
360,
319
],
"score": 0.92,
"content": "P _ { \\Pi ( \\pmb \\theta ) } ( Q ) > 0 \\quad \\forall Q \\in { \\bf Q } .",
"type": "interline_equation",
"image_path": "38f649600323f2cd8117cf3828720d0ed67716c406e417f4cd2dd9149045732f.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
251,
304,
360,
319
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
107,
323,
505,
357
],
"lines": [
{
"bbox": [
105,
321,
506,
338
],
"spans": [
{
"bbox": [
105,
321,
398,
338
],
"score": 1.0,
"content": "In accordance with the semantics, we seek to reward the right solutions",
"type": "text"
},
{
"bbox": [
398,
326,
422,
334
],
"score": 0.86,
"content": "v = c",
"type": "inline_equation"
},
{
"bbox": [
423,
321,
506,
338
],
"score": 1.0,
"content": "and penalize wrong",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
334,
506,
348
],
"spans": [
{
"bbox": [
105,
334,
127,
348
],
"score": 1.0,
"content": "ones",
"type": "text"
},
{
"bbox": [
127,
335,
151,
347
],
"score": 0.9,
"content": "v \\neq c",
"type": "inline_equation"
},
{
"bbox": [
151,
334,
282,
348
],
"score": 1.0,
"content": ". Referring to the probabilities in",
"type": "text"
},
{
"bbox": [
282,
335,
302,
345
],
"score": 0.86,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
302,
334,
506,
348
],
"score": 1.0,
"content": "(the set of logical rules denoting NPPs, see Def. 2)",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
346,
384,
359
],
"spans": [
{
"bbox": [
105,
346,
117,
359
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
117,
348,
125,
357
],
"score": 0.69,
"content": "\\mathbf { p }",
"type": "inline_equation"
},
{
"bbox": [
125,
346,
281,
359
],
"score": 1.0,
"content": ", one can compute their gradients w.r.t.",
"type": "text"
},
{
"bbox": [
281,
347,
288,
356
],
"score": 0.8,
"content": "\\pmb { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
288,
346,
384,
359
],
"score": 1.0,
"content": "via backpropagation as",
"type": "text"
}
],
"index": 20
}
],
"index": 19,
"bbox_fs": [
105,
321,
506,
359
]
},
{
"type": "interline_equation",
"bbox": [
189,
362,
424,
397
],
"lines": [
{
"bbox": [
189,
362,
424,
397
],
"spans": [
{
"bbox": [
189,
362,
424,
397
],
"score": 0.94,
"content": "\\sum _ { Q \\in { \\bf Q } } \\frac { \\partial \\log \\left( P _ { \\Pi ( \\pmb \\theta ) } ( Q ) \\right) } { \\partial \\pmb \\theta } = \\sum _ { Q \\in { \\bf Q } } \\frac { \\partial \\log \\left( P _ { \\Pi ( \\pmb \\theta ) } ( Q ) \\right) } { \\partial { \\bf p } } \\times \\frac { \\partial { \\bf p } } { \\partial \\pmb \\theta } .",
"type": "interline_equation",
"image_path": "c71b1b57c842f235633b5397660cc117df3a103d2b93a4c5089f527d193524eb.jpg"
}
]
}
],
"index": 21.5,
"virtual_lines": [
{
"bbox": [
189,
362,
424,
379.5
],
"spans": [],
"index": 21
},
{
"bbox": [
189,
379.5,
424,
397.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "text",
"bbox": [
108,
403,
504,
445
],
"lines": [
{
"bbox": [
106,
400,
506,
419
],
"spans": [
{
"bbox": [
106,
400,
145,
419
],
"score": 1.0,
"content": "The term",
"type": "text"
},
{
"bbox": [
145,
402,
158,
417
],
"score": 0.9,
"content": "\\textstyle { \\frac { \\partial \\mathbf { p } } { \\partial \\theta } }",
"type": "inline_equation"
},
{
"bbox": [
158,
400,
506,
419
],
"score": 1.0,
"content": "can now be computed as usual via backward propagation through the NPPs (see Eq. 13",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
413,
504,
427
],
"spans": [
{
"bbox": [
105,
413,
263,
427
],
"score": 1.0,
"content": "in the appendix for details). By letting",
"type": "text"
},
{
"bbox": [
263,
417,
270,
426
],
"score": 0.79,
"content": "p",
"type": "inline_equation"
},
{
"bbox": [
270,
413,
448,
427
],
"score": 1.0,
"content": "to be the label of the probability of an atom",
"type": "text"
},
{
"bbox": [
449,
416,
473,
425
],
"score": 0.88,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
473,
413,
484,
427
],
"score": 1.0,
"content": "in",
"type": "text"
},
{
"bbox": [
484,
415,
504,
425
],
"score": 0.87,
"content": "r ^ { n p p }",
"type": "inline_equation"
}
],
"index": 24
},
{
"bbox": [
106,
426,
499,
446
],
"spans": [
{
"bbox": [
106,
431,
160,
444
],
"score": 1.0,
"content": "and denoting",
"type": "text"
},
{
"bbox": [
161,
430,
216,
444
],
"score": 0.91,
"content": "P _ { \\Pi ( \\pmb { \\theta } ) } ( c = v )",
"type": "inline_equation"
},
{
"bbox": [
216,
431,
255,
444
],
"score": 1.0,
"content": ", the term",
"type": "text"
},
{
"bbox": [
256,
426,
315,
446
],
"score": 0.94,
"content": "\\frac { \\partial \\log \\left( P _ { \\mathrm { I I } ( \\pmb { \\theta } ) } ( Q ) \\right) } { \\partial \\mathbf { p } }",
"type": "inline_equation"
},
{
"bbox": [
313,
429,
499,
444
],
"score": 1.0,
"content": "follows from NeurASP (Yang et al., 2020) as",
"type": "text"
}
],
"index": 25
}
],
"index": 24,
"bbox_fs": [
105,
400,
506,
446
]
},
{
"type": "interline_equation",
"bbox": [
171,
450,
440,
509
],
"lines": [
{
"bbox": [
171,
450,
440,
509
],
"spans": [
{
"bbox": [
171,
450,
440,
509
],
"score": 0.95,
"content": "\\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q ) \\right) } { \\partial \\mathbf { p } } = \\frac { I \\underset { I : I \\mid = Q } { \\sum } \\frac { P _ { \\Pi ( \\theta ) } ( I ) } { P _ { \\Pi ( \\theta ) } ( c = v ) } - \\underset { I : I , v ^ { \\prime } \\mid = Q } { \\sum } \\frac { P _ { \\Pi ( \\theta ) } ( I ) } { P _ { \\Pi ( \\theta ) } ( c = v ^ { \\prime } ) } } { \\underset { I : I \\mid = Q } { \\sum } P _ { \\Pi ( \\theta ) } ( I ) } .",
"type": "interline_equation",
"image_path": "16ead2545e3a3b0901ab36b2cc5263e0781b7c70ac9da887d966ab918f15d447.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
171,
450,
440,
469.6666666666667
],
"spans": [],
"index": 26
},
{
"bbox": [
171,
469.6666666666667,
440,
489.33333333333337
],
"spans": [],
"index": 27
},
{
"bbox": [
171,
489.33333333333337,
440,
509.00000000000006
],
"spans": [],
"index": 28
}
]
},
{
"type": "text",
"bbox": [
105,
520,
505,
542
],
"lines": [
{
"bbox": [
105,
519,
507,
533
],
"spans": [
{
"bbox": [
105,
519,
507,
533
],
"score": 1.0,
"content": "This is sensible. For instance, if a query to be true is not likely to be entailed, the gradient is positive.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
530,
318,
543
],
"spans": [
{
"bbox": [
106,
530,
318,
543
],
"score": 1.0,
"content": "Putting everything together, the final loss function is",
"type": "text"
}
],
"index": 30
}
],
"index": 29.5,
"bbox_fs": [
105,
519,
507,
543
]
},
{
"type": "interline_equation",
"bbox": [
246,
549,
364,
561
],
"lines": [
{
"bbox": [
246,
549,
364,
561
],
"spans": [
{
"bbox": [
246,
549,
364,
561
],
"score": 0.9,
"content": "{ \\cal L } _ { S L A S H } = { \\cal L } _ { N P P } + { \\cal L } _ { E N T }",
"type": "interline_equation",
"image_path": "f145bb6a4a4cfb00ec9261d4033d6ebfe1f151e47554c538aff178796d7c3820.jpg"
}
]
}
],
"index": 31,
"virtual_lines": [
{
"bbox": [
246,
549,
364,
561
],
"spans": [],
"index": 31
}
]
},
{
"type": "text",
"bbox": [
105,
568,
505,
590
],
"lines": [
{
"bbox": [
106,
568,
505,
580
],
"spans": [
{
"bbox": [
106,
568,
505,
580
],
"score": 1.0,
"content": "and we perform training using coordinate descent, i.e., we train the NPPs, the train the program with",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
578,
355,
591
],
"spans": [
{
"bbox": [
105,
578,
355,
591
],
"score": 1.0,
"content": "fixed NPPs, train the NPPs with the program fixed, and so on.",
"type": "text"
}
],
"index": 33
}
],
"index": 32.5,
"bbox_fs": [
105,
568,
505,
591
]
},
{
"type": "text",
"bbox": [
106,
595,
505,
640
],
"lines": [
{
"bbox": [
105,
595,
505,
607
],
"spans": [
{
"bbox": [
105,
595,
505,
607
],
"score": 1.0,
"content": "In hindsight, rather than requiring a novel loss function for each individual task and data set, with",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
606,
505,
619
],
"spans": [
{
"bbox": [
106,
606,
505,
619
],
"score": 1.0,
"content": "SLASH, it is possible to simply incorporate the specific requirements into the logic program. The",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
617,
505,
631
],
"spans": [
{
"bbox": [
105,
617,
505,
631
],
"score": 1.0,
"content": "training loss, however, remains the same. We refer to the Appendix A for further details, including",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
629,
264,
641
],
"spans": [
{
"bbox": [
106,
629,
264,
641
],
"score": 1.0,
"content": "the derivation of the total loss gradient.",
"type": "text"
}
],
"index": 37
}
],
"index": 35.5,
"bbox_fs": [
105,
595,
505,
641
]
},
{
"type": "title",
"bbox": [
108,
656,
234,
669
],
"lines": [
{
"bbox": [
105,
655,
235,
671
],
"spans": [
{
"bbox": [
105,
655,
235,
671
],
"score": 1.0,
"content": "4 EMPIRICAL RESULTS",
"type": "text"
}
],
"index": 38
}
],
"index": 38
},
{
"type": "text",
"bbox": [
106,
681,
503,
704
],
"lines": [
{
"bbox": [
105,
680,
505,
694
],
"spans": [
{
"bbox": [
105,
680,
505,
694
],
"score": 1.0,
"content": "The advantage of SLASH lies in the efficient integration of neural, probabilistic and symbolic",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
693,
443,
705
],
"spans": [
{
"bbox": [
106,
693,
443,
705
],
"score": 1.0,
"content": "computations. To emphasize this, we conduct a variety of experimental evaluations.",
"type": "text"
}
],
"index": 40
}
],
"index": 39.5,
"bbox_fs": [
105,
680,
505,
705
]
},
{
"type": "text",
"bbox": [
106,
709,
503,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Experimental Details. We use two benchmark data sets, namely MNIST (LeCun et al., 1998b) for",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
720,
505,
733
],
"spans": [
{
"bbox": [
106,
720,
505,
733
],
"score": 1.0,
"content": "the task of MNIST-Addition and a variant of the ShapeWorld data set (Kuhnle & Copestake, 2017) for",
"type": "text"
}
],
"index": 42
}
],
"index": 41.5,
"bbox_fs": [
106,
709,
505,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
80,
505,
147
],
"lines": [
{
"bbox": [
105,
80,
506,
93
],
"spans": [
{
"bbox": [
105,
80,
506,
93
],
"score": 1.0,
"content": "Table 1: MNIST Addition Results. Test accuracy corresponds to the percentage of correctly classified",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
90,
505,
104
],
"spans": [
{
"bbox": [
105,
90,
505,
104
],
"score": 1.0,
"content": "test images. (a) Test accuracies in percent for the MNIST Addition task with various DPPLs, including",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
101,
505,
115
],
"spans": [
{
"bbox": [
105,
101,
505,
115
],
"score": 1.0,
"content": "SLASH with an NPP that models the joint probabilities (SLASH (PC)) and one that models only",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
113,
505,
126
],
"spans": [
{
"bbox": [
106,
113,
505,
126
],
"score": 1.0,
"content": "conditional probabilities (SLASH (DNN)). (b) Test accuracies in percent for the MNIST Addition",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
124,
505,
137
],
"spans": [
{
"bbox": [
105,
124,
505,
137
],
"score": 1.0,
"content": "task with missing data, comparing DeepProbLog with SLASH (PC). The amount of missing data was",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
134,
324,
149
],
"spans": [
{
"bbox": [
105,
134,
168,
149
],
"score": 1.0,
"content": "varied between",
"type": "text"
},
{
"bbox": [
169,
135,
189,
145
],
"score": 0.85,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
189,
134,
206,
149
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
207,
135,
226,
145
],
"score": 0.86,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
227,
134,
324,
149
],
"score": 1.0,
"content": "of the pixels per image.",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5
},
{
"type": "table",
"bbox": [
119,
171,
274,
232
],
"blocks": [
{
"type": "table_caption",
"bbox": [
140,
156,
250,
166
],
"group_id": 0,
"lines": [
{
"bbox": [
139,
155,
251,
167
],
"spans": [
{
"bbox": [
139,
155,
251,
167
],
"score": 1.0,
"content": "(a) Baseline MNIST Addition.",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "table_body",
"bbox": [
119,
171,
274,
232
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
171,
274,
232
],
"spans": [
{
"bbox": [
119,
171,
274,
232
],
"score": 0.972,
"html": " | Test Acc. (%) |
| DeepProbLog | 98.49±0.18 |
| NeurASP | 98.21 ± 0.30 |
| SLASH (PC) | 95.39 ± 0.29 |
| SLASH (DNN) | 98.74±0.21 |
",
"type": "table",
"image_path": "3788e55de7e7cfea5762aea98d8f13b0b9a4135918a954edc0d22b46010214d8.jpg"
}
]
}
],
"index": 11.0,
"virtual_lines": [
{
"bbox": [
119,
171,
274,
186.25
],
"spans": [],
"index": 8
},
{
"bbox": [
119,
186.25,
274,
201.5
],
"spans": [],
"index": 10
},
{
"bbox": [
119,
201.5,
274,
216.75
],
"spans": [],
"index": 12
},
{
"bbox": [
119,
216.75,
274,
232.0
],
"spans": [],
"index": 14
}
]
}
],
"index": 8.5
},
{
"type": "table",
"bbox": [
309,
171,
491,
232
],
"blocks": [
{
"type": "table_caption",
"bbox": [
334,
156,
461,
167
],
"group_id": 1,
"lines": [
{
"bbox": [
333,
155,
462,
168
],
"spans": [
{
"bbox": [
333,
155,
462,
168
],
"score": 1.0,
"content": "(b) Missing data MNIST Addition.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "table_body",
"bbox": [
309,
171,
491,
232
],
"group_id": 1,
"lines": [
{
"bbox": [
309,
171,
491,
232
],
"spans": [
{
"bbox": [
309,
171,
491,
232
],
"score": 0.975,
"html": "| DeepProbLog | SLASH (PC) |
| 50% | 97.73 ± 0.12 | 97.67±0.12 |
| 80% | 76.07 ± 18.38 | 96.72士0.05 |
| 90% | 69.15 ± 29.15 | 94.85士0.38 |
| 97% | 32.46 ± 22.48 | 82.57士4.66 |
",
"type": "table",
"image_path": "3c1f935051a7d747c53cdcf34ea4560b6fea51387ad5d49e7ef95ee58f377bb2.jpg"
}
]
}
],
"index": 12.0,
"virtual_lines": [
{
"bbox": [
309,
171,
491,
186.25
],
"spans": [],
"index": 9
},
{
"bbox": [
309,
186.25,
491,
201.5
],
"spans": [],
"index": 11
},
{
"bbox": [
309,
201.5,
491,
216.75
],
"spans": [],
"index": 13
},
{
"bbox": [
309,
216.75,
491,
232.0
],
"spans": [],
"index": 15
}
]
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
108,
257,
504,
279
],
"lines": [
{
"bbox": [
106,
258,
505,
269
],
"spans": [
{
"bbox": [
106,
258,
505,
269
],
"score": 1.0,
"content": "object-centric set prediction. For all experiments we present the average and the standard deviation",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
267,
388,
281
],
"spans": [
{
"bbox": [
105,
267,
388,
281
],
"score": 1.0,
"content": "over five runs with different random seeds for parameter initialization.",
"type": "text"
}
],
"index": 17
}
],
"index": 16.5
},
{
"type": "text",
"bbox": [
107,
285,
505,
341
],
"lines": [
{
"bbox": [
105,
285,
506,
298
],
"spans": [
{
"bbox": [
105,
285,
506,
298
],
"score": 1.0,
"content": "For ShapeWorld experiments, we generate a data set we refer to as ShapeWorld4. Images of",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
295,
506,
309
],
"spans": [
{
"bbox": [
105,
295,
506,
309
],
"score": 1.0,
"content": "ShapeWorld4 contain between one and four objects, with each object consisting of four attributes:",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
104,
306,
505,
321
],
"spans": [
{
"bbox": [
104,
306,
505,
321
],
"score": 1.0,
"content": "a color (red, blue, green, gray, brown, magenta, cyan or yellow), a shade (bright, or dark), a shape",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
318,
505,
330
],
"spans": [
{
"bbox": [
105,
318,
505,
330
],
"score": 1.0,
"content": "(circle, triangle or square) and a size (small or big). Thus, each object can be created from 84 different",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
329,
349,
342
],
"spans": [
{
"bbox": [
106,
329,
349,
342
],
"score": 1.0,
"content": "combinations of attributes. Fig. 1 depicts an example image.",
"type": "text"
}
],
"index": 22
}
],
"index": 20
},
{
"type": "text",
"bbox": [
107,
345,
505,
390
],
"lines": [
{
"bbox": [
105,
345,
507,
358
],
"spans": [
{
"bbox": [
105,
345,
507,
358
],
"score": 1.0,
"content": "We measure performance via classification accuracies in the MNIST-Addition task. In our Shape-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
356,
506,
369
],
"spans": [
{
"bbox": [
105,
356,
506,
369
],
"score": 1.0,
"content": "World4 experiments, we present the average precision. We refer to appendix B for the SLASH",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
368,
507,
380
],
"spans": [
{
"bbox": [
105,
368,
507,
380
],
"score": 1.0,
"content": "programs and queries of each experiment, and appendix C for a detailed description of hyperparame-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
379,
200,
390
],
"spans": [
{
"bbox": [
106,
379,
200,
390
],
"score": 1.0,
"content": "ters and further details.",
"type": "text"
}
],
"index": 26
}
],
"index": 24.5
},
{
"type": "text",
"bbox": [
106,
395,
505,
451
],
"lines": [
{
"bbox": [
105,
395,
507,
408
],
"spans": [
{
"bbox": [
105,
395,
507,
408
],
"score": 1.0,
"content": "Evaluation 1: SLASH outperforms SOTA DPPLs in MNIST-Addition. The task of MNIST-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
406,
506,
420
],
"spans": [
{
"bbox": [
105,
406,
506,
420
],
"score": 1.0,
"content": "Addition (Manhaeve et al., 2018) is to predict the sum of two MNIST digits, presented only as raw",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
417,
506,
430
],
"spans": [
{
"bbox": [
105,
417,
506,
430
],
"score": 1.0,
"content": "images. During test time, however, a model should classify the images directly. Thus, although a",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
429,
505,
442
],
"spans": [
{
"bbox": [
105,
429,
505,
442
],
"score": 1.0,
"content": "model does not receive explicit information about the depicted digits, it must learn to identify digits",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
439,
283,
453
],
"spans": [
{
"bbox": [
105,
439,
283,
453
],
"score": 1.0,
"content": "via indirect feedback on the sum prediction.",
"type": "text"
}
],
"index": 31
}
],
"index": 29
},
{
"type": "text",
"bbox": [
106,
456,
505,
523
],
"lines": [
{
"bbox": [
105,
455,
505,
470
],
"spans": [
{
"bbox": [
105,
455,
505,
470
],
"score": 1.0,
"content": "We compare the test accuracy after convergence between the three DPPLs: DeepProbLog (Manhaeve",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
104,
466,
506,
481
],
"spans": [
{
"bbox": [
104,
466,
506,
481
],
"score": 1.0,
"content": "et al., 2018), NeurASP (Yang et al., 2020) and SLASH, using a probabilistic circuit (PC) or a deep",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
478,
505,
491
],
"spans": [
{
"bbox": [
105,
478,
505,
491
],
"score": 1.0,
"content": "neural network (DNN) as NPP. Notably, the DNN used in SLASH (DNN) is the LeNet5 model",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
489,
505,
502
],
"spans": [
{
"bbox": [
105,
489,
505,
502
],
"score": 1.0,
"content": "(LeCun et al., 1998a) of DeepProbLog and NeurASP. We note that when using the PC as NPP, we",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
500,
506,
514
],
"spans": [
{
"bbox": [
105,
500,
302,
514
],
"score": 1.0,
"content": "have also extracted conditional class probabilities",
"type": "text"
},
{
"bbox": [
303,
500,
339,
512
],
"score": 0.92,
"content": "P ( C | X )",
"type": "inline_equation"
},
{
"bbox": [
339,
500,
484,
514
],
"score": 1.0,
"content": ", by marginalizing the class variables",
"type": "text"
},
{
"bbox": [
485,
501,
494,
510
],
"score": 0.84,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
494,
500,
506,
514
],
"score": 1.0,
"content": "to",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
511,
475,
524
],
"spans": [
{
"bbox": [
106,
511,
246,
524
],
"score": 1.0,
"content": "acquire the normalization constant",
"type": "text"
},
{
"bbox": [
247,
511,
272,
523
],
"score": 0.92,
"content": "P ( X )",
"type": "inline_equation"
},
{
"bbox": [
272,
511,
331,
524
],
"score": 1.0,
"content": "from the joint",
"type": "text"
},
{
"bbox": [
331,
512,
368,
523
],
"score": 0.93,
"content": "P ( X , C )",
"type": "inline_equation"
},
{
"bbox": [
368,
511,
435,
524
],
"score": 1.0,
"content": ", and calculating",
"type": "text"
},
{
"bbox": [
435,
511,
471,
523
],
"score": 0.93,
"content": "P ( X | C )",
"type": "inline_equation"
},
{
"bbox": [
471,
511,
475,
524
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 37
}
],
"index": 34.5
},
{
"type": "text",
"bbox": [
106,
528,
505,
628
],
"lines": [
{
"bbox": [
106,
528,
505,
540
],
"spans": [
{
"bbox": [
106,
528,
505,
540
],
"score": 1.0,
"content": "The results can be seen in Tab. 1a. We observe that training SLASH with a DNN NPP produces",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
540,
506,
551
],
"spans": [
{
"bbox": [
106,
540,
506,
551
],
"score": 1.0,
"content": "SOTA accuracies compared to DeepProbLog and NeurASP, confirming that SLASH’s batch-wise",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
551,
506,
563
],
"spans": [
{
"bbox": [
106,
551,
506,
563
],
"score": 1.0,
"content": "loss computation leads to improved performances. We further observe that the test accuracy of",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
560,
506,
574
],
"spans": [
{
"bbox": [
105,
560,
506,
574
],
"score": 1.0,
"content": "SLASH with a PC NPP is slightly below the other DPPLs, however we argue that this may be",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
572,
506,
585
],
"spans": [
{
"bbox": [
105,
572,
506,
585
],
"score": 1.0,
"content": "since a PC, in comparison to a DNN, is learning a true mixture density rather than just conditional",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
583,
506,
596
],
"spans": [
{
"bbox": [
106,
583,
506,
596
],
"score": 1.0,
"content": "probabilities. The advantages of doing so will be investigated in the next experiments. Note that,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
594,
505,
607
],
"spans": [
{
"bbox": [
105,
594,
505,
607
],
"score": 1.0,
"content": "optimal architecture search for PCs, e.g. for computer vision, is an open research question.These",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
605,
505,
617
],
"spans": [
{
"bbox": [
106,
605,
505,
617
],
"score": 1.0,
"content": "evaluations show SLASH’s advantages on the benchmark MNIST-Addition task. Additional benefits",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
616,
301,
629
],
"spans": [
{
"bbox": [
105,
616,
301,
629
],
"score": 1.0,
"content": "will be made clear in the following experiments.",
"type": "text"
}
],
"index": 46
}
],
"index": 42
},
{
"type": "text",
"bbox": [
106,
632,
505,
732
],
"lines": [
{
"bbox": [
105,
632,
505,
646
],
"spans": [
{
"bbox": [
105,
632,
505,
646
],
"score": 1.0,
"content": "Evaluation 2: Handling Missing Data with SLASH. SLASH offers the advantage of its flexibility",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
643,
505,
657
],
"spans": [
{
"bbox": [
105,
643,
505,
657
],
"score": 1.0,
"content": "to use various kinds of NPPs. Thus, in comparison to previous DPPLs, one can easily integrate",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
655,
505,
668
],
"spans": [
{
"bbox": [
105,
655,
505,
668
],
"score": 1.0,
"content": "NPPs into SLASH that perform joint probability estimation. For this evaluation, we consider the",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
106,
666,
505,
678
],
"spans": [
{
"bbox": [
106,
666,
505,
678
],
"score": 1.0,
"content": "task of MNIST-Addition with missing data. We trained SLASH (PC) and DeepProbLog with the",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
676,
506,
690
],
"spans": [
{
"bbox": [
105,
676,
506,
690
],
"score": 1.0,
"content": "MNIST-Addition task with images in which a percentage of pixels per image has been removed. It is",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
688,
505,
701
],
"spans": [
{
"bbox": [
105,
688,
505,
701
],
"score": 1.0,
"content": "important to mention here that whereas DeepProbLog handles the missing data simply as background",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
106,
699,
505,
712
],
"spans": [
{
"bbox": [
106,
699,
505,
712
],
"score": 1.0,
"content": "pixels, SLASH (PC) specifically models the missing data as uncertain data by marginalizing the",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
105,
709,
505,
723
],
"spans": [
{
"bbox": [
105,
709,
505,
723
],
"score": 1.0,
"content": "denoted pixels at inference time. We use DeepProbLog here representative of DPPLs without true",
"type": "text"
}
],
"index": 54
},
{
"bbox": [
106,
721,
183,
733
],
"spans": [
{
"bbox": [
106,
721,
183,
733
],
"score": 1.0,
"content": "density estimation.",
"type": "text"
}
],
"index": 55
}
],
"index": 51
}
],
"page_idx": 6,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
759
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "7",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
80,
505,
147
],
"lines": [
{
"bbox": [
105,
80,
506,
93
],
"spans": [
{
"bbox": [
105,
80,
506,
93
],
"score": 1.0,
"content": "Table 1: MNIST Addition Results. Test accuracy corresponds to the percentage of correctly classified",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
90,
505,
104
],
"spans": [
{
"bbox": [
105,
90,
505,
104
],
"score": 1.0,
"content": "test images. (a) Test accuracies in percent for the MNIST Addition task with various DPPLs, including",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
101,
505,
115
],
"spans": [
{
"bbox": [
105,
101,
505,
115
],
"score": 1.0,
"content": "SLASH with an NPP that models the joint probabilities (SLASH (PC)) and one that models only",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
113,
505,
126
],
"spans": [
{
"bbox": [
106,
113,
505,
126
],
"score": 1.0,
"content": "conditional probabilities (SLASH (DNN)). (b) Test accuracies in percent for the MNIST Addition",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
124,
505,
137
],
"spans": [
{
"bbox": [
105,
124,
505,
137
],
"score": 1.0,
"content": "task with missing data, comparing DeepProbLog with SLASH (PC). The amount of missing data was",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
134,
324,
149
],
"spans": [
{
"bbox": [
105,
134,
168,
149
],
"score": 1.0,
"content": "varied between",
"type": "text"
},
{
"bbox": [
169,
135,
189,
145
],
"score": 0.85,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
189,
134,
206,
149
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
207,
135,
226,
145
],
"score": 0.86,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
227,
134,
324,
149
],
"score": 1.0,
"content": "of the pixels per image.",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5,
"bbox_fs": [
105,
80,
506,
149
]
},
{
"type": "table",
"bbox": [
119,
171,
274,
232
],
"blocks": [
{
"type": "table_caption",
"bbox": [
140,
156,
250,
166
],
"group_id": 0,
"lines": [
{
"bbox": [
139,
155,
251,
167
],
"spans": [
{
"bbox": [
139,
155,
251,
167
],
"score": 1.0,
"content": "(a) Baseline MNIST Addition.",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "table_body",
"bbox": [
119,
171,
274,
232
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
171,
274,
232
],
"spans": [
{
"bbox": [
119,
171,
274,
232
],
"score": 0.972,
"html": " | Test Acc. (%) |
| DeepProbLog | 98.49±0.18 |
| NeurASP | 98.21 ± 0.30 |
| SLASH (PC) | 95.39 ± 0.29 |
| SLASH (DNN) | 98.74±0.21 |
",
"type": "table",
"image_path": "3788e55de7e7cfea5762aea98d8f13b0b9a4135918a954edc0d22b46010214d8.jpg"
}
]
}
],
"index": 11.0,
"virtual_lines": [
{
"bbox": [
119,
171,
274,
186.25
],
"spans": [],
"index": 8
},
{
"bbox": [
119,
186.25,
274,
201.5
],
"spans": [],
"index": 10
},
{
"bbox": [
119,
201.5,
274,
216.75
],
"spans": [],
"index": 12
},
{
"bbox": [
119,
216.75,
274,
232.0
],
"spans": [],
"index": 14
}
]
}
],
"index": 8.5
},
{
"type": "table",
"bbox": [
309,
171,
491,
232
],
"blocks": [
{
"type": "table_caption",
"bbox": [
334,
156,
461,
167
],
"group_id": 1,
"lines": [
{
"bbox": [
333,
155,
462,
168
],
"spans": [
{
"bbox": [
333,
155,
462,
168
],
"score": 1.0,
"content": "(b) Missing data MNIST Addition.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "table_body",
"bbox": [
309,
171,
491,
232
],
"group_id": 1,
"lines": [
{
"bbox": [
309,
171,
491,
232
],
"spans": [
{
"bbox": [
309,
171,
491,
232
],
"score": 0.975,
"html": "| DeepProbLog | SLASH (PC) |
| 50% | 97.73 ± 0.12 | 97.67±0.12 |
| 80% | 76.07 ± 18.38 | 96.72士0.05 |
| 90% | 69.15 ± 29.15 | 94.85士0.38 |
| 97% | 32.46 ± 22.48 | 82.57士4.66 |
",
"type": "table",
"image_path": "3c1f935051a7d747c53cdcf34ea4560b6fea51387ad5d49e7ef95ee58f377bb2.jpg"
}
]
}
],
"index": 12.0,
"virtual_lines": [
{
"bbox": [
309,
171,
491,
186.25
],
"spans": [],
"index": 9
},
{
"bbox": [
309,
186.25,
491,
201.5
],
"spans": [],
"index": 11
},
{
"bbox": [
309,
201.5,
491,
216.75
],
"spans": [],
"index": 13
},
{
"bbox": [
309,
216.75,
491,
232.0
],
"spans": [],
"index": 15
}
]
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
108,
257,
504,
279
],
"lines": [
{
"bbox": [
106,
258,
505,
269
],
"spans": [
{
"bbox": [
106,
258,
505,
269
],
"score": 1.0,
"content": "object-centric set prediction. For all experiments we present the average and the standard deviation",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
267,
388,
281
],
"spans": [
{
"bbox": [
105,
267,
388,
281
],
"score": 1.0,
"content": "over five runs with different random seeds for parameter initialization.",
"type": "text"
}
],
"index": 17
}
],
"index": 16.5,
"bbox_fs": [
105,
258,
505,
281
]
},
{
"type": "text",
"bbox": [
107,
285,
505,
341
],
"lines": [
{
"bbox": [
105,
285,
506,
298
],
"spans": [
{
"bbox": [
105,
285,
506,
298
],
"score": 1.0,
"content": "For ShapeWorld experiments, we generate a data set we refer to as ShapeWorld4. Images of",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
295,
506,
309
],
"spans": [
{
"bbox": [
105,
295,
506,
309
],
"score": 1.0,
"content": "ShapeWorld4 contain between one and four objects, with each object consisting of four attributes:",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
104,
306,
505,
321
],
"spans": [
{
"bbox": [
104,
306,
505,
321
],
"score": 1.0,
"content": "a color (red, blue, green, gray, brown, magenta, cyan or yellow), a shade (bright, or dark), a shape",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
318,
505,
330
],
"spans": [
{
"bbox": [
105,
318,
505,
330
],
"score": 1.0,
"content": "(circle, triangle or square) and a size (small or big). Thus, each object can be created from 84 different",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
329,
349,
342
],
"spans": [
{
"bbox": [
106,
329,
349,
342
],
"score": 1.0,
"content": "combinations of attributes. Fig. 1 depicts an example image.",
"type": "text"
}
],
"index": 22
}
],
"index": 20,
"bbox_fs": [
104,
285,
506,
342
]
},
{
"type": "text",
"bbox": [
107,
345,
505,
390
],
"lines": [
{
"bbox": [
105,
345,
507,
358
],
"spans": [
{
"bbox": [
105,
345,
507,
358
],
"score": 1.0,
"content": "We measure performance via classification accuracies in the MNIST-Addition task. In our Shape-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
356,
506,
369
],
"spans": [
{
"bbox": [
105,
356,
506,
369
],
"score": 1.0,
"content": "World4 experiments, we present the average precision. We refer to appendix B for the SLASH",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
368,
507,
380
],
"spans": [
{
"bbox": [
105,
368,
507,
380
],
"score": 1.0,
"content": "programs and queries of each experiment, and appendix C for a detailed description of hyperparame-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
379,
200,
390
],
"spans": [
{
"bbox": [
106,
379,
200,
390
],
"score": 1.0,
"content": "ters and further details.",
"type": "text"
}
],
"index": 26
}
],
"index": 24.5,
"bbox_fs": [
105,
345,
507,
390
]
},
{
"type": "text",
"bbox": [
106,
395,
505,
451
],
"lines": [
{
"bbox": [
105,
395,
507,
408
],
"spans": [
{
"bbox": [
105,
395,
507,
408
],
"score": 1.0,
"content": "Evaluation 1: SLASH outperforms SOTA DPPLs in MNIST-Addition. The task of MNIST-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
406,
506,
420
],
"spans": [
{
"bbox": [
105,
406,
506,
420
],
"score": 1.0,
"content": "Addition (Manhaeve et al., 2018) is to predict the sum of two MNIST digits, presented only as raw",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
417,
506,
430
],
"spans": [
{
"bbox": [
105,
417,
506,
430
],
"score": 1.0,
"content": "images. During test time, however, a model should classify the images directly. Thus, although a",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
429,
505,
442
],
"spans": [
{
"bbox": [
105,
429,
505,
442
],
"score": 1.0,
"content": "model does not receive explicit information about the depicted digits, it must learn to identify digits",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
439,
283,
453
],
"spans": [
{
"bbox": [
105,
439,
283,
453
],
"score": 1.0,
"content": "via indirect feedback on the sum prediction.",
"type": "text"
}
],
"index": 31
}
],
"index": 29,
"bbox_fs": [
105,
395,
507,
453
]
},
{
"type": "text",
"bbox": [
106,
456,
505,
523
],
"lines": [
{
"bbox": [
105,
455,
505,
470
],
"spans": [
{
"bbox": [
105,
455,
505,
470
],
"score": 1.0,
"content": "We compare the test accuracy after convergence between the three DPPLs: DeepProbLog (Manhaeve",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
104,
466,
506,
481
],
"spans": [
{
"bbox": [
104,
466,
506,
481
],
"score": 1.0,
"content": "et al., 2018), NeurASP (Yang et al., 2020) and SLASH, using a probabilistic circuit (PC) or a deep",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
478,
505,
491
],
"spans": [
{
"bbox": [
105,
478,
505,
491
],
"score": 1.0,
"content": "neural network (DNN) as NPP. Notably, the DNN used in SLASH (DNN) is the LeNet5 model",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
489,
505,
502
],
"spans": [
{
"bbox": [
105,
489,
505,
502
],
"score": 1.0,
"content": "(LeCun et al., 1998a) of DeepProbLog and NeurASP. We note that when using the PC as NPP, we",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
500,
506,
514
],
"spans": [
{
"bbox": [
105,
500,
302,
514
],
"score": 1.0,
"content": "have also extracted conditional class probabilities",
"type": "text"
},
{
"bbox": [
303,
500,
339,
512
],
"score": 0.92,
"content": "P ( C | X )",
"type": "inline_equation"
},
{
"bbox": [
339,
500,
484,
514
],
"score": 1.0,
"content": ", by marginalizing the class variables",
"type": "text"
},
{
"bbox": [
485,
501,
494,
510
],
"score": 0.84,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
494,
500,
506,
514
],
"score": 1.0,
"content": "to",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
511,
475,
524
],
"spans": [
{
"bbox": [
106,
511,
246,
524
],
"score": 1.0,
"content": "acquire the normalization constant",
"type": "text"
},
{
"bbox": [
247,
511,
272,
523
],
"score": 0.92,
"content": "P ( X )",
"type": "inline_equation"
},
{
"bbox": [
272,
511,
331,
524
],
"score": 1.0,
"content": "from the joint",
"type": "text"
},
{
"bbox": [
331,
512,
368,
523
],
"score": 0.93,
"content": "P ( X , C )",
"type": "inline_equation"
},
{
"bbox": [
368,
511,
435,
524
],
"score": 1.0,
"content": ", and calculating",
"type": "text"
},
{
"bbox": [
435,
511,
471,
523
],
"score": 0.93,
"content": "P ( X | C )",
"type": "inline_equation"
},
{
"bbox": [
471,
511,
475,
524
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 37
}
],
"index": 34.5,
"bbox_fs": [
104,
455,
506,
524
]
},
{
"type": "text",
"bbox": [
106,
528,
505,
628
],
"lines": [
{
"bbox": [
106,
528,
505,
540
],
"spans": [
{
"bbox": [
106,
528,
505,
540
],
"score": 1.0,
"content": "The results can be seen in Tab. 1a. We observe that training SLASH with a DNN NPP produces",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
540,
506,
551
],
"spans": [
{
"bbox": [
106,
540,
506,
551
],
"score": 1.0,
"content": "SOTA accuracies compared to DeepProbLog and NeurASP, confirming that SLASH’s batch-wise",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
551,
506,
563
],
"spans": [
{
"bbox": [
106,
551,
506,
563
],
"score": 1.0,
"content": "loss computation leads to improved performances. We further observe that the test accuracy of",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
560,
506,
574
],
"spans": [
{
"bbox": [
105,
560,
506,
574
],
"score": 1.0,
"content": "SLASH with a PC NPP is slightly below the other DPPLs, however we argue that this may be",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
572,
506,
585
],
"spans": [
{
"bbox": [
105,
572,
506,
585
],
"score": 1.0,
"content": "since a PC, in comparison to a DNN, is learning a true mixture density rather than just conditional",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
583,
506,
596
],
"spans": [
{
"bbox": [
106,
583,
506,
596
],
"score": 1.0,
"content": "probabilities. The advantages of doing so will be investigated in the next experiments. Note that,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
594,
505,
607
],
"spans": [
{
"bbox": [
105,
594,
505,
607
],
"score": 1.0,
"content": "optimal architecture search for PCs, e.g. for computer vision, is an open research question.These",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
605,
505,
617
],
"spans": [
{
"bbox": [
106,
605,
505,
617
],
"score": 1.0,
"content": "evaluations show SLASH’s advantages on the benchmark MNIST-Addition task. Additional benefits",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
616,
301,
629
],
"spans": [
{
"bbox": [
105,
616,
301,
629
],
"score": 1.0,
"content": "will be made clear in the following experiments.",
"type": "text"
}
],
"index": 46
}
],
"index": 42,
"bbox_fs": [
105,
528,
506,
629
]
},
{
"type": "text",
"bbox": [
106,
632,
505,
732
],
"lines": [
{
"bbox": [
105,
632,
505,
646
],
"spans": [
{
"bbox": [
105,
632,
505,
646
],
"score": 1.0,
"content": "Evaluation 2: Handling Missing Data with SLASH. SLASH offers the advantage of its flexibility",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
643,
505,
657
],
"spans": [
{
"bbox": [
105,
643,
505,
657
],
"score": 1.0,
"content": "to use various kinds of NPPs. Thus, in comparison to previous DPPLs, one can easily integrate",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
655,
505,
668
],
"spans": [
{
"bbox": [
105,
655,
505,
668
],
"score": 1.0,
"content": "NPPs into SLASH that perform joint probability estimation. For this evaluation, we consider the",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
106,
666,
505,
678
],
"spans": [
{
"bbox": [
106,
666,
505,
678
],
"score": 1.0,
"content": "task of MNIST-Addition with missing data. We trained SLASH (PC) and DeepProbLog with the",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
676,
506,
690
],
"spans": [
{
"bbox": [
105,
676,
506,
690
],
"score": 1.0,
"content": "MNIST-Addition task with images in which a percentage of pixels per image has been removed. It is",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
688,
505,
701
],
"spans": [
{
"bbox": [
105,
688,
505,
701
],
"score": 1.0,
"content": "important to mention here that whereas DeepProbLog handles the missing data simply as background",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
106,
699,
505,
712
],
"spans": [
{
"bbox": [
106,
699,
505,
712
],
"score": 1.0,
"content": "pixels, SLASH (PC) specifically models the missing data as uncertain data by marginalizing the",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
105,
709,
505,
723
],
"spans": [
{
"bbox": [
105,
709,
505,
723
],
"score": 1.0,
"content": "denoted pixels at inference time. We use DeepProbLog here representative of DPPLs without true",
"type": "text"
}
],
"index": 54
},
{
"bbox": [
106,
721,
183,
733
],
"spans": [
{
"bbox": [
106,
721,
183,
733
],
"score": 1.0,
"content": "density estimation.",
"type": "text"
}
],
"index": 55
}
],
"index": 51,
"bbox_fs": [
105,
632,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
119,
106,
317,
183
],
"blocks": [
{
"type": "table_caption",
"bbox": [
119,
83,
288,
104
],
"group_id": 0,
"lines": [
{
"bbox": [
118,
83,
288,
93
],
"spans": [
{
"bbox": [
118,
83,
288,
93
],
"score": 0.982,
"content": "(a) ShapeWorld4 and ShapeWorld4 CoGenT Test",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
118,
93,
171,
105
],
"spans": [
{
"bbox": [
118,
93,
171,
105
],
"score": 1.0,
"content": "Avg.Precision",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "table_body",
"bbox": [
119,
106,
317,
183
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
106,
317,
183
],
"spans": [
{
"bbox": [
119,
106,
317,
183
],
"score": 0.976,
"html": " | Slot Att. | SLASH Att. |
| Test Set | ShapeWorld4 90.24 ± 0.93 | 95.58 ± 0.61 |
| CoGenT | |
| Test Cond. A | 90.37 ± 2.19 | 96.85 ± 0.43 |
| Test Cond.B | 27.15 ± 2.36 | 40.58 ± 1.99 |
",
"type": "table",
"image_path": "0b06c8954d45055a4eff3ce528d6a37b02d5a2f22cb9bd72bf07133146fdad99.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
119,
106,
317,
121.4
],
"spans": [],
"index": 2
},
{
"bbox": [
119,
121.4,
317,
136.8
],
"spans": [],
"index": 3
},
{
"bbox": [
119,
136.8,
317,
152.20000000000002
],
"spans": [],
"index": 4
},
{
"bbox": [
119,
152.20000000000002,
317,
167.60000000000002
],
"spans": [],
"index": 5
},
{
"bbox": [
119,
167.60000000000002,
317,
183.00000000000003
],
"spans": [],
"index": 6
}
]
},
{
"type": "table_caption",
"bbox": [
106,
193,
505,
282
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
194,
506,
207
],
"spans": [
{
"bbox": [
105,
194,
506,
207
],
"score": 1.0,
"content": "Figure 3: ShapeWorld4 Experiments. (a) Converged test average precision scores for the set prediction",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
204,
505,
218
],
"spans": [
{
"bbox": [
105,
204,
505,
218
],
"score": 1.0,
"content": "task with ShapeWorld4 (top) and ShapeWorld4 CoGenT (bottom). (b) Test average precision scores",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
215,
505,
228
],
"spans": [
{
"bbox": [
105,
215,
505,
228
],
"score": 1.0,
"content": "for set prediction with ShapeWorld4 over the training epochs. In these experiments we compared",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
227,
505,
238
],
"spans": [
{
"bbox": [
106,
227,
505,
238
],
"score": 1.0,
"content": "a baseline slot encoder versus SLASH Attention with slot attention and PC-based NPPs. For the",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
238,
506,
250
],
"spans": [
{
"bbox": [
105,
238,
506,
250
],
"score": 1.0,
"content": "CoGenT experiments, a model is trained on one training set and tested on two separate test conditions.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
248,
505,
261
],
"spans": [
{
"bbox": [
105,
248,
505,
261
],
"score": 1.0,
"content": "The Condition A test set contains attribute compositions which were also seen during training. The",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
259,
506,
273
],
"spans": [
{
"bbox": [
105,
259,
506,
273
],
"score": 1.0,
"content": "Condition B test set contains attribute compositions which were not seen during training, e.g. yellow",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
271,
418,
283
],
"spans": [
{
"bbox": [
106,
271,
418,
283
],
"score": 1.0,
"content": "circles were not present in the training set, but present in Condition B test set.",
"type": "text"
}
],
"index": 21
}
],
"index": 17.5
}
],
"index": 4
},
{
"type": "image",
"bbox": [
323,
84,
493,
171
],
"blocks": [
{
"type": "image_body",
"bbox": [
323,
84,
493,
171
],
"group_id": 0,
"lines": [
{
"bbox": [
323,
84,
493,
171
],
"spans": [
{
"bbox": [
323,
84,
493,
171
],
"score": 0.959,
"type": "image",
"image_path": "0eaf19403876aa7fbb58219357ca9c041a53e7685ffb819aff2c0d5f28605742.jpg"
}
]
}
],
"index": 9.5,
"virtual_lines": [
{
"bbox": [
323,
84,
493,
98.5
],
"spans": [],
"index": 7
},
{
"bbox": [
323,
98.5,
493,
113.0
],
"spans": [],
"index": 8
},
{
"bbox": [
323,
113.0,
493,
127.5
],
"spans": [],
"index": 9
},
{
"bbox": [
323,
127.5,
493,
142.0
],
"spans": [],
"index": 10
},
{
"bbox": [
323,
142.0,
493,
156.5
],
"spans": [],
"index": 11
},
{
"bbox": [
323,
156.5,
493,
171.0
],
"spans": [],
"index": 12
}
]
},
{
"type": "image_caption",
"bbox": [
333,
173,
485,
183
],
"group_id": 0,
"lines": [
{
"bbox": [
333,
173,
486,
184
],
"spans": [
{
"bbox": [
333,
173,
486,
184
],
"score": 0.98,
"content": "(b) Test Avg.Precision over Training Epochs",
"type": "text"
}
],
"index": 13
}
],
"index": 13
}
],
"index": 11.25
},
{
"type": "text",
"bbox": [
107,
295,
505,
384
],
"lines": [
{
"bbox": [
105,
294,
506,
309
],
"spans": [
{
"bbox": [
105,
294,
252,
309
],
"score": 1.0,
"content": "The results can be seen in Tab. 1b for",
"type": "text"
},
{
"bbox": [
252,
296,
272,
307
],
"score": 0.88,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
272,
294,
275,
309
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
275,
295,
294,
307
],
"score": 0.86,
"content": "8 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
295,
294,
298,
309
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
298,
296,
318,
307
],
"score": 0.86,
"content": "9 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
318,
294,
334,
309
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
335,
296,
354,
307
],
"score": 0.88,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
355,
294,
506,
309
],
"score": 1.0,
"content": "missing pixels per image. We observe",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
306,
506,
321
],
"spans": [
{
"bbox": [
105,
306,
133,
321
],
"score": 1.0,
"content": "that at",
"type": "text"
},
{
"bbox": [
133,
307,
153,
318
],
"score": 0.87,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
153,
306,
419,
321
],
"score": 1.0,
"content": ", DeepProbLog and SLASH produce almost equal accuracies. With",
"type": "text"
},
{
"bbox": [
420,
307,
439,
318
],
"score": 0.88,
"content": "8 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
440,
306,
506,
321
],
"score": 1.0,
"content": "percent missing",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
317,
506,
331
],
"spans": [
{
"bbox": [
105,
317,
506,
331
],
"score": 1.0,
"content": "pixels, there is a substantial difference in the ability of the two DPPLs to correctly classify images,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
329,
505,
341
],
"spans": [
{
"bbox": [
105,
329,
505,
341
],
"score": 1.0,
"content": "with SLASH being very stable. By further increasing the percentage of missing pixels, this difference",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
339,
506,
352
],
"spans": [
{
"bbox": [
105,
339,
352,
352
],
"score": 1.0,
"content": "becomes even more substantial with SLASH still reaching a",
"type": "text"
},
{
"bbox": [
352,
340,
372,
351
],
"score": 0.89,
"content": "8 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
372,
339,
473,
352
],
"score": 1.0,
"content": "test accuracy even when",
"type": "text"
},
{
"bbox": [
473,
340,
493,
351
],
"score": 0.88,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
493,
339,
506,
352
],
"score": 1.0,
"content": "of",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
349,
507,
365
],
"spans": [
{
"bbox": [
105,
349,
431,
365
],
"score": 1.0,
"content": "the pixels per image are missing, whereas DeepProbLog degrades to an average of",
"type": "text"
},
{
"bbox": [
432,
351,
451,
362
],
"score": 0.89,
"content": "3 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
451,
349,
507,
365
],
"score": 1.0,
"content": "test accuracy.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
361,
506,
375
],
"spans": [
{
"bbox": [
105,
361,
506,
375
],
"score": 1.0,
"content": "We further note that SLASH, in comparison to DeepProbLog, produces largely reduced standard",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
372,
192,
385
],
"spans": [
{
"bbox": [
106,
372,
192,
385
],
"score": 1.0,
"content": "deviations over runs.",
"type": "text"
}
],
"index": 29
}
],
"index": 25.5
},
{
"type": "text",
"bbox": [
107,
389,
505,
434
],
"lines": [
{
"bbox": [
106,
390,
505,
402
],
"spans": [
{
"bbox": [
106,
390,
505,
402
],
"score": 1.0,
"content": "Thus, by utilizing the power of true density estimation SLASH, with an appropriate NPP, can produce",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
401,
505,
413
],
"spans": [
{
"bbox": [
105,
401,
505,
413
],
"score": 1.0,
"content": "more robust results in comparison to other DPPLs. Further, we refer to Appendix D, which contains",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
412,
506,
425
],
"spans": [
{
"bbox": [
105,
412,
506,
425
],
"score": 1.0,
"content": "results of additional experiments where training is performed with the full MNIST data set whereas",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
423,
330,
435
],
"spans": [
{
"bbox": [
105,
423,
330,
435
],
"score": 1.0,
"content": "only the test set entails different rates of missing pixels.",
"type": "text"
}
],
"index": 33
}
],
"index": 31.5
},
{
"type": "text",
"bbox": [
107,
439,
505,
506
],
"lines": [
{
"bbox": [
106,
440,
506,
451
],
"spans": [
{
"bbox": [
106,
440,
506,
451
],
"score": 1.0,
"content": "Evaluation 3: Improved Concept Learning via SLASH. We show that SLASH can be very ef-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
451,
505,
462
],
"spans": [
{
"bbox": [
105,
451,
505,
462
],
"score": 1.0,
"content": "fective for the complex task of set prediction, which previous DPPLs have not tackled. We revert",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
461,
505,
474
],
"spans": [
{
"bbox": [
105,
461,
505,
474
],
"score": 1.0,
"content": "to the ShapeWorld4 data set for this setting. For set prediction, a model is trained to predict the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
471,
505,
485
],
"spans": [
{
"bbox": [
105,
471,
505,
485
],
"score": 1.0,
"content": "discrete attributes of a set of objects in an image (cf. Fig. 1 for an example ShapeWorld4 image). The",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
482,
505,
496
],
"spans": [
{
"bbox": [
105,
482,
505,
496
],
"score": 1.0,
"content": "difficulty for the model lies therein that it must match an unordered set of corresponding attributes",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
493,
478,
508
],
"spans": [
{
"bbox": [
105,
493,
478,
508
],
"score": 1.0,
"content": "(with varying number of entities over samples) with its internal representations of the image.",
"type": "text"
}
],
"index": 39
}
],
"index": 36.5
},
{
"type": "text",
"bbox": [
107,
511,
505,
566
],
"lines": [
{
"bbox": [
106,
511,
505,
523
],
"spans": [
{
"bbox": [
106,
511,
505,
523
],
"score": 1.0,
"content": "The slot attention module introduced by Locatello et al. (2020) allows for an attractive object-centric",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
522,
505,
535
],
"spans": [
{
"bbox": [
105,
522,
505,
535
],
"score": 1.0,
"content": "approach to this task. Specifically, this module represents a pluggable, differentiable module that can",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
533,
507,
546
],
"spans": [
{
"bbox": [
106,
533,
507,
546
],
"score": 1.0,
"content": "be easily added to any architecture and, through a competitive softmax-based attention mechanism,",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
545,
506,
557
],
"spans": [
{
"bbox": [
106,
545,
506,
557
],
"score": 1.0,
"content": "can enforce the binding of specific parts of a latent representation into permutation-invariant, task-",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
556,
222,
567
],
"spans": [
{
"bbox": [
106,
556,
222,
567
],
"score": 1.0,
"content": "specific vectors, called slots.",
"type": "text"
}
],
"index": 44
}
],
"index": 42
},
{
"type": "text",
"bbox": [
107,
572,
505,
638
],
"lines": [
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "In our experiments, we wish to show that by adding logical constraints to the training setting, one",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
582,
506,
596
],
"spans": [
{
"bbox": [
105,
582,
506,
596
],
"score": 1.0,
"content": "can improve the overall performances and generalization properties of such a model. For this, we",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
594,
506,
607
],
"spans": [
{
"bbox": [
105,
594,
506,
607
],
"score": 1.0,
"content": "train SLASH with NPPs as depicted in Fig. 1 consisting of a shared slot encoder and separate PCs,",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
604,
506,
618
],
"spans": [
{
"bbox": [
105,
604,
506,
618
],
"score": 1.0,
"content": "each modelling the mixture of latent slot variables and the attributes of one category, e.g. color.",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
616,
506,
628
],
"spans": [
{
"bbox": [
106,
616,
506,
628
],
"score": 1.0,
"content": "For ShapeWorld4, we thereby have altogether four NPPs. SLASH is trained via queries of the kind",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
627,
465,
639
],
"spans": [
{
"bbox": [
105,
627,
465,
639
],
"score": 1.0,
"content": "exemplified in Fig. 7 in the Appendix. We refer to this configuration as SLASH Attention.",
"type": "text"
}
],
"index": 50
}
],
"index": 47.5
},
{
"type": "text",
"bbox": [
107,
643,
505,
732
],
"lines": [
{
"bbox": [
106,
644,
505,
656
],
"spans": [
{
"bbox": [
106,
644,
505,
656
],
"score": 1.0,
"content": "We compare SLASH Attention to a baseline slot attention encoder using an MLP and Hungarian",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
655,
505,
667
],
"spans": [
{
"bbox": [
105,
655,
505,
667
],
"score": 1.0,
"content": "loss for predicting the object properties from the slot encodings as in Locatello et al. (2020). The",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
105,
666,
506,
678
],
"spans": [
{
"bbox": [
105,
666,
506,
678
],
"score": 1.0,
"content": "results of these experiments can be found in Fig. 3a (top). We observe that the average precision after",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
105,
677,
505,
689
],
"spans": [
{
"bbox": [
105,
677,
505,
689
],
"score": 1.0,
"content": "convergence on the held-out test set with SLASH Attention is greatly improved to that of the baseline",
"type": "text"
}
],
"index": 54
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "model. Additionally, in Fig. 3b we observe that SLASH Attention reaches the average precision",
"type": "text"
}
],
"index": 55
},
{
"bbox": [
105,
698,
505,
712
],
"spans": [
{
"bbox": [
105,
698,
505,
712
],
"score": 1.0,
"content": "value of the baseline model in much fewer number of epochs. Thus, we can summarize that adding",
"type": "text"
}
],
"index": 56
},
{
"bbox": [
105,
710,
506,
722
],
"spans": [
{
"bbox": [
105,
710,
506,
722
],
"score": 1.0,
"content": "logical knowledge in the training procedure via SLASH can greatly improve the capabilities of a",
"type": "text"
}
],
"index": 57
},
{
"bbox": [
105,
721,
239,
733
],
"spans": [
{
"bbox": [
105,
721,
239,
733
],
"score": 1.0,
"content": "neural module for set prediction.",
"type": "text"
}
],
"index": 58
}
],
"index": 54.5
}
],
"page_idx": 7,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"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"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
119,
106,
317,
183
],
"blocks": [
{
"type": "table_caption",
"bbox": [
119,
83,
288,
104
],
"group_id": 0,
"lines": [
{
"bbox": [
118,
83,
288,
93
],
"spans": [
{
"bbox": [
118,
83,
288,
93
],
"score": 0.982,
"content": "(a) ShapeWorld4 and ShapeWorld4 CoGenT Test",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
118,
93,
171,
105
],
"spans": [
{
"bbox": [
118,
93,
171,
105
],
"score": 1.0,
"content": "Avg.Precision",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "table_body",
"bbox": [
119,
106,
317,
183
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
106,
317,
183
],
"spans": [
{
"bbox": [
119,
106,
317,
183
],
"score": 0.976,
"html": " | Slot Att. | SLASH Att. |
| Test Set | ShapeWorld4 90.24 ± 0.93 | 95.58 ± 0.61 |
| CoGenT | |
| Test Cond. A | 90.37 ± 2.19 | 96.85 ± 0.43 |
| Test Cond.B | 27.15 ± 2.36 | 40.58 ± 1.99 |
",
"type": "table",
"image_path": "0b06c8954d45055a4eff3ce528d6a37b02d5a2f22cb9bd72bf07133146fdad99.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
119,
106,
317,
121.4
],
"spans": [],
"index": 2
},
{
"bbox": [
119,
121.4,
317,
136.8
],
"spans": [],
"index": 3
},
{
"bbox": [
119,
136.8,
317,
152.20000000000002
],
"spans": [],
"index": 4
},
{
"bbox": [
119,
152.20000000000002,
317,
167.60000000000002
],
"spans": [],
"index": 5
},
{
"bbox": [
119,
167.60000000000002,
317,
183.00000000000003
],
"spans": [],
"index": 6
}
]
},
{
"type": "table_caption",
"bbox": [
106,
193,
505,
282
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
194,
506,
207
],
"spans": [
{
"bbox": [
105,
194,
506,
207
],
"score": 1.0,
"content": "Figure 3: ShapeWorld4 Experiments. (a) Converged test average precision scores for the set prediction",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
204,
505,
218
],
"spans": [
{
"bbox": [
105,
204,
505,
218
],
"score": 1.0,
"content": "task with ShapeWorld4 (top) and ShapeWorld4 CoGenT (bottom). (b) Test average precision scores",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
215,
505,
228
],
"spans": [
{
"bbox": [
105,
215,
505,
228
],
"score": 1.0,
"content": "for set prediction with ShapeWorld4 over the training epochs. In these experiments we compared",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
227,
505,
238
],
"spans": [
{
"bbox": [
106,
227,
505,
238
],
"score": 1.0,
"content": "a baseline slot encoder versus SLASH Attention with slot attention and PC-based NPPs. For the",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
238,
506,
250
],
"spans": [
{
"bbox": [
105,
238,
506,
250
],
"score": 1.0,
"content": "CoGenT experiments, a model is trained on one training set and tested on two separate test conditions.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
248,
505,
261
],
"spans": [
{
"bbox": [
105,
248,
505,
261
],
"score": 1.0,
"content": "The Condition A test set contains attribute compositions which were also seen during training. The",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
259,
506,
273
],
"spans": [
{
"bbox": [
105,
259,
506,
273
],
"score": 1.0,
"content": "Condition B test set contains attribute compositions which were not seen during training, e.g. yellow",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
271,
418,
283
],
"spans": [
{
"bbox": [
106,
271,
418,
283
],
"score": 1.0,
"content": "circles were not present in the training set, but present in Condition B test set.",
"type": "text"
}
],
"index": 21
}
],
"index": 17.5
}
],
"index": 4
},
{
"type": "image",
"bbox": [
323,
84,
493,
171
],
"blocks": [
{
"type": "image_body",
"bbox": [
323,
84,
493,
171
],
"group_id": 0,
"lines": [
{
"bbox": [
323,
84,
493,
171
],
"spans": [
{
"bbox": [
323,
84,
493,
171
],
"score": 0.959,
"type": "image",
"image_path": "0eaf19403876aa7fbb58219357ca9c041a53e7685ffb819aff2c0d5f28605742.jpg"
}
]
}
],
"index": 9.5,
"virtual_lines": [
{
"bbox": [
323,
84,
493,
98.5
],
"spans": [],
"index": 7
},
{
"bbox": [
323,
98.5,
493,
113.0
],
"spans": [],
"index": 8
},
{
"bbox": [
323,
113.0,
493,
127.5
],
"spans": [],
"index": 9
},
{
"bbox": [
323,
127.5,
493,
142.0
],
"spans": [],
"index": 10
},
{
"bbox": [
323,
142.0,
493,
156.5
],
"spans": [],
"index": 11
},
{
"bbox": [
323,
156.5,
493,
171.0
],
"spans": [],
"index": 12
}
]
},
{
"type": "image_caption",
"bbox": [
333,
173,
485,
183
],
"group_id": 0,
"lines": [
{
"bbox": [
333,
173,
486,
184
],
"spans": [
{
"bbox": [
333,
173,
486,
184
],
"score": 0.98,
"content": "(b) Test Avg.Precision over Training Epochs",
"type": "text"
}
],
"index": 13
}
],
"index": 13
}
],
"index": 11.25
},
{
"type": "text",
"bbox": [
107,
295,
505,
384
],
"lines": [
{
"bbox": [
105,
294,
506,
309
],
"spans": [
{
"bbox": [
105,
294,
252,
309
],
"score": 1.0,
"content": "The results can be seen in Tab. 1b for",
"type": "text"
},
{
"bbox": [
252,
296,
272,
307
],
"score": 0.88,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
272,
294,
275,
309
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
275,
295,
294,
307
],
"score": 0.86,
"content": "8 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
295,
294,
298,
309
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
298,
296,
318,
307
],
"score": 0.86,
"content": "9 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
318,
294,
334,
309
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
335,
296,
354,
307
],
"score": 0.88,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
355,
294,
506,
309
],
"score": 1.0,
"content": "missing pixels per image. We observe",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
306,
506,
321
],
"spans": [
{
"bbox": [
105,
306,
133,
321
],
"score": 1.0,
"content": "that at",
"type": "text"
},
{
"bbox": [
133,
307,
153,
318
],
"score": 0.87,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
153,
306,
419,
321
],
"score": 1.0,
"content": ", DeepProbLog and SLASH produce almost equal accuracies. With",
"type": "text"
},
{
"bbox": [
420,
307,
439,
318
],
"score": 0.88,
"content": "8 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
440,
306,
506,
321
],
"score": 1.0,
"content": "percent missing",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
317,
506,
331
],
"spans": [
{
"bbox": [
105,
317,
506,
331
],
"score": 1.0,
"content": "pixels, there is a substantial difference in the ability of the two DPPLs to correctly classify images,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
329,
505,
341
],
"spans": [
{
"bbox": [
105,
329,
505,
341
],
"score": 1.0,
"content": "with SLASH being very stable. By further increasing the percentage of missing pixels, this difference",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
339,
506,
352
],
"spans": [
{
"bbox": [
105,
339,
352,
352
],
"score": 1.0,
"content": "becomes even more substantial with SLASH still reaching a",
"type": "text"
},
{
"bbox": [
352,
340,
372,
351
],
"score": 0.89,
"content": "8 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
372,
339,
473,
352
],
"score": 1.0,
"content": "test accuracy even when",
"type": "text"
},
{
"bbox": [
473,
340,
493,
351
],
"score": 0.88,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
493,
339,
506,
352
],
"score": 1.0,
"content": "of",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
349,
507,
365
],
"spans": [
{
"bbox": [
105,
349,
431,
365
],
"score": 1.0,
"content": "the pixels per image are missing, whereas DeepProbLog degrades to an average of",
"type": "text"
},
{
"bbox": [
432,
351,
451,
362
],
"score": 0.89,
"content": "3 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
451,
349,
507,
365
],
"score": 1.0,
"content": "test accuracy.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
361,
506,
375
],
"spans": [
{
"bbox": [
105,
361,
506,
375
],
"score": 1.0,
"content": "We further note that SLASH, in comparison to DeepProbLog, produces largely reduced standard",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
372,
192,
385
],
"spans": [
{
"bbox": [
106,
372,
192,
385
],
"score": 1.0,
"content": "deviations over runs.",
"type": "text"
}
],
"index": 29
}
],
"index": 25.5,
"bbox_fs": [
105,
294,
507,
385
]
},
{
"type": "text",
"bbox": [
107,
389,
505,
434
],
"lines": [
{
"bbox": [
106,
390,
505,
402
],
"spans": [
{
"bbox": [
106,
390,
505,
402
],
"score": 1.0,
"content": "Thus, by utilizing the power of true density estimation SLASH, with an appropriate NPP, can produce",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
401,
505,
413
],
"spans": [
{
"bbox": [
105,
401,
505,
413
],
"score": 1.0,
"content": "more robust results in comparison to other DPPLs. Further, we refer to Appendix D, which contains",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
412,
506,
425
],
"spans": [
{
"bbox": [
105,
412,
506,
425
],
"score": 1.0,
"content": "results of additional experiments where training is performed with the full MNIST data set whereas",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
423,
330,
435
],
"spans": [
{
"bbox": [
105,
423,
330,
435
],
"score": 1.0,
"content": "only the test set entails different rates of missing pixels.",
"type": "text"
}
],
"index": 33
}
],
"index": 31.5,
"bbox_fs": [
105,
390,
506,
435
]
},
{
"type": "text",
"bbox": [
107,
439,
505,
506
],
"lines": [
{
"bbox": [
106,
440,
506,
451
],
"spans": [
{
"bbox": [
106,
440,
506,
451
],
"score": 1.0,
"content": "Evaluation 3: Improved Concept Learning via SLASH. We show that SLASH can be very ef-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
451,
505,
462
],
"spans": [
{
"bbox": [
105,
451,
505,
462
],
"score": 1.0,
"content": "fective for the complex task of set prediction, which previous DPPLs have not tackled. We revert",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
461,
505,
474
],
"spans": [
{
"bbox": [
105,
461,
505,
474
],
"score": 1.0,
"content": "to the ShapeWorld4 data set for this setting. For set prediction, a model is trained to predict the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
471,
505,
485
],
"spans": [
{
"bbox": [
105,
471,
505,
485
],
"score": 1.0,
"content": "discrete attributes of a set of objects in an image (cf. Fig. 1 for an example ShapeWorld4 image). The",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
482,
505,
496
],
"spans": [
{
"bbox": [
105,
482,
505,
496
],
"score": 1.0,
"content": "difficulty for the model lies therein that it must match an unordered set of corresponding attributes",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
493,
478,
508
],
"spans": [
{
"bbox": [
105,
493,
478,
508
],
"score": 1.0,
"content": "(with varying number of entities over samples) with its internal representations of the image.",
"type": "text"
}
],
"index": 39
}
],
"index": 36.5,
"bbox_fs": [
105,
440,
506,
508
]
},
{
"type": "text",
"bbox": [
107,
511,
505,
566
],
"lines": [
{
"bbox": [
106,
511,
505,
523
],
"spans": [
{
"bbox": [
106,
511,
505,
523
],
"score": 1.0,
"content": "The slot attention module introduced by Locatello et al. (2020) allows for an attractive object-centric",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
522,
505,
535
],
"spans": [
{
"bbox": [
105,
522,
505,
535
],
"score": 1.0,
"content": "approach to this task. Specifically, this module represents a pluggable, differentiable module that can",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
533,
507,
546
],
"spans": [
{
"bbox": [
106,
533,
507,
546
],
"score": 1.0,
"content": "be easily added to any architecture and, through a competitive softmax-based attention mechanism,",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
545,
506,
557
],
"spans": [
{
"bbox": [
106,
545,
506,
557
],
"score": 1.0,
"content": "can enforce the binding of specific parts of a latent representation into permutation-invariant, task-",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
556,
222,
567
],
"spans": [
{
"bbox": [
106,
556,
222,
567
],
"score": 1.0,
"content": "specific vectors, called slots.",
"type": "text"
}
],
"index": 44
}
],
"index": 42,
"bbox_fs": [
105,
511,
507,
567
]
},
{
"type": "text",
"bbox": [
107,
572,
505,
638
],
"lines": [
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "In our experiments, we wish to show that by adding logical constraints to the training setting, one",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
582,
506,
596
],
"spans": [
{
"bbox": [
105,
582,
506,
596
],
"score": 1.0,
"content": "can improve the overall performances and generalization properties of such a model. For this, we",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
594,
506,
607
],
"spans": [
{
"bbox": [
105,
594,
506,
607
],
"score": 1.0,
"content": "train SLASH with NPPs as depicted in Fig. 1 consisting of a shared slot encoder and separate PCs,",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
604,
506,
618
],
"spans": [
{
"bbox": [
105,
604,
506,
618
],
"score": 1.0,
"content": "each modelling the mixture of latent slot variables and the attributes of one category, e.g. color.",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
616,
506,
628
],
"spans": [
{
"bbox": [
106,
616,
506,
628
],
"score": 1.0,
"content": "For ShapeWorld4, we thereby have altogether four NPPs. SLASH is trained via queries of the kind",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
627,
465,
639
],
"spans": [
{
"bbox": [
105,
627,
465,
639
],
"score": 1.0,
"content": "exemplified in Fig. 7 in the Appendix. We refer to this configuration as SLASH Attention.",
"type": "text"
}
],
"index": 50
}
],
"index": 47.5,
"bbox_fs": [
105,
572,
506,
639
]
},
{
"type": "text",
"bbox": [
107,
643,
505,
732
],
"lines": [
{
"bbox": [
106,
644,
505,
656
],
"spans": [
{
"bbox": [
106,
644,
505,
656
],
"score": 1.0,
"content": "We compare SLASH Attention to a baseline slot attention encoder using an MLP and Hungarian",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
655,
505,
667
],
"spans": [
{
"bbox": [
105,
655,
505,
667
],
"score": 1.0,
"content": "loss for predicting the object properties from the slot encodings as in Locatello et al. (2020). The",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
105,
666,
506,
678
],
"spans": [
{
"bbox": [
105,
666,
506,
678
],
"score": 1.0,
"content": "results of these experiments can be found in Fig. 3a (top). We observe that the average precision after",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
105,
677,
505,
689
],
"spans": [
{
"bbox": [
105,
677,
505,
689
],
"score": 1.0,
"content": "convergence on the held-out test set with SLASH Attention is greatly improved to that of the baseline",
"type": "text"
}
],
"index": 54
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "model. Additionally, in Fig. 3b we observe that SLASH Attention reaches the average precision",
"type": "text"
}
],
"index": 55
},
{
"bbox": [
105,
698,
505,
712
],
"spans": [
{
"bbox": [
105,
698,
505,
712
],
"score": 1.0,
"content": "value of the baseline model in much fewer number of epochs. Thus, we can summarize that adding",
"type": "text"
}
],
"index": 56
},
{
"bbox": [
105,
710,
506,
722
],
"spans": [
{
"bbox": [
105,
710,
506,
722
],
"score": 1.0,
"content": "logical knowledge in the training procedure via SLASH can greatly improve the capabilities of a",
"type": "text"
}
],
"index": 57
},
{
"bbox": [
105,
721,
239,
733
],
"spans": [
{
"bbox": [
105,
721,
239,
733
],
"score": 1.0,
"content": "neural module for set prediction.",
"type": "text"
}
],
"index": 58
}
],
"index": 54.5,
"bbox_fs": [
105,
644,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
137
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "Evaluation 4: Improved Compositional Generalization with SLASH. To test the hypothesis that",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
94,
506,
106
],
"spans": [
{
"bbox": [
106,
94,
506,
106
],
"score": 1.0,
"content": "SLASH Attention possesses improved generalization properties in comparison to the baseline model,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
104,
506,
117
],
"spans": [
{
"bbox": [
105,
104,
506,
117
],
"score": 1.0,
"content": "we ran experiments on a variant of ShapeWorld4 similar to the CLEVR Compositional Generalization",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
114,
505,
129
],
"spans": [
{
"bbox": [
105,
114,
505,
129
],
"score": 1.0,
"content": "Test (CoGenT) (Johnson et al., 2017). The goal of CoGenT is to investigate a model’s ability to",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
125,
405,
141
],
"spans": [
{
"bbox": [
105,
125,
405,
141
],
"score": 1.0,
"content": "handle novel combinations of attributes that were not seen during training.",
"type": "text"
}
],
"index": 4
}
],
"index": 2
},
{
"type": "text",
"bbox": [
106,
143,
505,
221
],
"lines": [
{
"bbox": [
106,
143,
505,
155
],
"spans": [
{
"bbox": [
106,
143,
505,
155
],
"score": 1.0,
"content": "For this purpose, we established two conditions within a ShapeWorld4 CoGenT data set: Condition",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
154,
506,
167
],
"spans": [
{
"bbox": [
105,
154,
506,
167
],
"score": 1.0,
"content": "(A) – the training and test data set contains squares with the colors gray, blue, brown, or yellow,",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
165,
505,
178
],
"spans": [
{
"bbox": [
105,
165,
505,
178
],
"score": 1.0,
"content": "triangles with the colors red, green, magenta, or cyan and circles of all colors. Condition (B) – the",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
176,
507,
190
],
"spans": [
{
"bbox": [
105,
176,
507,
190
],
"score": 1.0,
"content": "training set is as in Condition (A). However, the test set contains squares with the colors red, green,",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
505,
199
],
"spans": [
{
"bbox": [
105,
187,
505,
199
],
"score": 1.0,
"content": "magenta, or cyan, triangles with the colors gray, blue, brown, or yellow and circles of all colors. The",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
198,
507,
210
],
"spans": [
{
"bbox": [
105,
198,
507,
210
],
"score": 1.0,
"content": "goal is to investigate how well a model can generalize that, e.g., also squares can have the color red,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
209,
350,
223
],
"spans": [
{
"bbox": [
105,
209,
350,
223
],
"score": 1.0,
"content": "although never having seen evidence for this during training.",
"type": "text"
}
],
"index": 11
}
],
"index": 8
},
{
"type": "text",
"bbox": [
107,
226,
505,
292
],
"lines": [
{
"bbox": [
105,
225,
506,
239
],
"spans": [
{
"bbox": [
105,
225,
506,
239
],
"score": 1.0,
"content": "The resulting average precision test scores are presented in Fig. 3a (bottom). We observe that, even",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
236,
505,
250
],
"spans": [
{
"bbox": [
105,
236,
505,
250
],
"score": 1.0,
"content": "though the SLASH Program used for this experiment was not explicitly written to handle composition",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
104,
246,
506,
261
],
"spans": [
{
"bbox": [
104,
246,
506,
261
],
"score": 1.0,
"content": "generalization, SLASH Attention shows greatly improved generalization capabilities. This can be",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
259,
505,
272
],
"spans": [
{
"bbox": [
105,
259,
184,
272
],
"score": 1.0,
"content": "seen in the approx.",
"type": "text"
},
{
"bbox": [
184,
259,
204,
270
],
"score": 0.88,
"content": "1 3 \\%",
"type": "inline_equation"
},
{
"bbox": [
204,
259,
505,
272
],
"score": 1.0,
"content": "higher average precision scores on the Condition (B) test set in comparison",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
269,
505,
282
],
"spans": [
{
"bbox": [
105,
269,
505,
282
],
"score": 1.0,
"content": "to the baseline model. Importantly, this trend still holds even when subtracting the higher precision",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
281,
244,
293
],
"spans": [
{
"bbox": [
105,
281,
244,
293
],
"score": 1.0,
"content": "scores observed in Condition (A).",
"type": "text"
}
],
"index": 17
}
],
"index": 14.5
},
{
"type": "text",
"bbox": [
107,
298,
505,
375
],
"lines": [
{
"bbox": [
105,
297,
506,
311
],
"spans": [
{
"bbox": [
105,
297,
506,
311
],
"score": 1.0,
"content": "To summarize our findings from the experiments on set prediction: we observe that adding prior",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
309,
505,
321
],
"spans": [
{
"bbox": [
106,
309,
505,
321
],
"score": 1.0,
"content": "knowledge in the form of logical constraints via SLASH can greatly improve a neural module in",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
320,
506,
332
],
"spans": [
{
"bbox": [
106,
320,
506,
332
],
"score": 1.0,
"content": "terms of performance and generalizability. On a side note: training neural networks for novel tasks,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
331,
506,
343
],
"spans": [
{
"bbox": [
106,
331,
506,
343
],
"score": 1.0,
"content": "often involves defining explicit loss functions, e.g. Hungarian loss for set prediction. In contrast",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
342,
506,
354
],
"spans": [
{
"bbox": [
105,
342,
506,
354
],
"score": 1.0,
"content": "with SLASH, no matter the choice of NPP and underlying task, the training loss remains the same.",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
352,
505,
365
],
"spans": [
{
"bbox": [
105,
352,
505,
365
],
"score": 1.0,
"content": "Task-related requirements simply need to be added as lines of code to the SLASH program. This",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
363,
342,
376
],
"spans": [
{
"bbox": [
105,
363,
342,
376
],
"score": 1.0,
"content": "additionally highlights SLASH’s versatility and flexibility.",
"type": "text"
}
],
"index": 24
}
],
"index": 21
},
{
"type": "text",
"bbox": [
107,
380,
505,
425
],
"lines": [
{
"bbox": [
105,
380,
506,
393
],
"spans": [
{
"bbox": [
105,
380,
506,
393
],
"score": 1.0,
"content": "Summary of all Empirical Results. All empirical results together demonstrate that the flexibility of",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
391,
506,
404
],
"spans": [
{
"bbox": [
105,
391,
506,
404
],
"score": 1.0,
"content": "SLASH is highly beneficial and can easily outperform state-of-the-art: one can freely combine what",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
403,
505,
416
],
"spans": [
{
"bbox": [
105,
403,
505,
416
],
"score": 1.0,
"content": "is required to solve the underlying task — (deep) neural networks, PCs, and logic. Particularly, the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
414,
395,
425
],
"spans": [
{
"bbox": [
106,
414,
395,
425
],
"score": 1.0,
"content": "results indicate the potential of integrating PCs via SLASH into DPPLs.",
"type": "text"
}
],
"index": 28
}
],
"index": 26.5
},
{
"type": "title",
"bbox": [
107,
465,
302,
479
],
"lines": [
{
"bbox": [
105,
465,
304,
480
],
"spans": [
{
"bbox": [
105,
465,
304,
480
],
"score": 1.0,
"content": "5 CONCLUSION AND FUTURE WORK",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
506,
505,
660
],
"lines": [
{
"bbox": [
105,
505,
507,
519
],
"spans": [
{
"bbox": [
105,
505,
507,
519
],
"score": 1.0,
"content": "We introduce SLASH, a novel DPPL that integrates neural computations with tractable probability es-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
518,
505,
529
],
"spans": [
{
"bbox": [
106,
518,
505,
529
],
"score": 1.0,
"content": "timates and logical statements. The key ingredient of SLASH to achieve this are Neural-Probabilistic",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
528,
505,
540
],
"spans": [
{
"bbox": [
105,
528,
505,
540
],
"score": 1.0,
"content": "Predicates (NPPs) that can be flexibly constructed out of neural and/or probabilistic circuit modules",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
538,
506,
553
],
"spans": [
{
"bbox": [
105,
538,
506,
553
],
"score": 1.0,
"content": "based on the data and underlying task. With these NPPs, one can produce task-specific probability",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
550,
506,
563
],
"spans": [
{
"bbox": [
106,
550,
506,
563
],
"score": 1.0,
"content": "estimates. The details and additional prior knowledge of a task are neatly encompassed within a",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
561,
506,
574
],
"spans": [
{
"bbox": [
105,
561,
506,
574
],
"score": 1.0,
"content": "SLASH program with only few lines of code. Finally, via Answer Set Programming and Weighted",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
571,
506,
585
],
"spans": [
{
"bbox": [
105,
571,
506,
585
],
"score": 1.0,
"content": "Model Counting, the logical SLASH program and probability estimates from the NPPs are combined",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
582,
506,
597
],
"spans": [
{
"bbox": [
105,
582,
506,
597
],
"score": 1.0,
"content": "to estimate the truth value of a task-specific query. Our experiments show the power and efficiency",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
594,
506,
607
],
"spans": [
{
"bbox": [
105,
594,
506,
607
],
"score": 1.0,
"content": "of SLASH, improving upon previous DPPLs in the benchmark MNIST-Addition task in terms of",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
605,
506,
618
],
"spans": [
{
"bbox": [
105,
605,
506,
618
],
"score": 1.0,
"content": "performance, efficiency and robustness. Importantly, by integrating a SOTA slot attention encoder",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
616,
506,
629
],
"spans": [
{
"bbox": [
106,
616,
506,
629
],
"score": 1.0,
"content": "into NPPs and adding few logical constraints, SLASH demonstrates improved performances and",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
627,
506,
640
],
"spans": [
{
"bbox": [
105,
627,
506,
640
],
"score": 1.0,
"content": "generalizability in comparison to the pure slot encoder for the task of object-centric set prediction;",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
637,
505,
650
],
"spans": [
{
"bbox": [
105,
637,
505,
650
],
"score": 1.0,
"content": "a setting no DPPL has tackled yet. This shows the great potential of DPPLs to elegantly combine",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
649,
387,
662
],
"spans": [
{
"bbox": [
106,
649,
387,
662
],
"score": 1.0,
"content": "logical reasoning with neural computations and uncertainty estimates.",
"type": "text"
}
],
"index": 43
}
],
"index": 36.5
},
{
"type": "text",
"bbox": [
107,
666,
505,
732
],
"lines": [
{
"bbox": [
105,
665,
505,
678
],
"spans": [
{
"bbox": [
105,
665,
505,
678
],
"score": 1.0,
"content": "Interesting avenues for future work include benchmarking SLASH on additional data types and",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
677,
505,
690
],
"spans": [
{
"bbox": [
105,
677,
505,
690
],
"score": 1.0,
"content": "tasks. One should explore unsupervised and weakly supervised learning using logic with SLASH and",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
106,
688,
505,
700
],
"spans": [
{
"bbox": [
106,
688,
505,
700
],
"score": 1.0,
"content": "investigate how far logical constraints can help unsupervised object discovery. In direct alignment",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
699,
505,
711
],
"spans": [
{
"bbox": [
105,
699,
505,
711
],
"score": 1.0,
"content": "with our work, one should also investigate image generation via the beneficial feature of PCs to",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
506,
722
],
"score": 1.0,
"content": "generate random samples. Actually, it should be possible to generate images that encapsulate logical",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
721,
435,
732
],
"spans": [
{
"bbox": [
106,
721,
435,
732
],
"score": 1.0,
"content": "knowledge bases. This is important to move from data-rich to knowledge-rich AI.",
"type": "text"
}
],
"index": 49
}
],
"index": 46.5
}
],
"page_idx": 8,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
308,
38
],
"spans": [
{
"bbox": [
106,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
759
],
"lines": [
{
"bbox": [
302,
751,
309,
762
],
"spans": [
{
"bbox": [
302,
751,
309,
762
],
"score": 1.0,
"content": "9",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
137
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "Evaluation 4: Improved Compositional Generalization with SLASH. To test the hypothesis that",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
94,
506,
106
],
"spans": [
{
"bbox": [
106,
94,
506,
106
],
"score": 1.0,
"content": "SLASH Attention possesses improved generalization properties in comparison to the baseline model,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
104,
506,
117
],
"spans": [
{
"bbox": [
105,
104,
506,
117
],
"score": 1.0,
"content": "we ran experiments on a variant of ShapeWorld4 similar to the CLEVR Compositional Generalization",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
114,
505,
129
],
"spans": [
{
"bbox": [
105,
114,
505,
129
],
"score": 1.0,
"content": "Test (CoGenT) (Johnson et al., 2017). The goal of CoGenT is to investigate a model’s ability to",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
125,
405,
141
],
"spans": [
{
"bbox": [
105,
125,
405,
141
],
"score": 1.0,
"content": "handle novel combinations of attributes that were not seen during training.",
"type": "text"
}
],
"index": 4
}
],
"index": 2,
"bbox_fs": [
105,
83,
506,
141
]
},
{
"type": "text",
"bbox": [
106,
143,
505,
221
],
"lines": [
{
"bbox": [
106,
143,
505,
155
],
"spans": [
{
"bbox": [
106,
143,
505,
155
],
"score": 1.0,
"content": "For this purpose, we established two conditions within a ShapeWorld4 CoGenT data set: Condition",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
154,
506,
167
],
"spans": [
{
"bbox": [
105,
154,
506,
167
],
"score": 1.0,
"content": "(A) – the training and test data set contains squares with the colors gray, blue, brown, or yellow,",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
165,
505,
178
],
"spans": [
{
"bbox": [
105,
165,
505,
178
],
"score": 1.0,
"content": "triangles with the colors red, green, magenta, or cyan and circles of all colors. Condition (B) – the",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
176,
507,
190
],
"spans": [
{
"bbox": [
105,
176,
507,
190
],
"score": 1.0,
"content": "training set is as in Condition (A). However, the test set contains squares with the colors red, green,",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
505,
199
],
"spans": [
{
"bbox": [
105,
187,
505,
199
],
"score": 1.0,
"content": "magenta, or cyan, triangles with the colors gray, blue, brown, or yellow and circles of all colors. The",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
198,
507,
210
],
"spans": [
{
"bbox": [
105,
198,
507,
210
],
"score": 1.0,
"content": "goal is to investigate how well a model can generalize that, e.g., also squares can have the color red,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
209,
350,
223
],
"spans": [
{
"bbox": [
105,
209,
350,
223
],
"score": 1.0,
"content": "although never having seen evidence for this during training.",
"type": "text"
}
],
"index": 11
}
],
"index": 8,
"bbox_fs": [
105,
143,
507,
223
]
},
{
"type": "text",
"bbox": [
107,
226,
505,
292
],
"lines": [
{
"bbox": [
105,
225,
506,
239
],
"spans": [
{
"bbox": [
105,
225,
506,
239
],
"score": 1.0,
"content": "The resulting average precision test scores are presented in Fig. 3a (bottom). We observe that, even",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
236,
505,
250
],
"spans": [
{
"bbox": [
105,
236,
505,
250
],
"score": 1.0,
"content": "though the SLASH Program used for this experiment was not explicitly written to handle composition",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
104,
246,
506,
261
],
"spans": [
{
"bbox": [
104,
246,
506,
261
],
"score": 1.0,
"content": "generalization, SLASH Attention shows greatly improved generalization capabilities. This can be",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
259,
505,
272
],
"spans": [
{
"bbox": [
105,
259,
184,
272
],
"score": 1.0,
"content": "seen in the approx.",
"type": "text"
},
{
"bbox": [
184,
259,
204,
270
],
"score": 0.88,
"content": "1 3 \\%",
"type": "inline_equation"
},
{
"bbox": [
204,
259,
505,
272
],
"score": 1.0,
"content": "higher average precision scores on the Condition (B) test set in comparison",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
269,
505,
282
],
"spans": [
{
"bbox": [
105,
269,
505,
282
],
"score": 1.0,
"content": "to the baseline model. Importantly, this trend still holds even when subtracting the higher precision",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
281,
244,
293
],
"spans": [
{
"bbox": [
105,
281,
244,
293
],
"score": 1.0,
"content": "scores observed in Condition (A).",
"type": "text"
}
],
"index": 17
}
],
"index": 14.5,
"bbox_fs": [
104,
225,
506,
293
]
},
{
"type": "text",
"bbox": [
107,
298,
505,
375
],
"lines": [
{
"bbox": [
105,
297,
506,
311
],
"spans": [
{
"bbox": [
105,
297,
506,
311
],
"score": 1.0,
"content": "To summarize our findings from the experiments on set prediction: we observe that adding prior",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
309,
505,
321
],
"spans": [
{
"bbox": [
106,
309,
505,
321
],
"score": 1.0,
"content": "knowledge in the form of logical constraints via SLASH can greatly improve a neural module in",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
320,
506,
332
],
"spans": [
{
"bbox": [
106,
320,
506,
332
],
"score": 1.0,
"content": "terms of performance and generalizability. On a side note: training neural networks for novel tasks,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
331,
506,
343
],
"spans": [
{
"bbox": [
106,
331,
506,
343
],
"score": 1.0,
"content": "often involves defining explicit loss functions, e.g. Hungarian loss for set prediction. In contrast",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
342,
506,
354
],
"spans": [
{
"bbox": [
105,
342,
506,
354
],
"score": 1.0,
"content": "with SLASH, no matter the choice of NPP and underlying task, the training loss remains the same.",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
352,
505,
365
],
"spans": [
{
"bbox": [
105,
352,
505,
365
],
"score": 1.0,
"content": "Task-related requirements simply need to be added as lines of code to the SLASH program. This",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
363,
342,
376
],
"spans": [
{
"bbox": [
105,
363,
342,
376
],
"score": 1.0,
"content": "additionally highlights SLASH’s versatility and flexibility.",
"type": "text"
}
],
"index": 24
}
],
"index": 21,
"bbox_fs": [
105,
297,
506,
376
]
},
{
"type": "text",
"bbox": [
107,
380,
505,
425
],
"lines": [
{
"bbox": [
105,
380,
506,
393
],
"spans": [
{
"bbox": [
105,
380,
506,
393
],
"score": 1.0,
"content": "Summary of all Empirical Results. All empirical results together demonstrate that the flexibility of",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
391,
506,
404
],
"spans": [
{
"bbox": [
105,
391,
506,
404
],
"score": 1.0,
"content": "SLASH is highly beneficial and can easily outperform state-of-the-art: one can freely combine what",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
403,
505,
416
],
"spans": [
{
"bbox": [
105,
403,
505,
416
],
"score": 1.0,
"content": "is required to solve the underlying task — (deep) neural networks, PCs, and logic. Particularly, the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
414,
395,
425
],
"spans": [
{
"bbox": [
106,
414,
395,
425
],
"score": 1.0,
"content": "results indicate the potential of integrating PCs via SLASH into DPPLs.",
"type": "text"
}
],
"index": 28
}
],
"index": 26.5,
"bbox_fs": [
105,
380,
506,
425
]
},
{
"type": "title",
"bbox": [
107,
465,
302,
479
],
"lines": [
{
"bbox": [
105,
465,
304,
480
],
"spans": [
{
"bbox": [
105,
465,
304,
480
],
"score": 1.0,
"content": "5 CONCLUSION AND FUTURE WORK",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
506,
505,
660
],
"lines": [
{
"bbox": [
105,
505,
507,
519
],
"spans": [
{
"bbox": [
105,
505,
507,
519
],
"score": 1.0,
"content": "We introduce SLASH, a novel DPPL that integrates neural computations with tractable probability es-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
518,
505,
529
],
"spans": [
{
"bbox": [
106,
518,
505,
529
],
"score": 1.0,
"content": "timates and logical statements. The key ingredient of SLASH to achieve this are Neural-Probabilistic",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
528,
505,
540
],
"spans": [
{
"bbox": [
105,
528,
505,
540
],
"score": 1.0,
"content": "Predicates (NPPs) that can be flexibly constructed out of neural and/or probabilistic circuit modules",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
538,
506,
553
],
"spans": [
{
"bbox": [
105,
538,
506,
553
],
"score": 1.0,
"content": "based on the data and underlying task. With these NPPs, one can produce task-specific probability",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
550,
506,
563
],
"spans": [
{
"bbox": [
106,
550,
506,
563
],
"score": 1.0,
"content": "estimates. The details and additional prior knowledge of a task are neatly encompassed within a",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
561,
506,
574
],
"spans": [
{
"bbox": [
105,
561,
506,
574
],
"score": 1.0,
"content": "SLASH program with only few lines of code. Finally, via Answer Set Programming and Weighted",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
571,
506,
585
],
"spans": [
{
"bbox": [
105,
571,
506,
585
],
"score": 1.0,
"content": "Model Counting, the logical SLASH program and probability estimates from the NPPs are combined",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
582,
506,
597
],
"spans": [
{
"bbox": [
105,
582,
506,
597
],
"score": 1.0,
"content": "to estimate the truth value of a task-specific query. Our experiments show the power and efficiency",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
594,
506,
607
],
"spans": [
{
"bbox": [
105,
594,
506,
607
],
"score": 1.0,
"content": "of SLASH, improving upon previous DPPLs in the benchmark MNIST-Addition task in terms of",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
605,
506,
618
],
"spans": [
{
"bbox": [
105,
605,
506,
618
],
"score": 1.0,
"content": "performance, efficiency and robustness. Importantly, by integrating a SOTA slot attention encoder",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
616,
506,
629
],
"spans": [
{
"bbox": [
106,
616,
506,
629
],
"score": 1.0,
"content": "into NPPs and adding few logical constraints, SLASH demonstrates improved performances and",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
627,
506,
640
],
"spans": [
{
"bbox": [
105,
627,
506,
640
],
"score": 1.0,
"content": "generalizability in comparison to the pure slot encoder for the task of object-centric set prediction;",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
637,
505,
650
],
"spans": [
{
"bbox": [
105,
637,
505,
650
],
"score": 1.0,
"content": "a setting no DPPL has tackled yet. This shows the great potential of DPPLs to elegantly combine",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
649,
387,
662
],
"spans": [
{
"bbox": [
106,
649,
387,
662
],
"score": 1.0,
"content": "logical reasoning with neural computations and uncertainty estimates.",
"type": "text"
}
],
"index": 43
}
],
"index": 36.5,
"bbox_fs": [
105,
505,
507,
662
]
},
{
"type": "text",
"bbox": [
107,
666,
505,
732
],
"lines": [
{
"bbox": [
105,
665,
505,
678
],
"spans": [
{
"bbox": [
105,
665,
505,
678
],
"score": 1.0,
"content": "Interesting avenues for future work include benchmarking SLASH on additional data types and",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
677,
505,
690
],
"spans": [
{
"bbox": [
105,
677,
505,
690
],
"score": 1.0,
"content": "tasks. One should explore unsupervised and weakly supervised learning using logic with SLASH and",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
106,
688,
505,
700
],
"spans": [
{
"bbox": [
106,
688,
505,
700
],
"score": 1.0,
"content": "investigate how far logical constraints can help unsupervised object discovery. In direct alignment",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
699,
505,
711
],
"spans": [
{
"bbox": [
105,
699,
505,
711
],
"score": 1.0,
"content": "with our work, one should also investigate image generation via the beneficial feature of PCs to",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
506,
722
],
"score": 1.0,
"content": "generate random samples. Actually, it should be possible to generate images that encapsulate logical",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
721,
435,
732
],
"spans": [
{
"bbox": [
106,
721,
435,
732
],
"score": 1.0,
"content": "knowledge bases. This is important to move from data-rich to knowledge-rich AI.",
"type": "text"
}
],
"index": 49
}
],
"index": 46.5,
"bbox_fs": [
105,
665,
506,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
81,
211,
94
],
"lines": [
{
"bbox": [
106,
81,
212,
95
],
"spans": [
{
"bbox": [
106,
81,
212,
95
],
"score": 1.0,
"content": "ETHICS STATEMENT",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
105,
505,
172
],
"lines": [
{
"bbox": [
106,
106,
505,
118
],
"spans": [
{
"bbox": [
106,
106,
505,
118
],
"score": 1.0,
"content": "With our work, we have shown that one can add prior knowledge and logical constraints to the",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
117,
505,
129
],
"spans": [
{
"bbox": [
106,
117,
505,
129
],
"score": 1.0,
"content": "training of learning systems. We postulate that SLASH can therefore additionally be used to identify",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
127,
506,
141
],
"spans": [
{
"bbox": [
105,
127,
506,
141
],
"score": 1.0,
"content": "and remove biases or undesirable behavior, by adding constraints within the SLASH program. We",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
137,
507,
153
],
"spans": [
{
"bbox": [
105,
137,
507,
153
],
"score": 1.0,
"content": "observe that this feature, however, also has the potential danger to be used in the opposite way, e.g.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
150,
505,
163
],
"spans": [
{
"bbox": [
106,
150,
505,
163
],
"score": 1.0,
"content": "explicitly adding bias and discriminatory factors to a system. To the best of our knowledge, our study",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
162,
370,
173
],
"spans": [
{
"bbox": [
106,
162,
370,
173
],
"score": 1.0,
"content": "does not raise any ethical, privacy or conflict of interest concerns.",
"type": "text"
}
],
"index": 6
}
],
"index": 3.5
},
{
"type": "title",
"bbox": [
108,
188,
267,
200
],
"lines": [
{
"bbox": [
106,
187,
270,
202
],
"spans": [
{
"bbox": [
106,
187,
270,
202
],
"score": 1.0,
"content": "REPRODUCIBILITY STATEMENT",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
107,
213,
505,
268
],
"lines": [
{
"bbox": [
106,
213,
504,
225
],
"spans": [
{
"bbox": [
106,
213,
504,
225
],
"score": 1.0,
"content": "An official, curated GitHub repository will be made public with the final version, containing the code",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
223,
506,
236
],
"spans": [
{
"bbox": [
105,
223,
506,
236
],
"score": 1.0,
"content": "of SLASH, as well as scripts to reproduce the experiments and generate data sets. In addition to this,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
235,
505,
247
],
"spans": [
{
"bbox": [
106,
235,
505,
247
],
"score": 1.0,
"content": "architectural details and hyperparameters are included in the appendix. Preliminary code will be",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
246,
505,
259
],
"spans": [
{
"bbox": [
106,
246,
505,
259
],
"score": 1.0,
"content": "uploaded upon submission. Lastly, details on the evaluation metrics and relevant data sets are given",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
257,
253,
269
],
"spans": [
{
"bbox": [
106,
257,
253,
269
],
"score": 1.0,
"content": "in the main text as well as appendix.",
"type": "text"
}
],
"index": 12
}
],
"index": 10
},
{
"type": "title",
"bbox": [
108,
285,
175,
297
],
"lines": [
{
"bbox": [
106,
284,
176,
298
],
"spans": [
{
"bbox": [
106,
284,
176,
298
],
"score": 1.0,
"content": "REFERENCES",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
303,
504,
326
],
"lines": [
{
"bbox": [
106,
302,
506,
316
],
"spans": [
{
"bbox": [
106,
302,
506,
316
],
"score": 1.0,
"content": "Yoshua Bengio. From System 1 Deep Learning to System 2 Deep Learning. Invited talk NeurIPS,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
116,
313,
426,
327
],
"spans": [
{
"bbox": [
116,
313,
350,
327
],
"score": 1.0,
"content": "2019. URL https://www.youtube.com/watch?v",
"type": "text"
},
{
"bbox": [
351,
316,
357,
324
],
"score": 0.29,
"content": "=",
"type": "inline_equation"
},
{
"bbox": [
358,
313,
426,
327
],
"score": 1.0,
"content": "T3sxeTgT4qc.",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5
},
{
"type": "text",
"bbox": [
107,
332,
506,
366
],
"lines": [
{
"bbox": [
106,
333,
505,
345
],
"spans": [
{
"bbox": [
106,
333,
505,
345
],
"score": 1.0,
"content": "Eli Bingham, Jonathan P. Chen, Martin Jankowiak, Fritz Obermeyer, Neeraj Pradhan, Theofanis",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
115,
342,
506,
358
],
"spans": [
{
"bbox": [
115,
342,
506,
358
],
"score": 1.0,
"content": "Karaletsos, Rohit Singh, Paul A. Szerlip, Paul Horsfall, and Noah D. Goodman. Pyro: Deep",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
116,
354,
507,
367
],
"spans": [
{
"bbox": [
116,
354,
507,
367
],
"score": 1.0,
"content": "universal probabilistic programming. Journal of Machine Learning Research, pp. 28:1–28:6, 2019.",
"type": "text"
}
],
"index": 18
}
],
"index": 17
},
{
"type": "text",
"bbox": [
107,
372,
506,
406
],
"lines": [
{
"bbox": [
105,
372,
505,
386
],
"spans": [
{
"bbox": [
105,
372,
505,
386
],
"score": 1.0,
"content": "Christopher P. Burgess, Loïc Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
115,
383,
506,
397
],
"spans": [
{
"bbox": [
115,
383,
506,
397
],
"score": 1.0,
"content": "Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
116,
395,
171,
406
],
"spans": [
{
"bbox": [
116,
395,
171,
406
],
"score": 1.0,
"content": "CoRR, 2019.",
"type": "text"
}
],
"index": 21
}
],
"index": 20
},
{
"type": "text",
"bbox": [
106,
413,
506,
457
],
"lines": [
{
"bbox": [
106,
414,
506,
425
],
"spans": [
{
"bbox": [
106,
414,
506,
425
],
"score": 1.0,
"content": "Francesco Calimeri, Wolfgang Faber, Martin Gebser, Giovambattista Ianni, Roland Kaminski,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
115,
424,
506,
437
],
"spans": [
{
"bbox": [
115,
424,
506,
437
],
"score": 1.0,
"content": "Thomas Krennwallner, Nicola Leone, Marco Maratea, Francesco Ricca, and Torsten Schaub.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
435,
507,
448
],
"spans": [
{
"bbox": [
115,
435,
507,
448
],
"score": 1.0,
"content": "Asp-core-2 input language format. Theory and Practice of Logic Programming, pp. 294–309,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
114,
445,
142,
458
],
"spans": [
{
"bbox": [
114,
445,
142,
458
],
"score": 1.0,
"content": "2020.",
"type": "text"
}
],
"index": 25
}
],
"index": 23.5
},
{
"type": "text",
"bbox": [
108,
465,
504,
488
],
"lines": [
{
"bbox": [
106,
464,
506,
479
],
"spans": [
{
"bbox": [
106,
464,
506,
479
],
"score": 1.0,
"content": "YooJung Choi, Antonio Vergari, and Guy Van den Broeck. Probabilistic circuits: A unifying",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
116,
476,
465,
488
],
"spans": [
{
"bbox": [
116,
476,
465,
488
],
"score": 1.0,
"content": "framework for tractable probabilistic models. Technical report, Technical report, 2020.",
"type": "text"
}
],
"index": 27
}
],
"index": 26.5
},
{
"type": "text",
"bbox": [
106,
495,
505,
517
],
"lines": [
{
"bbox": [
106,
494,
505,
507
],
"spans": [
{
"bbox": [
106,
494,
505,
507
],
"score": 1.0,
"content": "Gabriele Ciravegna, Francesco Giannini, Marco Gori, Marco Maggini, and Stefano Melacci. Human-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
505,
417,
518
],
"spans": [
{
"bbox": [
115,
505,
417,
518
],
"score": 1.0,
"content": "driven FOL explanations of deep learning. In IJCAI, pp. 2234–2240, 2020.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
},
{
"type": "text",
"bbox": [
106,
524,
370,
536
],
"lines": [
{
"bbox": [
106,
524,
371,
537
],
"spans": [
{
"bbox": [
106,
524,
371,
537
],
"score": 1.0,
"content": "W. F. Clocksin and Chris Mellish. Programming in Prolog. 1981.",
"type": "text"
}
],
"index": 30
}
],
"index": 30
},
{
"type": "text",
"bbox": [
109,
543,
475,
555
],
"lines": [
{
"bbox": [
106,
542,
475,
556
],
"spans": [
{
"bbox": [
106,
542,
475,
556
],
"score": 1.0,
"content": "Alain Colmerauer and Philippe Roussel. The birth of prolog. In HOPL-II, pp. 37–52, 1993.",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
108,
561,
506,
584
],
"lines": [
{
"bbox": [
105,
560,
506,
575
],
"spans": [
{
"bbox": [
105,
560,
506,
575
],
"score": 1.0,
"content": "Adnan Darwiche. SDD: A new canonical representation of propositional knowledge bases. In IJCAI,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
573,
197,
584
],
"spans": [
{
"bbox": [
115,
573,
197,
584
],
"score": 1.0,
"content": "pp. 819–826, 2011.",
"type": "text"
}
],
"index": 33
}
],
"index": 32.5
},
{
"type": "text",
"bbox": [
106,
591,
462,
603
],
"lines": [
{
"bbox": [
106,
590,
462,
604
],
"spans": [
{
"bbox": [
106,
590,
462,
604
],
"score": 1.0,
"content": "Artur d’Avila Garcez and Luís C. Lamb. Neurosymbolic AI: the 3rd wave. CoRR, 2020.",
"type": "text"
}
],
"index": 34
}
],
"index": 34
},
{
"type": "text",
"bbox": [
108,
609,
504,
632
],
"lines": [
{
"bbox": [
105,
607,
507,
624
],
"spans": [
{
"bbox": [
105,
607,
507,
624
],
"score": 1.0,
"content": "Artur S. d’Avila Garcez, Luís C. Lamb, and Dov M. Gabbay. Neural-Symbolic Cognitive Reasoning.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
116,
621,
241,
633
],
"spans": [
{
"bbox": [
116,
621,
241,
633
],
"score": 1.0,
"content": "Cognitive Technologies. 2009.",
"type": "text"
}
],
"index": 36
}
],
"index": 35.5
},
{
"type": "text",
"bbox": [
107,
639,
505,
673
],
"lines": [
{
"bbox": [
106,
639,
506,
651
],
"spans": [
{
"bbox": [
106,
639,
506,
651
],
"score": 1.0,
"content": "Artur S. d’Avila Garcez, Marco Gori, Luís C. Lamb, Luciano Serafini, Michael Spranger, and Son N.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
115,
650,
505,
663
],
"spans": [
{
"bbox": [
115,
650,
505,
663
],
"score": 1.0,
"content": "Tran. Neural-symbolic computing: An effective methodology for principled integration of machine",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
116,
662,
322,
674
],
"spans": [
{
"bbox": [
116,
662,
322,
674
],
"score": 1.0,
"content": "learning and reasoning. FLAP, pp. 611–632, 2019.",
"type": "text"
}
],
"index": 39
}
],
"index": 38
},
{
"type": "text",
"bbox": [
108,
680,
504,
703
],
"lines": [
{
"bbox": [
106,
679,
505,
693
],
"spans": [
{
"bbox": [
106,
679,
505,
693
],
"score": 1.0,
"content": "Yannis Dimopoulos, Bernhard Nebel, and Jana Koehler. Encoding planning problems in nonmono-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
115,
691,
466,
704
],
"spans": [
{
"bbox": [
115,
691,
466,
704
],
"score": 1.0,
"content": "tonic logic programs. In ECP, Lecture Notes in Computer Science, pp. 169–181, 1997.",
"type": "text"
}
],
"index": 41
}
],
"index": 40.5
},
{
"type": "text",
"bbox": [
108,
709,
503,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Martin Engelcke, Adam R. Kosiorek, Oiwi Parker Jones, and Ingmar Posner. GENESIS: generative",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
115,
721,
468,
732
],
"spans": [
{
"bbox": [
115,
721,
468,
732
],
"score": 1.0,
"content": "scene inference and sampling with object-centric latent representations. In ICLR, 2020.",
"type": "text"
}
],
"index": 43
}
],
"index": 42.5
}
],
"page_idx": 9,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"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": "10",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
81,
211,
94
],
"lines": [
{
"bbox": [
106,
81,
212,
95
],
"spans": [
{
"bbox": [
106,
81,
212,
95
],
"score": 1.0,
"content": "ETHICS STATEMENT",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
105,
505,
172
],
"lines": [
{
"bbox": [
106,
106,
505,
118
],
"spans": [
{
"bbox": [
106,
106,
505,
118
],
"score": 1.0,
"content": "With our work, we have shown that one can add prior knowledge and logical constraints to the",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
117,
505,
129
],
"spans": [
{
"bbox": [
106,
117,
505,
129
],
"score": 1.0,
"content": "training of learning systems. We postulate that SLASH can therefore additionally be used to identify",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
127,
506,
141
],
"spans": [
{
"bbox": [
105,
127,
506,
141
],
"score": 1.0,
"content": "and remove biases or undesirable behavior, by adding constraints within the SLASH program. We",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
137,
507,
153
],
"spans": [
{
"bbox": [
105,
137,
507,
153
],
"score": 1.0,
"content": "observe that this feature, however, also has the potential danger to be used in the opposite way, e.g.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
150,
505,
163
],
"spans": [
{
"bbox": [
106,
150,
505,
163
],
"score": 1.0,
"content": "explicitly adding bias and discriminatory factors to a system. To the best of our knowledge, our study",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
162,
370,
173
],
"spans": [
{
"bbox": [
106,
162,
370,
173
],
"score": 1.0,
"content": "does not raise any ethical, privacy or conflict of interest concerns.",
"type": "text"
}
],
"index": 6
}
],
"index": 3.5,
"bbox_fs": [
105,
106,
507,
173
]
},
{
"type": "title",
"bbox": [
108,
188,
267,
200
],
"lines": [
{
"bbox": [
106,
187,
270,
202
],
"spans": [
{
"bbox": [
106,
187,
270,
202
],
"score": 1.0,
"content": "REPRODUCIBILITY STATEMENT",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
107,
213,
505,
268
],
"lines": [
{
"bbox": [
106,
213,
504,
225
],
"spans": [
{
"bbox": [
106,
213,
504,
225
],
"score": 1.0,
"content": "An official, curated GitHub repository will be made public with the final version, containing the code",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
223,
506,
236
],
"spans": [
{
"bbox": [
105,
223,
506,
236
],
"score": 1.0,
"content": "of SLASH, as well as scripts to reproduce the experiments and generate data sets. In addition to this,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
235,
505,
247
],
"spans": [
{
"bbox": [
106,
235,
505,
247
],
"score": 1.0,
"content": "architectural details and hyperparameters are included in the appendix. Preliminary code will be",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
246,
505,
259
],
"spans": [
{
"bbox": [
106,
246,
505,
259
],
"score": 1.0,
"content": "uploaded upon submission. Lastly, details on the evaluation metrics and relevant data sets are given",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
257,
253,
269
],
"spans": [
{
"bbox": [
106,
257,
253,
269
],
"score": 1.0,
"content": "in the main text as well as appendix.",
"type": "text"
}
],
"index": 12
}
],
"index": 10,
"bbox_fs": [
105,
213,
506,
269
]
},
{
"type": "title",
"bbox": [
108,
285,
175,
297
],
"lines": [
{
"bbox": [
106,
284,
176,
298
],
"spans": [
{
"bbox": [
106,
284,
176,
298
],
"score": 1.0,
"content": "REFERENCES",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
303,
504,
326
],
"lines": [
{
"bbox": [
106,
302,
506,
316
],
"spans": [
{
"bbox": [
106,
302,
506,
316
],
"score": 1.0,
"content": "Yoshua Bengio. From System 1 Deep Learning to System 2 Deep Learning. Invited talk NeurIPS,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
116,
313,
426,
327
],
"spans": [
{
"bbox": [
116,
313,
350,
327
],
"score": 1.0,
"content": "2019. URL https://www.youtube.com/watch?v",
"type": "text"
},
{
"bbox": [
351,
316,
357,
324
],
"score": 0.29,
"content": "=",
"type": "inline_equation"
},
{
"bbox": [
358,
313,
426,
327
],
"score": 1.0,
"content": "T3sxeTgT4qc.",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5,
"bbox_fs": [
106,
302,
506,
327
]
},
{
"type": "text",
"bbox": [
107,
332,
506,
366
],
"lines": [
{
"bbox": [
106,
333,
505,
345
],
"spans": [
{
"bbox": [
106,
333,
505,
345
],
"score": 1.0,
"content": "Eli Bingham, Jonathan P. Chen, Martin Jankowiak, Fritz Obermeyer, Neeraj Pradhan, Theofanis",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
115,
342,
506,
358
],
"spans": [
{
"bbox": [
115,
342,
506,
358
],
"score": 1.0,
"content": "Karaletsos, Rohit Singh, Paul A. Szerlip, Paul Horsfall, and Noah D. Goodman. Pyro: Deep",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
116,
354,
507,
367
],
"spans": [
{
"bbox": [
116,
354,
507,
367
],
"score": 1.0,
"content": "universal probabilistic programming. Journal of Machine Learning Research, pp. 28:1–28:6, 2019.",
"type": "text"
}
],
"index": 18
}
],
"index": 17,
"bbox_fs": [
106,
333,
507,
367
]
},
{
"type": "text",
"bbox": [
107,
372,
506,
406
],
"lines": [
{
"bbox": [
105,
372,
505,
386
],
"spans": [
{
"bbox": [
105,
372,
505,
386
],
"score": 1.0,
"content": "Christopher P. Burgess, Loïc Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
115,
383,
506,
397
],
"spans": [
{
"bbox": [
115,
383,
506,
397
],
"score": 1.0,
"content": "Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
116,
395,
171,
406
],
"spans": [
{
"bbox": [
116,
395,
171,
406
],
"score": 1.0,
"content": "CoRR, 2019.",
"type": "text"
}
],
"index": 21
}
],
"index": 20,
"bbox_fs": [
105,
372,
506,
406
]
},
{
"type": "text",
"bbox": [
106,
413,
506,
457
],
"lines": [
{
"bbox": [
106,
414,
506,
425
],
"spans": [
{
"bbox": [
106,
414,
506,
425
],
"score": 1.0,
"content": "Francesco Calimeri, Wolfgang Faber, Martin Gebser, Giovambattista Ianni, Roland Kaminski,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
115,
424,
506,
437
],
"spans": [
{
"bbox": [
115,
424,
506,
437
],
"score": 1.0,
"content": "Thomas Krennwallner, Nicola Leone, Marco Maratea, Francesco Ricca, and Torsten Schaub.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
435,
507,
448
],
"spans": [
{
"bbox": [
115,
435,
507,
448
],
"score": 1.0,
"content": "Asp-core-2 input language format. Theory and Practice of Logic Programming, pp. 294–309,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
114,
445,
142,
458
],
"spans": [
{
"bbox": [
114,
445,
142,
458
],
"score": 1.0,
"content": "2020.",
"type": "text"
}
],
"index": 25
}
],
"index": 23.5,
"bbox_fs": [
106,
414,
507,
458
]
},
{
"type": "text",
"bbox": [
108,
465,
504,
488
],
"lines": [
{
"bbox": [
106,
464,
506,
479
],
"spans": [
{
"bbox": [
106,
464,
506,
479
],
"score": 1.0,
"content": "YooJung Choi, Antonio Vergari, and Guy Van den Broeck. Probabilistic circuits: A unifying",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
116,
476,
465,
488
],
"spans": [
{
"bbox": [
116,
476,
465,
488
],
"score": 1.0,
"content": "framework for tractable probabilistic models. Technical report, Technical report, 2020.",
"type": "text"
}
],
"index": 27
}
],
"index": 26.5,
"bbox_fs": [
106,
464,
506,
488
]
},
{
"type": "text",
"bbox": [
106,
495,
505,
517
],
"lines": [
{
"bbox": [
106,
494,
505,
507
],
"spans": [
{
"bbox": [
106,
494,
505,
507
],
"score": 1.0,
"content": "Gabriele Ciravegna, Francesco Giannini, Marco Gori, Marco Maggini, and Stefano Melacci. Human-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
505,
417,
518
],
"spans": [
{
"bbox": [
115,
505,
417,
518
],
"score": 1.0,
"content": "driven FOL explanations of deep learning. In IJCAI, pp. 2234–2240, 2020.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
106,
494,
505,
518
]
},
{
"type": "text",
"bbox": [
106,
524,
370,
536
],
"lines": [
{
"bbox": [
106,
524,
371,
537
],
"spans": [
{
"bbox": [
106,
524,
371,
537
],
"score": 1.0,
"content": "W. F. Clocksin and Chris Mellish. Programming in Prolog. 1981.",
"type": "text"
}
],
"index": 30
}
],
"index": 30,
"bbox_fs": [
106,
524,
371,
537
]
},
{
"type": "text",
"bbox": [
109,
543,
475,
555
],
"lines": [
{
"bbox": [
106,
542,
475,
556
],
"spans": [
{
"bbox": [
106,
542,
475,
556
],
"score": 1.0,
"content": "Alain Colmerauer and Philippe Roussel. The birth of prolog. In HOPL-II, pp. 37–52, 1993.",
"type": "text"
}
],
"index": 31
}
],
"index": 31,
"bbox_fs": [
106,
542,
475,
556
]
},
{
"type": "text",
"bbox": [
108,
561,
506,
584
],
"lines": [
{
"bbox": [
105,
560,
506,
575
],
"spans": [
{
"bbox": [
105,
560,
506,
575
],
"score": 1.0,
"content": "Adnan Darwiche. SDD: A new canonical representation of propositional knowledge bases. In IJCAI,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
573,
197,
584
],
"spans": [
{
"bbox": [
115,
573,
197,
584
],
"score": 1.0,
"content": "pp. 819–826, 2011.",
"type": "text"
}
],
"index": 33
}
],
"index": 32.5,
"bbox_fs": [
105,
560,
506,
584
]
},
{
"type": "text",
"bbox": [
106,
591,
462,
603
],
"lines": [
{
"bbox": [
106,
590,
462,
604
],
"spans": [
{
"bbox": [
106,
590,
462,
604
],
"score": 1.0,
"content": "Artur d’Avila Garcez and Luís C. Lamb. Neurosymbolic AI: the 3rd wave. CoRR, 2020.",
"type": "text"
}
],
"index": 34
}
],
"index": 34,
"bbox_fs": [
106,
590,
462,
604
]
},
{
"type": "text",
"bbox": [
108,
609,
504,
632
],
"lines": [
{
"bbox": [
105,
607,
507,
624
],
"spans": [
{
"bbox": [
105,
607,
507,
624
],
"score": 1.0,
"content": "Artur S. d’Avila Garcez, Luís C. Lamb, and Dov M. Gabbay. Neural-Symbolic Cognitive Reasoning.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
116,
621,
241,
633
],
"spans": [
{
"bbox": [
116,
621,
241,
633
],
"score": 1.0,
"content": "Cognitive Technologies. 2009.",
"type": "text"
}
],
"index": 36
}
],
"index": 35.5,
"bbox_fs": [
105,
607,
507,
633
]
},
{
"type": "text",
"bbox": [
107,
639,
505,
673
],
"lines": [
{
"bbox": [
106,
639,
506,
651
],
"spans": [
{
"bbox": [
106,
639,
506,
651
],
"score": 1.0,
"content": "Artur S. d’Avila Garcez, Marco Gori, Luís C. Lamb, Luciano Serafini, Michael Spranger, and Son N.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
115,
650,
505,
663
],
"spans": [
{
"bbox": [
115,
650,
505,
663
],
"score": 1.0,
"content": "Tran. Neural-symbolic computing: An effective methodology for principled integration of machine",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
116,
662,
322,
674
],
"spans": [
{
"bbox": [
116,
662,
322,
674
],
"score": 1.0,
"content": "learning and reasoning. FLAP, pp. 611–632, 2019.",
"type": "text"
}
],
"index": 39
}
],
"index": 38,
"bbox_fs": [
106,
639,
506,
674
]
},
{
"type": "text",
"bbox": [
108,
680,
504,
703
],
"lines": [
{
"bbox": [
106,
679,
505,
693
],
"spans": [
{
"bbox": [
106,
679,
505,
693
],
"score": 1.0,
"content": "Yannis Dimopoulos, Bernhard Nebel, and Jana Koehler. Encoding planning problems in nonmono-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
115,
691,
466,
704
],
"spans": [
{
"bbox": [
115,
691,
466,
704
],
"score": 1.0,
"content": "tonic logic programs. In ECP, Lecture Notes in Computer Science, pp. 169–181, 1997.",
"type": "text"
}
],
"index": 41
}
],
"index": 40.5,
"bbox_fs": [
106,
679,
505,
704
]
},
{
"type": "text",
"bbox": [
108,
709,
503,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Martin Engelcke, Adam R. Kosiorek, Oiwi Parker Jones, and Ingmar Posner. GENESIS: generative",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
115,
721,
468,
732
],
"spans": [
{
"bbox": [
115,
721,
468,
732
],
"score": 1.0,
"content": "scene inference and sampling with object-centric latent representations. In ICLR, 2020.",
"type": "text"
}
],
"index": 43
}
],
"index": 42.5,
"bbox_fs": [
106,
709,
505,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
82,
505,
105
],
"lines": [
{
"bbox": [
105,
81,
507,
96
],
"spans": [
{
"bbox": [
105,
81,
507,
96
],
"score": 1.0,
"content": "Jerry A Fodor and Zenon W Pylyshyn. Connectionism and cognitive architecture: A critical analysis.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
116,
93,
226,
106
],
"spans": [
{
"bbox": [
116,
93,
226,
106
],
"score": 1.0,
"content": "Cognition, pp. 3–71, 1988.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
108,
112,
503,
146
],
"lines": [
{
"bbox": [
105,
111,
505,
126
],
"spans": [
{
"bbox": [
105,
111,
505,
126
],
"score": 1.0,
"content": "Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
115,
123,
505,
136
],
"spans": [
{
"bbox": [
115,
123,
505,
136
],
"score": 1.0,
"content": "Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
134,
424,
147
],
"spans": [
{
"bbox": [
115,
134,
424,
147
],
"score": 1.0,
"content": "learning with iterative variational inference. In ICML, pp. 2424–2433, 2019.",
"type": "text"
}
],
"index": 4
}
],
"index": 3
},
{
"type": "text",
"bbox": [
104,
153,
504,
176
],
"lines": [
{
"bbox": [
105,
153,
506,
167
],
"spans": [
{
"bbox": [
105,
153,
506,
167
],
"score": 1.0,
"content": "Klaus Greff, Sjoerd van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
116,
165,
239,
176
],
"spans": [
{
"bbox": [
116,
165,
239,
176
],
"score": 1.0,
"content": "neural networks. CoRR, 2020.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5
},
{
"type": "text",
"bbox": [
105,
183,
505,
207
],
"lines": [
{
"bbox": [
105,
183,
507,
197
],
"spans": [
{
"bbox": [
105,
183,
507,
197
],
"score": 1.0,
"content": "Drew A. Hudson and Christopher D. Manning. Learning by abstraction: The neural state machine.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
116,
195,
256,
207
],
"spans": [
{
"bbox": [
116,
195,
256,
207
],
"score": 1.0,
"content": "In NeurIPS, pp. 5901–5914, 2019.",
"type": "text"
}
],
"index": 8
}
],
"index": 7.5
},
{
"type": "text",
"bbox": [
104,
213,
464,
226
],
"lines": [
{
"bbox": [
105,
213,
464,
226
],
"spans": [
{
"bbox": [
105,
213,
464,
226
],
"score": 1.0,
"content": "Jindong Jiang and Sungjin Ahn. Generative neurosymbolic machines. In NeurIPS, 2020.",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
108,
232,
505,
267
],
"lines": [
{
"bbox": [
106,
233,
505,
245
],
"spans": [
{
"bbox": [
106,
233,
505,
245
],
"score": 1.0,
"content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
115,
244,
505,
257
],
"spans": [
{
"bbox": [
115,
244,
505,
257
],
"score": 1.0,
"content": "Ross B. Girshick. CLEVR: A diagnostic dataset for compositional language and elementary visual",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
115,
255,
292,
267
],
"spans": [
{
"bbox": [
115,
255,
292,
267
],
"score": 1.0,
"content": "reasoning. In CVPR, pp. 1988–1997, 2017.",
"type": "text"
}
],
"index": 12
}
],
"index": 11
},
{
"type": "text",
"bbox": [
105,
273,
505,
297
],
"lines": [
{
"bbox": [
105,
273,
506,
287
],
"spans": [
{
"bbox": [
105,
273,
506,
287
],
"score": 1.0,
"content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR 2015,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
115,
284,
143,
297
],
"spans": [
{
"bbox": [
115,
284,
143,
297
],
"score": 1.0,
"content": "2015.",
"type": "text"
}
],
"index": 14
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
106,
304,
504,
327
],
"lines": [
{
"bbox": [
106,
303,
505,
317
],
"spans": [
{
"bbox": [
106,
303,
505,
317
],
"score": 1.0,
"content": "Alexander Kuhnle and Ann A. Copestake. Shapeworld - A new test methodology for multimodal",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
316,
272,
327
],
"spans": [
{
"bbox": [
115,
316,
272,
327
],
"score": 1.0,
"content": "language understanding. CoRR, 2017.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
106,
334,
504,
357
],
"lines": [
{
"bbox": [
106,
334,
505,
347
],
"spans": [
{
"bbox": [
106,
334,
505,
347
],
"score": 1.0,
"content": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
116,
345,
304,
357
],
"spans": [
{
"bbox": [
116,
345,
304,
357
],
"score": 1.0,
"content": "document recognition. pp. 2278–2324, 1998a.",
"type": "text"
}
],
"index": 18
}
],
"index": 17.5
},
{
"type": "text",
"bbox": [
107,
364,
504,
388
],
"lines": [
{
"bbox": [
107,
364,
505,
377
],
"spans": [
{
"bbox": [
107,
364,
505,
377
],
"score": 1.0,
"content": "Yann LeCun, Corinna Cortes, and Christopher J.C. Burges. THE MNIST DATABASE of handwritten",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
116,
375,
400,
388
],
"spans": [
{
"bbox": [
116,
375,
400,
388
],
"score": 1.0,
"content": "digits, 1998b. URL http://yann.lecun.com/exdb/mnist/.",
"type": "text"
}
],
"index": 20
}
],
"index": 19.5
},
{
"type": "text",
"bbox": [
106,
394,
504,
417
],
"lines": [
{
"bbox": [
105,
393,
506,
407
],
"spans": [
{
"bbox": [
105,
393,
506,
407
],
"score": 1.0,
"content": "Joohyung Lee and Yi Wang. Weighted rules under the stable model semantics. In KR, pp. 145–154,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
114,
404,
143,
418
],
"spans": [
{
"bbox": [
114,
404,
143,
418
],
"score": 1.0,
"content": "2016.",
"type": "text"
}
],
"index": 22
}
],
"index": 21.5
},
{
"type": "text",
"bbox": [
106,
424,
505,
458
],
"lines": [
{
"bbox": [
105,
423,
505,
438
],
"spans": [
{
"bbox": [
105,
423,
505,
438
],
"score": 1.0,
"content": "Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
436,
505,
448
],
"spans": [
{
"bbox": [
115,
436,
505,
448
],
"score": 1.0,
"content": "Jiang, and Sungjin Ahn. SPACE: unsupervised object-oriented scene representation via spatial",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
116,
447,
297,
458
],
"spans": [
{
"bbox": [
116,
447,
297,
458
],
"score": 1.0,
"content": "attention and decomposition. In ICLR, 2020.",
"type": "text"
}
],
"index": 25
}
],
"index": 24
},
{
"type": "text",
"bbox": [
106,
465,
505,
499
],
"lines": [
{
"bbox": [
105,
464,
506,
479
],
"spans": [
{
"bbox": [
105,
464,
506,
479
],
"score": 1.0,
"content": "Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
115,
477,
507,
489
],
"spans": [
{
"bbox": [
115,
477,
507,
489
],
"score": 1.0,
"content": "Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
116,
488,
191,
499
],
"spans": [
{
"bbox": [
116,
488,
191,
499
],
"score": 1.0,
"content": "In NeurIPS, 2020.",
"type": "text"
}
],
"index": 28
}
],
"index": 27
},
{
"type": "text",
"bbox": [
105,
506,
505,
531
],
"lines": [
{
"bbox": [
106,
507,
506,
519
],
"spans": [
{
"bbox": [
106,
507,
506,
519
],
"score": 1.0,
"content": "Robin Manhaeve, Sebastijan Dumancic, Angelika Kimmig, Thomas Demeester, and Luc De Raedt.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
115,
517,
429,
531
],
"spans": [
{
"bbox": [
115,
517,
429,
531
],
"score": 1.0,
"content": "Deepproblog: Neural probabilistic logic programming. pp. 3753–3763, 2018.",
"type": "text"
}
],
"index": 30
}
],
"index": 29.5
},
{
"type": "text",
"bbox": [
106,
536,
505,
570
],
"lines": [
{
"bbox": [
105,
536,
506,
550
],
"spans": [
{
"bbox": [
105,
536,
506,
550
],
"score": 1.0,
"content": "Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The neuro-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
547,
505,
561
],
"spans": [
{
"bbox": [
115,
547,
505,
561
],
"score": 1.0,
"content": "symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. In",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
559,
168,
571
],
"spans": [
{
"bbox": [
115,
559,
168,
571
],
"score": 1.0,
"content": "ICLR, 2019.",
"type": "text"
}
],
"index": 33
}
],
"index": 32
},
{
"type": "text",
"bbox": [
105,
577,
482,
591
],
"lines": [
{
"bbox": [
105,
577,
481,
591
],
"spans": [
{
"bbox": [
105,
577,
481,
591
],
"score": 1.0,
"content": "Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science. 2019.",
"type": "text"
}
],
"index": 34
}
],
"index": 34
},
{
"type": "text",
"bbox": [
107,
596,
505,
631
],
"lines": [
{
"bbox": [
105,
595,
505,
611
],
"spans": [
{
"bbox": [
105,
595,
505,
611
],
"score": 1.0,
"content": "Victor W. Marek and Miroslaw Truszczynski. Stable models and an alternative logic programming",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
115,
608,
507,
622
],
"spans": [
{
"bbox": [
115,
608,
507,
622
],
"score": 1.0,
"content": "paradigm. In The Logic Programming Paradigm - A 25-Year Perspective, Artificial Intelligence,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
114,
619,
196,
631
],
"spans": [
{
"bbox": [
114,
619,
196,
631
],
"score": 1.0,
"content": "pp. 375–398. 1999.",
"type": "text"
}
],
"index": 37
}
],
"index": 36
},
{
"type": "text",
"bbox": [
107,
638,
505,
672
],
"lines": [
{
"bbox": [
104,
636,
506,
653
],
"spans": [
{
"bbox": [
104,
636,
506,
653
],
"score": 1.0,
"content": "Robert Peharz, Steven Lang, Antonio Vergari, Karl Stelzner, Alejandro Molina, Martin Trapp,",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
116,
649,
505,
662
],
"spans": [
{
"bbox": [
116,
649,
505,
662
],
"score": 1.0,
"content": "Guy Van den Broeck, Kristian Kersting, and Zoubin Ghahramani. Einsum networks: Fast and",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
116,
660,
451,
673
],
"spans": [
{
"bbox": [
116,
660,
451,
673
],
"score": 1.0,
"content": "scalable learning of tractable probabilistic circuits. In ICML, pp. 7563–7574, 2020.",
"type": "text"
}
],
"index": 40
}
],
"index": 39
},
{
"type": "text",
"bbox": [
105,
679,
505,
702
],
"lines": [
{
"bbox": [
105,
679,
506,
692
],
"spans": [
{
"bbox": [
105,
679,
506,
692
],
"score": 1.0,
"content": "Hoifung Poon and Pedro M. Domingos. Sum-product networks: A new deep architecture. In UAI,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
115,
690,
196,
702
],
"spans": [
{
"bbox": [
115,
690,
196,
702
],
"score": 1.0,
"content": "pp. 337–346, 2011.",
"type": "text"
}
],
"index": 42
}
],
"index": 41.5
},
{
"type": "text",
"bbox": [
106,
709,
504,
732
],
"lines": [
{
"bbox": [
104,
706,
506,
724
],
"spans": [
{
"bbox": [
104,
706,
506,
724
],
"score": 1.0,
"content": "Matthew Richardson and Pedro M. Domingos. Markov logic networks. Machine Learning, pp.",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
116,
720,
180,
732
],
"spans": [
{
"bbox": [
116,
720,
180,
732
],
"score": 1.0,
"content": "107–136, 2006.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5
}
],
"page_idx": 10,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
765
],
"spans": [
{
"bbox": [
299,
750,
312,
765
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
105,
82,
505,
105
],
"lines": [
{
"bbox": [
105,
81,
507,
96
],
"spans": [
{
"bbox": [
105,
81,
507,
96
],
"score": 1.0,
"content": "Jerry A Fodor and Zenon W Pylyshyn. Connectionism and cognitive architecture: A critical analysis.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
116,
93,
226,
106
],
"spans": [
{
"bbox": [
116,
93,
226,
106
],
"score": 1.0,
"content": "Cognition, pp. 3–71, 1988.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
105,
81,
507,
106
]
},
{
"type": "text",
"bbox": [
108,
112,
503,
146
],
"lines": [
{
"bbox": [
105,
111,
505,
126
],
"spans": [
{
"bbox": [
105,
111,
505,
126
],
"score": 1.0,
"content": "Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
115,
123,
505,
136
],
"spans": [
{
"bbox": [
115,
123,
505,
136
],
"score": 1.0,
"content": "Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
134,
424,
147
],
"spans": [
{
"bbox": [
115,
134,
424,
147
],
"score": 1.0,
"content": "learning with iterative variational inference. In ICML, pp. 2424–2433, 2019.",
"type": "text"
}
],
"index": 4
}
],
"index": 3,
"bbox_fs": [
105,
111,
505,
147
]
},
{
"type": "text",
"bbox": [
104,
153,
504,
176
],
"lines": [
{
"bbox": [
105,
153,
506,
167
],
"spans": [
{
"bbox": [
105,
153,
506,
167
],
"score": 1.0,
"content": "Klaus Greff, Sjoerd van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
116,
165,
239,
176
],
"spans": [
{
"bbox": [
116,
165,
239,
176
],
"score": 1.0,
"content": "neural networks. CoRR, 2020.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5,
"bbox_fs": [
105,
153,
506,
176
]
},
{
"type": "text",
"bbox": [
105,
183,
505,
207
],
"lines": [
{
"bbox": [
105,
183,
507,
197
],
"spans": [
{
"bbox": [
105,
183,
507,
197
],
"score": 1.0,
"content": "Drew A. Hudson and Christopher D. Manning. Learning by abstraction: The neural state machine.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
116,
195,
256,
207
],
"spans": [
{
"bbox": [
116,
195,
256,
207
],
"score": 1.0,
"content": "In NeurIPS, pp. 5901–5914, 2019.",
"type": "text"
}
],
"index": 8
}
],
"index": 7.5,
"bbox_fs": [
105,
183,
507,
207
]
},
{
"type": "text",
"bbox": [
104,
213,
464,
226
],
"lines": [
{
"bbox": [
105,
213,
464,
226
],
"spans": [
{
"bbox": [
105,
213,
464,
226
],
"score": 1.0,
"content": "Jindong Jiang and Sungjin Ahn. Generative neurosymbolic machines. In NeurIPS, 2020.",
"type": "text"
}
],
"index": 9
}
],
"index": 9,
"bbox_fs": [
105,
213,
464,
226
]
},
{
"type": "text",
"bbox": [
108,
232,
505,
267
],
"lines": [
{
"bbox": [
106,
233,
505,
245
],
"spans": [
{
"bbox": [
106,
233,
505,
245
],
"score": 1.0,
"content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
115,
244,
505,
257
],
"spans": [
{
"bbox": [
115,
244,
505,
257
],
"score": 1.0,
"content": "Ross B. Girshick. CLEVR: A diagnostic dataset for compositional language and elementary visual",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
115,
255,
292,
267
],
"spans": [
{
"bbox": [
115,
255,
292,
267
],
"score": 1.0,
"content": "reasoning. In CVPR, pp. 1988–1997, 2017.",
"type": "text"
}
],
"index": 12
}
],
"index": 11,
"bbox_fs": [
106,
233,
505,
267
]
},
{
"type": "text",
"bbox": [
105,
273,
505,
297
],
"lines": [
{
"bbox": [
105,
273,
506,
287
],
"spans": [
{
"bbox": [
105,
273,
506,
287
],
"score": 1.0,
"content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR 2015,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
115,
284,
143,
297
],
"spans": [
{
"bbox": [
115,
284,
143,
297
],
"score": 1.0,
"content": "2015.",
"type": "text"
}
],
"index": 14
}
],
"index": 13.5,
"bbox_fs": [
105,
273,
506,
297
]
},
{
"type": "text",
"bbox": [
106,
304,
504,
327
],
"lines": [
{
"bbox": [
106,
303,
505,
317
],
"spans": [
{
"bbox": [
106,
303,
505,
317
],
"score": 1.0,
"content": "Alexander Kuhnle and Ann A. Copestake. Shapeworld - A new test methodology for multimodal",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
316,
272,
327
],
"spans": [
{
"bbox": [
115,
316,
272,
327
],
"score": 1.0,
"content": "language understanding. CoRR, 2017.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5,
"bbox_fs": [
106,
303,
505,
327
]
},
{
"type": "text",
"bbox": [
106,
334,
504,
357
],
"lines": [
{
"bbox": [
106,
334,
505,
347
],
"spans": [
{
"bbox": [
106,
334,
505,
347
],
"score": 1.0,
"content": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
116,
345,
304,
357
],
"spans": [
{
"bbox": [
116,
345,
304,
357
],
"score": 1.0,
"content": "document recognition. pp. 2278–2324, 1998a.",
"type": "text"
}
],
"index": 18
}
],
"index": 17.5,
"bbox_fs": [
106,
334,
505,
357
]
},
{
"type": "text",
"bbox": [
107,
364,
504,
388
],
"lines": [
{
"bbox": [
107,
364,
505,
377
],
"spans": [
{
"bbox": [
107,
364,
505,
377
],
"score": 1.0,
"content": "Yann LeCun, Corinna Cortes, and Christopher J.C. Burges. THE MNIST DATABASE of handwritten",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
116,
375,
400,
388
],
"spans": [
{
"bbox": [
116,
375,
400,
388
],
"score": 1.0,
"content": "digits, 1998b. URL http://yann.lecun.com/exdb/mnist/.",
"type": "text"
}
],
"index": 20
}
],
"index": 19.5,
"bbox_fs": [
107,
364,
505,
388
]
},
{
"type": "text",
"bbox": [
106,
394,
504,
417
],
"lines": [
{
"bbox": [
105,
393,
506,
407
],
"spans": [
{
"bbox": [
105,
393,
506,
407
],
"score": 1.0,
"content": "Joohyung Lee and Yi Wang. Weighted rules under the stable model semantics. In KR, pp. 145–154,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
114,
404,
143,
418
],
"spans": [
{
"bbox": [
114,
404,
143,
418
],
"score": 1.0,
"content": "2016.",
"type": "text"
}
],
"index": 22
}
],
"index": 21.5,
"bbox_fs": [
105,
393,
506,
418
]
},
{
"type": "text",
"bbox": [
106,
424,
505,
458
],
"lines": [
{
"bbox": [
105,
423,
505,
438
],
"spans": [
{
"bbox": [
105,
423,
505,
438
],
"score": 1.0,
"content": "Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
436,
505,
448
],
"spans": [
{
"bbox": [
115,
436,
505,
448
],
"score": 1.0,
"content": "Jiang, and Sungjin Ahn. SPACE: unsupervised object-oriented scene representation via spatial",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
116,
447,
297,
458
],
"spans": [
{
"bbox": [
116,
447,
297,
458
],
"score": 1.0,
"content": "attention and decomposition. In ICLR, 2020.",
"type": "text"
}
],
"index": 25
}
],
"index": 24,
"bbox_fs": [
105,
423,
505,
458
]
},
{
"type": "text",
"bbox": [
106,
465,
505,
499
],
"lines": [
{
"bbox": [
105,
464,
506,
479
],
"spans": [
{
"bbox": [
105,
464,
506,
479
],
"score": 1.0,
"content": "Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
115,
477,
507,
489
],
"spans": [
{
"bbox": [
115,
477,
507,
489
],
"score": 1.0,
"content": "Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
116,
488,
191,
499
],
"spans": [
{
"bbox": [
116,
488,
191,
499
],
"score": 1.0,
"content": "In NeurIPS, 2020.",
"type": "text"
}
],
"index": 28
}
],
"index": 27,
"bbox_fs": [
105,
464,
507,
499
]
},
{
"type": "text",
"bbox": [
105,
506,
505,
531
],
"lines": [
{
"bbox": [
106,
507,
506,
519
],
"spans": [
{
"bbox": [
106,
507,
506,
519
],
"score": 1.0,
"content": "Robin Manhaeve, Sebastijan Dumancic, Angelika Kimmig, Thomas Demeester, and Luc De Raedt.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
115,
517,
429,
531
],
"spans": [
{
"bbox": [
115,
517,
429,
531
],
"score": 1.0,
"content": "Deepproblog: Neural probabilistic logic programming. pp. 3753–3763, 2018.",
"type": "text"
}
],
"index": 30
}
],
"index": 29.5,
"bbox_fs": [
106,
507,
506,
531
]
},
{
"type": "text",
"bbox": [
106,
536,
505,
570
],
"lines": [
{
"bbox": [
105,
536,
506,
550
],
"spans": [
{
"bbox": [
105,
536,
506,
550
],
"score": 1.0,
"content": "Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The neuro-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
547,
505,
561
],
"spans": [
{
"bbox": [
115,
547,
505,
561
],
"score": 1.0,
"content": "symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. In",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
559,
168,
571
],
"spans": [
{
"bbox": [
115,
559,
168,
571
],
"score": 1.0,
"content": "ICLR, 2019.",
"type": "text"
}
],
"index": 33
}
],
"index": 32,
"bbox_fs": [
105,
536,
506,
571
]
},
{
"type": "text",
"bbox": [
105,
577,
482,
591
],
"lines": [
{
"bbox": [
105,
577,
481,
591
],
"spans": [
{
"bbox": [
105,
577,
481,
591
],
"score": 1.0,
"content": "Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science. 2019.",
"type": "text"
}
],
"index": 34
}
],
"index": 34,
"bbox_fs": [
105,
577,
481,
591
]
},
{
"type": "text",
"bbox": [
107,
596,
505,
631
],
"lines": [
{
"bbox": [
105,
595,
505,
611
],
"spans": [
{
"bbox": [
105,
595,
505,
611
],
"score": 1.0,
"content": "Victor W. Marek and Miroslaw Truszczynski. Stable models and an alternative logic programming",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
115,
608,
507,
622
],
"spans": [
{
"bbox": [
115,
608,
507,
622
],
"score": 1.0,
"content": "paradigm. In The Logic Programming Paradigm - A 25-Year Perspective, Artificial Intelligence,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
114,
619,
196,
631
],
"spans": [
{
"bbox": [
114,
619,
196,
631
],
"score": 1.0,
"content": "pp. 375–398. 1999.",
"type": "text"
}
],
"index": 37
}
],
"index": 36,
"bbox_fs": [
105,
595,
507,
631
]
},
{
"type": "text",
"bbox": [
107,
638,
505,
672
],
"lines": [
{
"bbox": [
104,
636,
506,
653
],
"spans": [
{
"bbox": [
104,
636,
506,
653
],
"score": 1.0,
"content": "Robert Peharz, Steven Lang, Antonio Vergari, Karl Stelzner, Alejandro Molina, Martin Trapp,",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
116,
649,
505,
662
],
"spans": [
{
"bbox": [
116,
649,
505,
662
],
"score": 1.0,
"content": "Guy Van den Broeck, Kristian Kersting, and Zoubin Ghahramani. Einsum networks: Fast and",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
116,
660,
451,
673
],
"spans": [
{
"bbox": [
116,
660,
451,
673
],
"score": 1.0,
"content": "scalable learning of tractable probabilistic circuits. In ICML, pp. 7563–7574, 2020.",
"type": "text"
}
],
"index": 40
}
],
"index": 39,
"bbox_fs": [
104,
636,
506,
673
]
},
{
"type": "text",
"bbox": [
105,
679,
505,
702
],
"lines": [
{
"bbox": [
105,
679,
506,
692
],
"spans": [
{
"bbox": [
105,
679,
506,
692
],
"score": 1.0,
"content": "Hoifung Poon and Pedro M. Domingos. Sum-product networks: A new deep architecture. In UAI,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
115,
690,
196,
702
],
"spans": [
{
"bbox": [
115,
690,
196,
702
],
"score": 1.0,
"content": "pp. 337–346, 2011.",
"type": "text"
}
],
"index": 42
}
],
"index": 41.5,
"bbox_fs": [
105,
679,
506,
702
]
},
{
"type": "text",
"bbox": [
106,
709,
504,
732
],
"lines": [
{
"bbox": [
104,
706,
506,
724
],
"spans": [
{
"bbox": [
104,
706,
506,
724
],
"score": 1.0,
"content": "Matthew Richardson and Pedro M. Domingos. Markov logic networks. Machine Learning, pp.",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
116,
720,
180,
732
],
"spans": [
{
"bbox": [
116,
720,
180,
732
],
"score": 1.0,
"content": "107–136, 2006.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5,
"bbox_fs": [
104,
706,
506,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
104,
82,
505,
105
],
"lines": [
{
"bbox": [
105,
81,
506,
96
],
"spans": [
{
"bbox": [
105,
81,
506,
96
],
"score": 1.0,
"content": "Arseny Skryagin, Karl Stelzner, Alejandro Molina, Fabrizio Ventola, and Kristian Kersting. Splog:",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
116,
93,
276,
106
],
"spans": [
{
"bbox": [
116,
93,
276,
106
],
"score": 1.0,
"content": "Sum-product logic. In ProbProg, 2020.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
106,
112,
504,
135
],
"lines": [
{
"bbox": [
106,
111,
505,
125
],
"spans": [
{
"bbox": [
106,
111,
505,
125
],
"score": 1.0,
"content": "Timo Soininen and Ilkka Niemelä. Developing a declarative rule language for applications in product",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
116,
123,
442,
136
],
"spans": [
{
"bbox": [
116,
123,
442,
136
],
"score": 1.0,
"content": "configuration. In PADL, Lecture Notes in Computer Science, pp. 305–319, 1999.",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
106,
142,
504,
165
],
"lines": [
{
"bbox": [
106,
140,
506,
156
],
"spans": [
{
"bbox": [
106,
140,
506,
156
],
"score": 1.0,
"content": "Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for the right concept: Revising",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
115,
153,
505,
166
],
"spans": [
{
"bbox": [
115,
153,
505,
166
],
"score": 1.0,
"content": "neuro-symbolic concepts by interacting with their explanations. In CVPR, pp. 3619–3629, 2021.",
"type": "text"
}
],
"index": 5
}
],
"index": 4.5
},
{
"type": "text",
"bbox": [
106,
172,
505,
195
],
"lines": [
{
"bbox": [
105,
171,
506,
185
],
"spans": [
{
"bbox": [
105,
171,
506,
185
],
"score": 1.0,
"content": "Dustin Tran, Matthew D. Hoffman, Rif A. Saurous, Eugene Brevdo, Kevin Murphy, and David M.",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
182,
337,
195
],
"spans": [
{
"bbox": [
115,
182,
337,
195
],
"score": 1.0,
"content": "Blei. Deep probabilistic programming. In ICLR, 2017.",
"type": "text"
}
],
"index": 7
}
],
"index": 6.5
},
{
"type": "text",
"bbox": [
104,
201,
505,
225
],
"lines": [
{
"bbox": [
106,
201,
506,
214
],
"spans": [
{
"bbox": [
106,
201,
506,
214
],
"score": 1.0,
"content": "Zhun Yang, Adam Ishay, and Joohyung Lee. NeurASP: Embracing neural networks into answer set",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
213,
306,
225
],
"spans": [
{
"bbox": [
115,
213,
306,
225
],
"score": 1.0,
"content": "programming. In IJCAI, pp. 1755–1762, 2020.",
"type": "text"
}
],
"index": 9
}
],
"index": 8.5
},
{
"type": "text",
"bbox": [
106,
231,
505,
265
],
"lines": [
{
"bbox": [
107,
232,
505,
243
],
"spans": [
{
"bbox": [
107,
232,
505,
243
],
"score": 1.0,
"content": "Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
115,
242,
507,
257
],
"spans": [
{
"bbox": [
115,
242,
507,
257
],
"score": 1.0,
"content": "symbolic VQA: disentangling reasoning from vision and language understanding. In NeurIPS, pp.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
116,
254,
191,
265
],
"spans": [
{
"bbox": [
116,
254,
191,
265
],
"score": 1.0,
"content": "1039–1050, 2018.",
"type": "text"
}
],
"index": 12
}
],
"index": 11
},
{
"type": "text",
"bbox": [
106,
272,
504,
295
],
"lines": [
{
"bbox": [
106,
272,
506,
285
],
"spans": [
{
"bbox": [
106,
272,
506,
285
],
"score": 1.0,
"content": "Yan Zhang, Jonathon S. Hare, and Adam Prügel-Bennett. Deep set prediction networks. In NeurIPS,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
115,
284,
207,
295
],
"spans": [
{
"bbox": [
115,
284,
207,
295
],
"score": 1.0,
"content": "pp. 3207–3217, 2019.",
"type": "text"
}
],
"index": 14
}
],
"index": 13.5
}
],
"page_idx": 11,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"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": "12",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
104,
82,
505,
105
],
"lines": [
{
"bbox": [
105,
81,
506,
96
],
"spans": [
{
"bbox": [
105,
81,
506,
96
],
"score": 1.0,
"content": "Arseny Skryagin, Karl Stelzner, Alejandro Molina, Fabrizio Ventola, and Kristian Kersting. Splog:",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
116,
93,
276,
106
],
"spans": [
{
"bbox": [
116,
93,
276,
106
],
"score": 1.0,
"content": "Sum-product logic. In ProbProg, 2020.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
105,
81,
506,
106
]
},
{
"type": "text",
"bbox": [
106,
112,
504,
135
],
"lines": [
{
"bbox": [
106,
111,
505,
125
],
"spans": [
{
"bbox": [
106,
111,
505,
125
],
"score": 1.0,
"content": "Timo Soininen and Ilkka Niemelä. Developing a declarative rule language for applications in product",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
116,
123,
442,
136
],
"spans": [
{
"bbox": [
116,
123,
442,
136
],
"score": 1.0,
"content": "configuration. In PADL, Lecture Notes in Computer Science, pp. 305–319, 1999.",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5,
"bbox_fs": [
106,
111,
505,
136
]
},
{
"type": "text",
"bbox": [
106,
142,
504,
165
],
"lines": [
{
"bbox": [
106,
140,
506,
156
],
"spans": [
{
"bbox": [
106,
140,
506,
156
],
"score": 1.0,
"content": "Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for the right concept: Revising",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
115,
153,
505,
166
],
"spans": [
{
"bbox": [
115,
153,
505,
166
],
"score": 1.0,
"content": "neuro-symbolic concepts by interacting with their explanations. In CVPR, pp. 3619–3629, 2021.",
"type": "text"
}
],
"index": 5
}
],
"index": 4.5,
"bbox_fs": [
106,
140,
506,
166
]
},
{
"type": "text",
"bbox": [
106,
172,
505,
195
],
"lines": [
{
"bbox": [
105,
171,
506,
185
],
"spans": [
{
"bbox": [
105,
171,
506,
185
],
"score": 1.0,
"content": "Dustin Tran, Matthew D. Hoffman, Rif A. Saurous, Eugene Brevdo, Kevin Murphy, and David M.",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
182,
337,
195
],
"spans": [
{
"bbox": [
115,
182,
337,
195
],
"score": 1.0,
"content": "Blei. Deep probabilistic programming. In ICLR, 2017.",
"type": "text"
}
],
"index": 7
}
],
"index": 6.5,
"bbox_fs": [
105,
171,
506,
195
]
},
{
"type": "text",
"bbox": [
104,
201,
505,
225
],
"lines": [
{
"bbox": [
106,
201,
506,
214
],
"spans": [
{
"bbox": [
106,
201,
506,
214
],
"score": 1.0,
"content": "Zhun Yang, Adam Ishay, and Joohyung Lee. NeurASP: Embracing neural networks into answer set",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
213,
306,
225
],
"spans": [
{
"bbox": [
115,
213,
306,
225
],
"score": 1.0,
"content": "programming. In IJCAI, pp. 1755–1762, 2020.",
"type": "text"
}
],
"index": 9
}
],
"index": 8.5,
"bbox_fs": [
106,
201,
506,
225
]
},
{
"type": "text",
"bbox": [
106,
231,
505,
265
],
"lines": [
{
"bbox": [
107,
232,
505,
243
],
"spans": [
{
"bbox": [
107,
232,
505,
243
],
"score": 1.0,
"content": "Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
115,
242,
507,
257
],
"spans": [
{
"bbox": [
115,
242,
507,
257
],
"score": 1.0,
"content": "symbolic VQA: disentangling reasoning from vision and language understanding. In NeurIPS, pp.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
116,
254,
191,
265
],
"spans": [
{
"bbox": [
116,
254,
191,
265
],
"score": 1.0,
"content": "1039–1050, 2018.",
"type": "text"
}
],
"index": 12
}
],
"index": 11,
"bbox_fs": [
107,
232,
507,
265
]
},
{
"type": "text",
"bbox": [
106,
272,
504,
295
],
"lines": [
{
"bbox": [
106,
272,
506,
285
],
"spans": [
{
"bbox": [
106,
272,
506,
285
],
"score": 1.0,
"content": "Yan Zhang, Jonathon S. Hare, and Adam Prügel-Bennett. Deep set prediction networks. In NeurIPS,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
115,
284,
207,
295
],
"spans": [
{
"bbox": [
115,
284,
207,
295
],
"score": 1.0,
"content": "pp. 3207–3217, 2019.",
"type": "text"
}
],
"index": 14
}
],
"index": 13.5,
"bbox_fs": [
106,
272,
506,
295
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
81,
396,
94
],
"lines": [
{
"bbox": [
106,
81,
396,
96
],
"spans": [
{
"bbox": [
106,
81,
396,
96
],
"score": 1.0,
"content": "A APPENDIX A – DETAILS ON PARAMETER LEARNING",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
105,
506,
139
],
"lines": [
{
"bbox": [
106,
106,
506,
118
],
"spans": [
{
"bbox": [
106,
106,
506,
118
],
"score": 1.0,
"content": "In the Appendix, we want to discuss details on parameter learning in SLASH. Since we use co-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
116,
505,
129
],
"spans": [
{
"bbox": [
105,
116,
505,
129
],
"score": 1.0,
"content": "ordinate descent for training SLASH we present the derivative of each component of the loss function",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
128,
447,
140
],
"spans": [
{
"bbox": [
106,
128,
447,
140
],
"score": 1.0,
"content": "defined in equation 10 since while optimization, one component has to be kept fixed.",
"type": "text"
}
],
"index": 3
}
],
"index": 2
},
{
"type": "text",
"bbox": [
106,
144,
505,
167
],
"lines": [
{
"bbox": [
105,
143,
506,
158
],
"spans": [
{
"bbox": [
105,
143,
309,
158
],
"score": 1.0,
"content": "We start with the gradient of the NPP loss function",
"type": "text"
},
{
"bbox": [
309,
145,
336,
156
],
"score": 0.91,
"content": "L _ { N P P }",
"type": "inline_equation"
},
{
"bbox": [
337,
143,
506,
158
],
"score": 1.0,
"content": "i.e. the negative log-likelihood, defined in",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
104,
155,
152,
168
],
"spans": [
{
"bbox": [
104,
155,
152,
168
],
"score": 1.0,
"content": "equation 2",
"type": "text"
}
],
"index": 5
}
],
"index": 4.5
},
{
"type": "interline_equation",
"bbox": [
141,
169,
470,
245
],
"lines": [
{
"bbox": [
141,
169,
470,
245
],
"spans": [
{
"bbox": [
141,
169,
470,
245
],
"score": 0.94,
"content": "\\begin{array} { c } { { \\displaystyle { \\frac { \\partial } { \\partial \\xi } } L _ { N P P } = { \\frac { \\partial } { \\partial x _ { Q _ { i } } } \\cdot \\frac { \\partial x _ { Q _ { i } } } { \\partial \\xi } } L _ { N P P } = { \\frac { \\partial x _ { Q _ { i } } } { \\partial \\xi } } \\left( - { \\sum _ { i = 1 } ^ { n } } { \\frac { \\partial } { \\partial x _ { Q _ { i } } } \\log \\left( P _ { \\xi } ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) } \\right) } } \\\\ { { \\displaystyle { = \\frac { \\partial x _ { Q _ { i } } } { \\partial \\xi } } \\left( - { \\sum _ { i = 1 } ^ { n } } { \\frac { 1 } { \\left( P _ { \\xi } ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) } } \\frac { \\partial } { \\partial x _ { Q _ { i } } } \\left( P _ { \\xi } ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) \\right) } } \\end{array}",
"type": "interline_equation",
"image_path": "585d7e4126426f14bee1247b8486a079b7133b5af781821d1e81c73d2b32d03d.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
141,
169,
470,
194.33333333333334
],
"spans": [],
"index": 6
},
{
"bbox": [
141,
194.33333333333334,
470,
219.66666666666669
],
"spans": [],
"index": 7
},
{
"bbox": [
141,
219.66666666666669,
470,
245.00000000000003
],
"spans": [],
"index": 8
}
]
},
{
"type": "text",
"bbox": [
108,
246,
503,
273
],
"lines": [
{
"bbox": [
106,
244,
507,
266
],
"spans": [
{
"bbox": [
106,
248,
193,
262
],
"score": 1.0,
"content": "Here, we remark tha",
"type": "text"
},
{
"bbox": [
190,
244,
507,
266
],
"score": 1.0,
"content": "t ∂xQi∂ξ will be carried out by back-propagation and the expression after it is the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
260,
171,
276
],
"spans": [
{
"bbox": [
105,
260,
171,
276
],
"score": 1.0,
"content": "initial gradient.",
"type": "text"
}
],
"index": 10
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
107,
279,
505,
301
],
"lines": [
{
"bbox": [
106,
279,
507,
291
],
"spans": [
{
"bbox": [
106,
279,
375,
291
],
"score": 1.0,
"content": "Next, we derive the gradient of the logical entailment loss function",
"type": "text"
},
{
"bbox": [
376,
280,
403,
290
],
"score": 0.9,
"content": "\\mathit { L } _ { \\mathit { E N T } }",
"type": "inline_equation"
},
{
"bbox": [
403,
279,
507,
291
],
"score": 1.0,
"content": ", as defined in equation 7.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
290,
258,
303
],
"spans": [
{
"bbox": [
106,
290,
258,
303
],
"score": 1.0,
"content": "One estimates the gradient as follows",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5
},
{
"type": "interline_equation",
"bbox": [
175,
304,
435,
336
],
"lines": [
{
"bbox": [
175,
304,
435,
336
],
"spans": [
{
"bbox": [
175,
304,
435,
336
],
"score": 0.93,
"content": "\\frac { 1 } { n } \\frac { \\partial } { \\partial p } L _ { E N T } \\ge - \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\frac { \\partial \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) } { \\partial p } \\cdot \\log \\bigl ( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\bigr ) ,",
"type": "interline_equation",
"image_path": "25d52485cccf285e31823cec21b9741e5713aae7acecaaba5018b450f79cbcf7.jpg"
}
]
}
],
"index": 14,
"virtual_lines": [
{
"bbox": [
175,
304,
435,
314.6666666666667
],
"spans": [],
"index": 13
},
{
"bbox": [
175,
314.6666666666667,
435,
325.33333333333337
],
"spans": [],
"index": 14
},
{
"bbox": [
175,
325.33333333333337,
435,
336.00000000000006
],
"spans": [],
"index": 15
}
]
},
{
"type": "text",
"bbox": [
106,
338,
142,
350
],
"lines": [
{
"bbox": [
105,
336,
144,
352
],
"spans": [
{
"bbox": [
105,
336,
144,
352
],
"score": 1.0,
"content": "whereby",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
133,
357,
505,
468
],
"lines": [
{
"bbox": [
133,
357,
442,
371
],
"spans": [
{
"bbox": [
133,
357,
142,
371
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
358,
159,
370
],
"score": 0.89,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
159,
357,
428,
371
],
"score": 1.0,
"content": "is the set of random variables associated with the set of the queries",
"type": "text"
},
{
"bbox": [
428,
358,
438,
370
],
"score": 0.73,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
439,
357,
442,
371
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
133,
372,
505,
385
],
"spans": [
{
"bbox": [
133,
372,
142,
385
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
375,
159,
385
],
"score": 0.87,
"content": "x _ { Q _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
159,
372,
410,
385
],
"score": 1.0,
"content": "is a training sample, a realization of the set of random variables",
"type": "text"
},
{
"bbox": [
410,
372,
427,
385
],
"score": 0.89,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
428,
372,
505,
385
],
"score": 1.0,
"content": "associated with the",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
383,
225,
397
],
"spans": [
{
"bbox": [
141,
383,
208,
397
],
"score": 1.0,
"content": "particular query",
"type": "text"
},
{
"bbox": [
208,
384,
220,
395
],
"score": 0.89,
"content": "Q _ { i }",
"type": "inline_equation"
},
{
"bbox": [
221,
383,
225,
397
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
133,
396,
505,
413
],
"spans": [
{
"bbox": [
133,
396,
142,
413
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
397,
202,
411
],
"score": 0.93,
"content": "P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } )",
"type": "inline_equation"
},
{
"bbox": [
203,
396,
342,
413
],
"score": 1.0,
"content": "is the probability of the realization",
"type": "text"
},
{
"bbox": [
342,
401,
359,
411
],
"score": 0.89,
"content": "x _ { Q _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
359,
396,
505,
413
],
"score": 1.0,
"content": "estimated by the NPP modelling the",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
139,
409,
436,
422
],
"spans": [
{
"bbox": [
139,
409,
211,
422
],
"score": 1.0,
"content": "joint over the set",
"type": "text"
},
{
"bbox": [
211,
410,
228,
422
],
"score": 0.91,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
228,
409,
246,
422
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
246,
410,
255,
420
],
"score": 0.55,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
256,
409,
436,
422
],
"score": 1.0,
"content": "– the set of classes (the domain of the NPP),",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
132,
422,
505,
439
],
"spans": [
{
"bbox": [
132,
422,
142,
439
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
424,
206,
437
],
"score": 0.89,
"content": "\\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) )",
"type": "inline_equation"
},
{
"bbox": [
206,
422,
332,
439
],
"score": 1.0,
"content": "– the probability of the query",
"type": "text"
},
{
"bbox": [
332,
424,
344,
436
],
"score": 0.9,
"content": "Q _ { i }",
"type": "inline_equation"
},
{
"bbox": [
345,
422,
425,
439
],
"score": 1.0,
"content": "under the program",
"type": "text"
},
{
"bbox": [
425,
424,
447,
436
],
"score": 0.87,
"content": "\\Pi ( \\theta )",
"type": "inline_equation"
},
{
"bbox": [
447,
422,
505,
439
],
"score": 1.0,
"content": "calculated by",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
434,
336,
449
],
"spans": [
{
"bbox": [
141,
434,
336,
449
],
"score": 1.0,
"content": "SLASH (for the reference see the equation (5)),",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
132,
450,
358,
467
],
"spans": [
{
"bbox": [
132,
453,
160,
465
],
"score": 1.0,
"content": "• and",
"type": "text"
},
{
"bbox": [
158,
450,
358,
467
],
"score": 1.0,
"content": "∂ log(PΠ(θ)(Qi)) is the gradient as defined in Eq.9.",
"type": "text"
}
],
"index": 24
}
],
"index": 20.5
},
{
"type": "text",
"bbox": [
107,
475,
411,
487
],
"lines": [
{
"bbox": [
105,
473,
411,
489
],
"spans": [
{
"bbox": [
105,
473,
370,
489
],
"score": 1.0,
"content": "We begin with the definition of the cross-entropy for two vectors",
"type": "text"
},
{
"bbox": [
370,
477,
379,
487
],
"score": 0.84,
"content": "y _ { i }",
"type": "inline_equation"
},
{
"bbox": [
379,
473,
397,
489
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
397,
476,
407,
487
],
"score": 0.87,
"content": "\\hat { y } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
407,
473,
411,
489
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 25
}
],
"index": 25
},
{
"type": "interline_equation",
"bbox": [
115,
489,
495,
529
],
"lines": [
{
"bbox": [
115,
489,
495,
529
],
"spans": [
{
"bbox": [
115,
489,
495,
529
],
"score": 0.96,
"content": "H ( y _ { i } , \\hat { y } _ { i } ) : = \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log \\left( \\frac { 1 } { \\hat { y } _ { i j } } \\right) = \\sum _ { j = 1 } ^ { m } \\left( y _ { i j } \\cdot \\underbrace { \\log ( 1 ) } _ { = 0 } - y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) \\right) = - \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) .",
"type": "interline_equation",
"image_path": "14efd2278c7a27ae3caf3aa86dbfe8b6cd10cb590327c2af094e018d78f8a58a.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
115,
489,
495,
502.3333333333333
],
"spans": [],
"index": 26
},
{
"bbox": [
115,
502.3333333333333,
495,
515.6666666666666
],
"spans": [],
"index": 27
},
{
"bbox": [
115,
515.6666666666666,
495,
529.0
],
"spans": [],
"index": 28
}
]
},
{
"type": "text",
"bbox": [
106,
531,
201,
542
],
"lines": [
{
"bbox": [
106,
530,
201,
543
],
"spans": [
{
"bbox": [
106,
530,
201,
543
],
"score": 1.0,
"content": "Hereafter we substitute",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "interline_equation",
"bbox": [
192,
543,
419,
559
],
"lines": [
{
"bbox": [
192,
543,
419,
559
],
"spans": [
{
"bbox": [
192,
543,
419,
559
],
"score": 0.88,
"content": "y _ { i } = \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) \\qquad \\mathrm { ~ a n d ~ } \\qquad \\hat { y } _ { i } = P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } )",
"type": "interline_equation",
"image_path": "8d2dc436876b6a02f8a97fb08e101dcf909c86ec9964c1fd88d9161441971688.jpg"
}
]
}
],
"index": 30,
"virtual_lines": [
{
"bbox": [
192,
543,
419,
559
],
"spans": [],
"index": 30
}
]
},
{
"type": "text",
"bbox": [
106,
560,
151,
572
],
"lines": [
{
"bbox": [
105,
560,
151,
572
],
"spans": [
{
"bbox": [
105,
560,
151,
572
],
"score": 1.0,
"content": "and obtain",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "interline_equation",
"bbox": [
111,
572,
510,
606
],
"lines": [
{
"bbox": [
111,
572,
510,
606
],
"spans": [
{
"bbox": [
111,
572,
510,
606
],
"score": 0.92,
"content": "\\bar { \\cal I } ( y _ { i } , \\hat { y } _ { i } ) = \\cal H \\left( \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) , P ^ { ( X _ { \\mathbf { Q } } , { \\cal C } ) } ( x _ { Q _ { i } } ) \\right) = - \\sum _ { j = 1 } ^ { m } \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) \\cdot \\log \\Big ( P ^ { ( X _ { \\mathbf { Q } } , { \\cal C } ) } ( x _ { Q _ { i j } } ) \\Big ) .",
"type": "interline_equation",
"image_path": "2b9e3b0da6436f352dbb7d2b7302116d37bd724246a285e83381b2e506b7e554.jpg"
}
]
}
],
"index": 33,
"virtual_lines": [
{
"bbox": [
111,
572,
510,
583.3333333333334
],
"spans": [],
"index": 32
},
{
"bbox": [
111,
583.3333333333334,
510,
594.6666666666667
],
"spans": [],
"index": 33
},
{
"bbox": [
111,
594.6666666666667,
510,
606.0000000000001
],
"spans": [],
"index": 34
}
]
},
{
"type": "text",
"bbox": [
107,
614,
505,
660
],
"lines": [
{
"bbox": [
105,
614,
506,
627
],
"spans": [
{
"bbox": [
105,
614,
174,
627
],
"score": 1.0,
"content": "We remark that",
"type": "text"
},
{
"bbox": [
174,
617,
184,
625
],
"score": 0.73,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
185,
614,
506,
627
],
"score": 1.0,
"content": "represent the number of classes defined in the domain of an NPP. Now, we",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
626,
506,
639
],
"spans": [
{
"bbox": [
105,
626,
317,
639
],
"score": 1.0,
"content": "differentiate the equation (11) with the respect to",
"type": "text"
},
{
"bbox": [
317,
628,
324,
637
],
"score": 0.78,
"content": "p",
"type": "inline_equation"
},
{
"bbox": [
324,
626,
506,
639
],
"score": 1.0,
"content": "depicted as in Eq. 9 to be the label of the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
636,
506,
650
],
"spans": [
{
"bbox": [
105,
636,
201,
650
],
"score": 1.0,
"content": "probability of an atom",
"type": "text"
},
{
"bbox": [
201,
639,
227,
647
],
"score": 0.88,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
228,
636,
240,
650
],
"score": 1.0,
"content": "in",
"type": "text"
},
{
"bbox": [
240,
637,
259,
647
],
"score": 0.88,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
259,
636,
302,
650
],
"score": 1.0,
"content": ", denoting",
"type": "text"
},
{
"bbox": [
303,
637,
359,
650
],
"score": 0.91,
"content": "P _ { \\Pi ( \\pmb { \\theta } ) } ( c = v )",
"type": "inline_equation"
},
{
"bbox": [
360,
636,
506,
650
],
"score": 1.0,
"content": ". Since differentiation is linear, the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
648,
246,
661
],
"spans": [
{
"bbox": [
105,
648,
246,
661
],
"score": 1.0,
"content": "product rule is applicable directly:",
"type": "text"
}
],
"index": 38
}
],
"index": 36.5
},
{
"type": "interline_equation",
"bbox": [
160,
661,
452,
731
],
"lines": [
{
"bbox": [
160,
661,
452,
731
],
"spans": [
{
"bbox": [
160,
661,
452,
731
],
"score": 0.93,
"content": "\\begin{array} { c } { \\displaystyle \\frac { \\partial } { \\partial p } H \\left( y _ { i } , \\hat { y } _ { i } \\right) = - \\sum _ { j = 1 } ^ { m } \\left[ \\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) \\right) } { \\partial p } \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) \\right. } \\\\ { \\displaystyle \\left. + \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) \\cdot \\frac { \\partial \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) } { \\partial p } \\right] . } \\end{array}",
"type": "interline_equation",
"image_path": "3c0bbe5adcca47ae64eee1a8772ec5065bb21a604d9ef8d82b7d7725f3268047.jpg"
}
]
}
],
"index": 40,
"virtual_lines": [
{
"bbox": [
160,
661,
452,
684.3333333333334
],
"spans": [],
"index": 39
},
{
"bbox": [
160,
684.3333333333334,
452,
707.6666666666667
],
"spans": [],
"index": 40
},
{
"bbox": [
160,
707.6666666666667,
452,
731.0000000000001
],
"spans": [],
"index": 41
}
]
}
],
"page_idx": 12,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
26,
308,
38
],
"lines": [
{
"bbox": [
106,
25,
308,
39
],
"spans": [
{
"bbox": [
106,
25,
308,
39
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"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": "title",
"bbox": [
107,
81,
396,
94
],
"lines": [
{
"bbox": [
106,
81,
396,
96
],
"spans": [
{
"bbox": [
106,
81,
396,
96
],
"score": 1.0,
"content": "A APPENDIX A – DETAILS ON PARAMETER LEARNING",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
105,
506,
139
],
"lines": [
{
"bbox": [
106,
106,
506,
118
],
"spans": [
{
"bbox": [
106,
106,
506,
118
],
"score": 1.0,
"content": "In the Appendix, we want to discuss details on parameter learning in SLASH. Since we use co-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
116,
505,
129
],
"spans": [
{
"bbox": [
105,
116,
505,
129
],
"score": 1.0,
"content": "ordinate descent for training SLASH we present the derivative of each component of the loss function",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
128,
447,
140
],
"spans": [
{
"bbox": [
106,
128,
447,
140
],
"score": 1.0,
"content": "defined in equation 10 since while optimization, one component has to be kept fixed.",
"type": "text"
}
],
"index": 3
}
],
"index": 2,
"bbox_fs": [
105,
106,
506,
140
]
},
{
"type": "text",
"bbox": [
106,
144,
505,
167
],
"lines": [
{
"bbox": [
105,
143,
506,
158
],
"spans": [
{
"bbox": [
105,
143,
309,
158
],
"score": 1.0,
"content": "We start with the gradient of the NPP loss function",
"type": "text"
},
{
"bbox": [
309,
145,
336,
156
],
"score": 0.91,
"content": "L _ { N P P }",
"type": "inline_equation"
},
{
"bbox": [
337,
143,
506,
158
],
"score": 1.0,
"content": "i.e. the negative log-likelihood, defined in",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
104,
155,
152,
168
],
"spans": [
{
"bbox": [
104,
155,
152,
168
],
"score": 1.0,
"content": "equation 2",
"type": "text"
}
],
"index": 5
}
],
"index": 4.5,
"bbox_fs": [
104,
143,
506,
168
]
},
{
"type": "interline_equation",
"bbox": [
141,
169,
470,
245
],
"lines": [
{
"bbox": [
141,
169,
470,
245
],
"spans": [
{
"bbox": [
141,
169,
470,
245
],
"score": 0.94,
"content": "\\begin{array} { c } { { \\displaystyle { \\frac { \\partial } { \\partial \\xi } } L _ { N P P } = { \\frac { \\partial } { \\partial x _ { Q _ { i } } } \\cdot \\frac { \\partial x _ { Q _ { i } } } { \\partial \\xi } } L _ { N P P } = { \\frac { \\partial x _ { Q _ { i } } } { \\partial \\xi } } \\left( - { \\sum _ { i = 1 } ^ { n } } { \\frac { \\partial } { \\partial x _ { Q _ { i } } } \\log \\left( P _ { \\xi } ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) } \\right) } } \\\\ { { \\displaystyle { = \\frac { \\partial x _ { Q _ { i } } } { \\partial \\xi } } \\left( - { \\sum _ { i = 1 } ^ { n } } { \\frac { 1 } { \\left( P _ { \\xi } ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) } } \\frac { \\partial } { \\partial x _ { Q _ { i } } } \\left( P _ { \\xi } ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) \\right) } } \\end{array}",
"type": "interline_equation",
"image_path": "585d7e4126426f14bee1247b8486a079b7133b5af781821d1e81c73d2b32d03d.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
141,
169,
470,
194.33333333333334
],
"spans": [],
"index": 6
},
{
"bbox": [
141,
194.33333333333334,
470,
219.66666666666669
],
"spans": [],
"index": 7
},
{
"bbox": [
141,
219.66666666666669,
470,
245.00000000000003
],
"spans": [],
"index": 8
}
]
},
{
"type": "text",
"bbox": [
108,
246,
503,
273
],
"lines": [
{
"bbox": [
106,
244,
507,
266
],
"spans": [
{
"bbox": [
106,
248,
193,
262
],
"score": 1.0,
"content": "Here, we remark tha",
"type": "text"
},
{
"bbox": [
190,
244,
507,
266
],
"score": 1.0,
"content": "t ∂xQi∂ξ will be carried out by back-propagation and the expression after it is the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
260,
171,
276
],
"spans": [
{
"bbox": [
105,
260,
171,
276
],
"score": 1.0,
"content": "initial gradient.",
"type": "text"
}
],
"index": 10
}
],
"index": 9.5,
"bbox_fs": [
105,
244,
507,
276
]
},
{
"type": "text",
"bbox": [
107,
279,
505,
301
],
"lines": [
{
"bbox": [
106,
279,
507,
291
],
"spans": [
{
"bbox": [
106,
279,
375,
291
],
"score": 1.0,
"content": "Next, we derive the gradient of the logical entailment loss function",
"type": "text"
},
{
"bbox": [
376,
280,
403,
290
],
"score": 0.9,
"content": "\\mathit { L } _ { \\mathit { E N T } }",
"type": "inline_equation"
},
{
"bbox": [
403,
279,
507,
291
],
"score": 1.0,
"content": ", as defined in equation 7.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
290,
258,
303
],
"spans": [
{
"bbox": [
106,
290,
258,
303
],
"score": 1.0,
"content": "One estimates the gradient as follows",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5,
"bbox_fs": [
106,
279,
507,
303
]
},
{
"type": "interline_equation",
"bbox": [
175,
304,
435,
336
],
"lines": [
{
"bbox": [
175,
304,
435,
336
],
"spans": [
{
"bbox": [
175,
304,
435,
336
],
"score": 0.93,
"content": "\\frac { 1 } { n } \\frac { \\partial } { \\partial p } L _ { E N T } \\ge - \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\frac { \\partial \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) } { \\partial p } \\cdot \\log \\bigl ( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\bigr ) ,",
"type": "interline_equation",
"image_path": "25d52485cccf285e31823cec21b9741e5713aae7acecaaba5018b450f79cbcf7.jpg"
}
]
}
],
"index": 14,
"virtual_lines": [
{
"bbox": [
175,
304,
435,
314.6666666666667
],
"spans": [],
"index": 13
},
{
"bbox": [
175,
314.6666666666667,
435,
325.33333333333337
],
"spans": [],
"index": 14
},
{
"bbox": [
175,
325.33333333333337,
435,
336.00000000000006
],
"spans": [],
"index": 15
}
]
},
{
"type": "text",
"bbox": [
106,
338,
142,
350
],
"lines": [
{
"bbox": [
105,
336,
144,
352
],
"spans": [
{
"bbox": [
105,
336,
144,
352
],
"score": 1.0,
"content": "whereby",
"type": "text"
}
],
"index": 16
}
],
"index": 16,
"bbox_fs": [
105,
336,
144,
352
]
},
{
"type": "list",
"bbox": [
133,
357,
505,
468
],
"lines": [
{
"bbox": [
133,
357,
442,
371
],
"spans": [
{
"bbox": [
133,
357,
142,
371
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
358,
159,
370
],
"score": 0.89,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
159,
357,
428,
371
],
"score": 1.0,
"content": "is the set of random variables associated with the set of the queries",
"type": "text"
},
{
"bbox": [
428,
358,
438,
370
],
"score": 0.73,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
439,
357,
442,
371
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 17,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
133,
372,
505,
385
],
"spans": [
{
"bbox": [
133,
372,
142,
385
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
375,
159,
385
],
"score": 0.87,
"content": "x _ { Q _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
159,
372,
410,
385
],
"score": 1.0,
"content": "is a training sample, a realization of the set of random variables",
"type": "text"
},
{
"bbox": [
410,
372,
427,
385
],
"score": 0.89,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
428,
372,
505,
385
],
"score": 1.0,
"content": "associated with the",
"type": "text"
}
],
"index": 18,
"is_list_start_line": true
},
{
"bbox": [
141,
383,
225,
397
],
"spans": [
{
"bbox": [
141,
383,
208,
397
],
"score": 1.0,
"content": "particular query",
"type": "text"
},
{
"bbox": [
208,
384,
220,
395
],
"score": 0.89,
"content": "Q _ { i }",
"type": "inline_equation"
},
{
"bbox": [
221,
383,
225,
397
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 19,
"is_list_end_line": true
},
{
"bbox": [
133,
396,
505,
413
],
"spans": [
{
"bbox": [
133,
396,
142,
413
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
397,
202,
411
],
"score": 0.93,
"content": "P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } )",
"type": "inline_equation"
},
{
"bbox": [
203,
396,
342,
413
],
"score": 1.0,
"content": "is the probability of the realization",
"type": "text"
},
{
"bbox": [
342,
401,
359,
411
],
"score": 0.89,
"content": "x _ { Q _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
359,
396,
505,
413
],
"score": 1.0,
"content": "estimated by the NPP modelling the",
"type": "text"
}
],
"index": 20,
"is_list_start_line": true
},
{
"bbox": [
139,
409,
436,
422
],
"spans": [
{
"bbox": [
139,
409,
211,
422
],
"score": 1.0,
"content": "joint over the set",
"type": "text"
},
{
"bbox": [
211,
410,
228,
422
],
"score": 0.91,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
228,
409,
246,
422
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
246,
410,
255,
420
],
"score": 0.55,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
256,
409,
436,
422
],
"score": 1.0,
"content": "– the set of classes (the domain of the NPP),",
"type": "text"
}
],
"index": 21,
"is_list_end_line": true
},
{
"bbox": [
132,
422,
505,
439
],
"spans": [
{
"bbox": [
132,
422,
142,
439
],
"score": 1.0,
"content": "•",
"type": "text"
},
{
"bbox": [
142,
424,
206,
437
],
"score": 0.89,
"content": "\\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) )",
"type": "inline_equation"
},
{
"bbox": [
206,
422,
332,
439
],
"score": 1.0,
"content": "– the probability of the query",
"type": "text"
},
{
"bbox": [
332,
424,
344,
436
],
"score": 0.9,
"content": "Q _ { i }",
"type": "inline_equation"
},
{
"bbox": [
345,
422,
425,
439
],
"score": 1.0,
"content": "under the program",
"type": "text"
},
{
"bbox": [
425,
424,
447,
436
],
"score": 0.87,
"content": "\\Pi ( \\theta )",
"type": "inline_equation"
},
{
"bbox": [
447,
422,
505,
439
],
"score": 1.0,
"content": "calculated by",
"type": "text"
}
],
"index": 22,
"is_list_start_line": true
},
{
"bbox": [
141,
434,
336,
449
],
"spans": [
{
"bbox": [
141,
434,
336,
449
],
"score": 1.0,
"content": "SLASH (for the reference see the equation (5)),",
"type": "text"
}
],
"index": 23,
"is_list_end_line": true
},
{
"bbox": [
132,
450,
358,
467
],
"spans": [
{
"bbox": [
132,
453,
160,
465
],
"score": 1.0,
"content": "• and",
"type": "text"
},
{
"bbox": [
158,
450,
358,
467
],
"score": 1.0,
"content": "∂ log(PΠ(θ)(Qi)) is the gradient as defined in Eq.9.",
"type": "text"
}
],
"index": 24,
"is_list_start_line": true,
"is_list_end_line": true
}
],
"index": 20.5,
"bbox_fs": [
132,
357,
505,
467
]
},
{
"type": "text",
"bbox": [
107,
475,
411,
487
],
"lines": [
{
"bbox": [
105,
473,
411,
489
],
"spans": [
{
"bbox": [
105,
473,
370,
489
],
"score": 1.0,
"content": "We begin with the definition of the cross-entropy for two vectors",
"type": "text"
},
{
"bbox": [
370,
477,
379,
487
],
"score": 0.84,
"content": "y _ { i }",
"type": "inline_equation"
},
{
"bbox": [
379,
473,
397,
489
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
397,
476,
407,
487
],
"score": 0.87,
"content": "\\hat { y } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
407,
473,
411,
489
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 25
}
],
"index": 25,
"bbox_fs": [
105,
473,
411,
489
]
},
{
"type": "interline_equation",
"bbox": [
115,
489,
495,
529
],
"lines": [
{
"bbox": [
115,
489,
495,
529
],
"spans": [
{
"bbox": [
115,
489,
495,
529
],
"score": 0.96,
"content": "H ( y _ { i } , \\hat { y } _ { i } ) : = \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log \\left( \\frac { 1 } { \\hat { y } _ { i j } } \\right) = \\sum _ { j = 1 } ^ { m } \\left( y _ { i j } \\cdot \\underbrace { \\log ( 1 ) } _ { = 0 } - y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) \\right) = - \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) .",
"type": "interline_equation",
"image_path": "14efd2278c7a27ae3caf3aa86dbfe8b6cd10cb590327c2af094e018d78f8a58a.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
115,
489,
495,
502.3333333333333
],
"spans": [],
"index": 26
},
{
"bbox": [
115,
502.3333333333333,
495,
515.6666666666666
],
"spans": [],
"index": 27
},
{
"bbox": [
115,
515.6666666666666,
495,
529.0
],
"spans": [],
"index": 28
}
]
},
{
"type": "text",
"bbox": [
106,
531,
201,
542
],
"lines": [
{
"bbox": [
106,
530,
201,
543
],
"spans": [
{
"bbox": [
106,
530,
201,
543
],
"score": 1.0,
"content": "Hereafter we substitute",
"type": "text"
}
],
"index": 29
}
],
"index": 29,
"bbox_fs": [
106,
530,
201,
543
]
},
{
"type": "interline_equation",
"bbox": [
192,
543,
419,
559
],
"lines": [
{
"bbox": [
192,
543,
419,
559
],
"spans": [
{
"bbox": [
192,
543,
419,
559
],
"score": 0.88,
"content": "y _ { i } = \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) \\qquad \\mathrm { ~ a n d ~ } \\qquad \\hat { y } _ { i } = P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } )",
"type": "interline_equation",
"image_path": "8d2dc436876b6a02f8a97fb08e101dcf909c86ec9964c1fd88d9161441971688.jpg"
}
]
}
],
"index": 30,
"virtual_lines": [
{
"bbox": [
192,
543,
419,
559
],
"spans": [],
"index": 30
}
]
},
{
"type": "text",
"bbox": [
106,
560,
151,
572
],
"lines": [
{
"bbox": [
105,
560,
151,
572
],
"spans": [
{
"bbox": [
105,
560,
151,
572
],
"score": 1.0,
"content": "and obtain",
"type": "text"
}
],
"index": 31
}
],
"index": 31,
"bbox_fs": [
105,
560,
151,
572
]
},
{
"type": "interline_equation",
"bbox": [
111,
572,
510,
606
],
"lines": [
{
"bbox": [
111,
572,
510,
606
],
"spans": [
{
"bbox": [
111,
572,
510,
606
],
"score": 0.92,
"content": "\\bar { \\cal I } ( y _ { i } , \\hat { y } _ { i } ) = \\cal H \\left( \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) , P ^ { ( X _ { \\mathbf { Q } } , { \\cal C } ) } ( x _ { Q _ { i } } ) \\right) = - \\sum _ { j = 1 } ^ { m } \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) \\cdot \\log \\Big ( P ^ { ( X _ { \\mathbf { Q } } , { \\cal C } ) } ( x _ { Q _ { i j } } ) \\Big ) .",
"type": "interline_equation",
"image_path": "2b9e3b0da6436f352dbb7d2b7302116d37bd724246a285e83381b2e506b7e554.jpg"
}
]
}
],
"index": 33,
"virtual_lines": [
{
"bbox": [
111,
572,
510,
583.3333333333334
],
"spans": [],
"index": 32
},
{
"bbox": [
111,
583.3333333333334,
510,
594.6666666666667
],
"spans": [],
"index": 33
},
{
"bbox": [
111,
594.6666666666667,
510,
606.0000000000001
],
"spans": [],
"index": 34
}
]
},
{
"type": "text",
"bbox": [
107,
614,
505,
660
],
"lines": [
{
"bbox": [
105,
614,
506,
627
],
"spans": [
{
"bbox": [
105,
614,
174,
627
],
"score": 1.0,
"content": "We remark that",
"type": "text"
},
{
"bbox": [
174,
617,
184,
625
],
"score": 0.73,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
185,
614,
506,
627
],
"score": 1.0,
"content": "represent the number of classes defined in the domain of an NPP. Now, we",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
626,
506,
639
],
"spans": [
{
"bbox": [
105,
626,
317,
639
],
"score": 1.0,
"content": "differentiate the equation (11) with the respect to",
"type": "text"
},
{
"bbox": [
317,
628,
324,
637
],
"score": 0.78,
"content": "p",
"type": "inline_equation"
},
{
"bbox": [
324,
626,
506,
639
],
"score": 1.0,
"content": "depicted as in Eq. 9 to be the label of the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
636,
506,
650
],
"spans": [
{
"bbox": [
105,
636,
201,
650
],
"score": 1.0,
"content": "probability of an atom",
"type": "text"
},
{
"bbox": [
201,
639,
227,
647
],
"score": 0.88,
"content": "c = v",
"type": "inline_equation"
},
{
"bbox": [
228,
636,
240,
650
],
"score": 1.0,
"content": "in",
"type": "text"
},
{
"bbox": [
240,
637,
259,
647
],
"score": 0.88,
"content": "r ^ { n p p }",
"type": "inline_equation"
},
{
"bbox": [
259,
636,
302,
650
],
"score": 1.0,
"content": ", denoting",
"type": "text"
},
{
"bbox": [
303,
637,
359,
650
],
"score": 0.91,
"content": "P _ { \\Pi ( \\pmb { \\theta } ) } ( c = v )",
"type": "inline_equation"
},
{
"bbox": [
360,
636,
506,
650
],
"score": 1.0,
"content": ". Since differentiation is linear, the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
648,
246,
661
],
"spans": [
{
"bbox": [
105,
648,
246,
661
],
"score": 1.0,
"content": "product rule is applicable directly:",
"type": "text"
}
],
"index": 38
}
],
"index": 36.5,
"bbox_fs": [
105,
614,
506,
661
]
},
{
"type": "interline_equation",
"bbox": [
160,
661,
452,
731
],
"lines": [
{
"bbox": [
160,
661,
452,
731
],
"spans": [
{
"bbox": [
160,
661,
452,
731
],
"score": 0.93,
"content": "\\begin{array} { c } { \\displaystyle \\frac { \\partial } { \\partial p } H \\left( y _ { i } , \\hat { y } _ { i } \\right) = - \\sum _ { j = 1 } ^ { m } \\left[ \\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) \\right) } { \\partial p } \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) \\right. } \\\\ { \\displaystyle \\left. + \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) \\cdot \\frac { \\partial \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) } { \\partial p } \\right] . } \\end{array}",
"type": "interline_equation",
"image_path": "3c0bbe5adcca47ae64eee1a8772ec5065bb21a604d9ef8d82b7d7725f3268047.jpg"
}
]
}
],
"index": 40,
"virtual_lines": [
{
"bbox": [
160,
661,
452,
684.3333333333334
],
"spans": [],
"index": 39
},
{
"bbox": [
160,
684.3333333333334,
452,
707.6666666666667
],
"spans": [],
"index": 40
},
{
"bbox": [
160,
707.6666666666667,
452,
731.0000000000001
],
"spans": [],
"index": 41
}
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
80,
506,
127
],
"lines": [
{
"bbox": [
104,
78,
506,
102
],
"spans": [
{
"bbox": [
104,
78,
277,
102
],
"score": 1.0,
"content": "We do not wish to consider the latter term of",
"type": "text"
},
{
"bbox": [
278,
81,
422,
100
],
"score": 0.64,
"content": "\\begin{array} { r } { \\log ( P _ { \\mathrm { { I I } } ( \\theta ) } ( Q _ { i } ) ) \\cdot \\frac { \\partial \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) } { \\partial p } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
422,
78,
506,
102
],
"score": 1.0,
"content": "because it represents",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
95,
506,
119
],
"spans": [
{
"bbox": [
106,
95,
330,
119
],
"score": 1.0,
"content": "the rescaling and to keep the first since SLASH procure",
"type": "text"
},
{
"bbox": [
331,
100,
390,
117
],
"score": 0.92,
"content": "\\frac { \\partial \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) } { \\partial p }",
"type": "inline_equation"
},
{
"bbox": [
391,
95,
506,
119
],
"score": 1.0,
"content": "following Eq. 9. To achieve",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
116,
319,
127
],
"spans": [
{
"bbox": [
106,
116,
319,
127
],
"score": 1.0,
"content": "this, we estimate equation from above downwards as",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "interline_equation",
"bbox": [
171,
131,
441,
166
],
"lines": [
{
"bbox": [
171,
131,
441,
166
],
"spans": [
{
"bbox": [
171,
131,
441,
166
],
"score": 0.93,
"content": "\\frac { \\partial } { \\partial p } H \\left( y _ { i } , \\hat { y } _ { i } \\right) \\geq - \\sum _ { j = 1 } ^ { m } \\frac { \\partial \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) } { \\partial p } \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) .",
"type": "interline_equation",
"image_path": "b2e3c45fb5858cfba6c3db1fb458509db228f8ac82adf03458a47e91d2270af1.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
171,
131,
441,
142.66666666666666
],
"spans": [],
"index": 3
},
{
"bbox": [
171,
142.66666666666666,
441,
154.33333333333331
],
"spans": [],
"index": 4
},
{
"bbox": [
171,
154.33333333333331,
441,
165.99999999999997
],
"spans": [],
"index": 5
}
]
},
{
"type": "text",
"bbox": [
105,
170,
492,
183
],
"lines": [
{
"bbox": [
106,
170,
490,
183
],
"spans": [
{
"bbox": [
106,
170,
490,
183
],
"score": 1.0,
"content": "Furthermore, let us recall that under i.i.d assumption we obtain from the definition of likelihood",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "interline_equation",
"bbox": [
247,
187,
363,
220
],
"lines": [
{
"bbox": [
247,
187,
363,
220
],
"spans": [
{
"bbox": [
247,
187,
363,
220
],
"score": 0.94,
"content": "L H ( y , \\hat { y } ) = \\prod _ { i = 1 } ^ { n } L H ( y _ { i } , \\hat { y } _ { i } ) ,",
"type": "interline_equation",
"image_path": "b1cef1055651c022e4fabcdf7453f8c1f11569815c366ea8cad9ac5be1bcf777.jpg"
}
]
}
],
"index": 7.5,
"virtual_lines": [
{
"bbox": [
247,
187,
363,
203.5
],
"spans": [],
"index": 7
},
{
"bbox": [
247,
203.5,
363,
220.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "text",
"bbox": [
106,
224,
505,
248
],
"lines": [
{
"bbox": [
105,
223,
505,
238
],
"spans": [
{
"bbox": [
105,
223,
470,
238
],
"score": 1.0,
"content": "and following the negative likelihood coupled with the knowledge that the log-likelihood of",
"type": "text"
},
{
"bbox": [
471,
227,
480,
236
],
"score": 0.84,
"content": "y _ { i }",
"type": "inline_equation"
},
{
"bbox": [
480,
223,
505,
238
],
"score": 1.0,
"content": "is the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
235,
223,
249
],
"spans": [
{
"bbox": [
105,
235,
214,
249
],
"score": 1.0,
"content": "log of a particular entry of",
"type": "text"
},
{
"bbox": [
214,
237,
223,
248
],
"score": 0.87,
"content": "\\hat { y } _ { i }",
"type": "inline_equation"
}
],
"index": 10
}
],
"index": 9.5
},
{
"type": "interline_equation",
"bbox": [
145,
252,
466,
327
],
"lines": [
{
"bbox": [
145,
252,
466,
327
],
"spans": [
{
"bbox": [
145,
252,
466,
327
],
"score": 0.94,
"content": "\\begin{array} { c } { { { \\cal L } _ { E N T } = - \\displaystyle \\log { \\cal L } H ( y , \\hat { y } ) = - \\sum _ { i = 1 } ^ { n } \\log { \\cal L } H ( y _ { i } , \\hat { y } _ { i } ) = - \\sum _ { i = 1 } ^ { n } \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) = } } \\\\ { { \\displaystyle \\sum _ { i = 1 } ^ { n } \\left[ - \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) \\right] = \\sum _ { i = 1 } ^ { n } H ( y _ { i } , \\hat { y } _ { i } ) . } } \\end{array}",
"type": "interline_equation",
"image_path": "d0e566d68e2f20b4e54fc7c2b09a43dd98b00de71bda7849e2d912e6c8af305d.jpg"
}
]
}
],
"index": 12,
"virtual_lines": [
{
"bbox": [
145,
252,
466,
277.0
],
"spans": [],
"index": 11
},
{
"bbox": [
145,
277.0,
466,
302.0
],
"spans": [],
"index": 12
},
{
"bbox": [
145,
302.0,
466,
327.0
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
105,
331,
371,
344
],
"lines": [
{
"bbox": [
105,
330,
371,
345
],
"spans": [
{
"bbox": [
105,
330,
371,
345
],
"score": 1.0,
"content": "Finally, we obtain the following estimate applying inequality (12)",
"type": "text"
}
],
"index": 14
}
],
"index": 14
},
{
"type": "interline_equation",
"bbox": [
129,
348,
481,
381
],
"lines": [
{
"bbox": [
129,
348,
481,
381
],
"spans": [
{
"bbox": [
129,
348,
481,
381
],
"score": 0.94,
"content": "\\frac { 1 } { n } \\frac { \\partial } { \\partial p } L _ { E N T } = \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\frac { \\partial } { \\partial p } H ( y _ { i } , \\hat { y } _ { i } ) \\geq - \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\frac { \\partial \\log \\left( P _ { \\mathrm { { I I } } ( \\theta ) } ( Q _ { i } ) \\right) } { \\partial p } \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right)",
"type": "interline_equation",
"image_path": "e89568be2c91580ddf980db40d0e2dcf668767f5e62acf9ad3ab41e0da02c656.jpg"
}
]
}
],
"index": 16,
"virtual_lines": [
{
"bbox": [
129,
348,
481,
359.0
],
"spans": [],
"index": 15
},
{
"bbox": [
129,
359.0,
481,
370.0
],
"spans": [],
"index": 16
},
{
"bbox": [
129,
370.0,
481,
381.0
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
106,
385,
506,
453
],
"lines": [
{
"bbox": [
105,
385,
506,
398
],
"spans": [
{
"bbox": [
105,
385,
506,
398
],
"score": 1.0,
"content": "Also, we note that the mathematical transformations listed above hold for any type of NPP and the",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
397,
505,
410
],
"spans": [
{
"bbox": [
105,
397,
505,
410
],
"score": 1.0,
"content": "task dependent queries (NN with Softmax, PC or PC jointly with NN). The only difference will be",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
104,
406,
507,
424
],
"spans": [
{
"bbox": [
104,
406,
193,
424
],
"score": 1.0,
"content": "the second term, i.e.,",
"type": "text"
},
{
"bbox": [
193,
408,
277,
421
],
"score": 0.88,
"content": "\\log ( P ^ { ( C | X _ { \\mathbf { Q } } ) } ( x _ { Q _ { i j } } ) )",
"type": "inline_equation"
},
{
"bbox": [
277,
406,
289,
424
],
"score": 1.0,
"content": "or",
"type": "text"
},
{
"bbox": [
290,
408,
373,
421
],
"score": 0.9,
"content": "\\log ( P ^ { ( \\bar { X } _ { \\mathbf { Q } } | C ) } \\bar { ( x _ { Q _ { i j } } ) } )",
"type": "inline_equation"
},
{
"bbox": [
374,
406,
507,
424
],
"score": 1.0,
"content": "depending on the NPP and task.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
419,
507,
434
],
"spans": [
{
"bbox": [
104,
419,
334,
434
],
"score": 1.0,
"content": "The NPP in a form of a single PC modeling the joint over",
"type": "text"
},
{
"bbox": [
335,
421,
351,
432
],
"score": 0.87,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
351,
419,
369,
434
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
369,
421,
378,
430
],
"score": 0.82,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
378,
419,
507,
434
],
"score": 1.0,
"content": "was depicted to be the example.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
430,
506,
444
],
"spans": [
{
"bbox": [
105,
430,
506,
444
],
"score": 1.0,
"content": "With that, the derivation of gradients for both loss functions 2 and 7 is complete, and the training is",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
442,
249,
454
],
"spans": [
{
"bbox": [
105,
442,
249,
454
],
"score": 1.0,
"content": "carried out by coordinated descent.",
"type": "text"
}
],
"index": 23
}
],
"index": 20.5
},
{
"type": "text",
"bbox": [
107,
469,
505,
504
],
"lines": [
{
"bbox": [
105,
469,
505,
483
],
"spans": [
{
"bbox": [
105,
469,
440,
483
],
"score": 1.0,
"content": "Backpropagation for joint NN and PC NPPs: If within the SLASH program,",
"type": "text"
},
{
"bbox": [
441,
470,
462,
482
],
"score": 0.83,
"content": "\\Pi ( \\pmb \\theta )",
"type": "inline_equation"
},
{
"bbox": [
463,
469,
505,
483
],
"score": 1.0,
"content": ", the NPP",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
480,
505,
493
],
"spans": [
{
"bbox": [
106,
480,
505,
493
],
"score": 1.0,
"content": "forwards the data tensor through a NN first, i.e., the NPP models a joint over the NN’s output",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
492,
270,
504
],
"spans": [
{
"bbox": [
106,
492,
270,
504
],
"score": 1.0,
"content": "variables by a PC, then we rewrite (8) to",
"type": "text"
}
],
"index": 26
}
],
"index": 25
},
{
"type": "interline_equation",
"bbox": [
176,
508,
435,
541
],
"lines": [
{
"bbox": [
176,
508,
435,
541
],
"spans": [
{
"bbox": [
176,
508,
435,
541
],
"score": 0.95,
"content": "\\sum _ { i = 1 } ^ { n } { \\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) \\right) } { \\partial \\theta } } = \\sum _ { i = 1 } ^ { n } { \\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) \\right) } { \\partial \\mathbf { p } } } \\times { \\frac { \\partial \\mathbf { p } } { \\partial \\theta } } \\times { \\frac { \\partial \\theta } { \\partial \\gamma } } .",
"type": "interline_equation",
"image_path": "44e2cbe01c8e9cc1b050627c51a956c1b81d85f83dda7a0471ea236f4081c019.jpg"
}
]
}
],
"index": 28,
"virtual_lines": [
{
"bbox": [
176,
508,
435,
519.0
],
"spans": [],
"index": 27
},
{
"bbox": [
176,
519.0,
435,
530.0
],
"spans": [],
"index": 28
},
{
"bbox": [
176,
530.0,
435,
541.0
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
106,
546,
505,
572
],
"lines": [
{
"bbox": [
105,
546,
505,
562
],
"spans": [
{
"bbox": [
105,
546,
145,
561
],
"score": 1.0,
"content": "Thereby,",
"type": "text"
},
{
"bbox": [
145,
550,
154,
559
],
"score": 0.82,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
154,
546,
313,
561
],
"score": 1.0,
"content": "is the set of the NN’s parameters and",
"type": "text"
},
{
"bbox": [
314,
546,
327,
562
],
"score": 0.9,
"content": "\\frac { \\partial \\pmb { \\theta } } { \\partial \\gamma }",
"type": "inline_equation"
},
{
"bbox": [
327,
546,
505,
561
],
"score": 1.0,
"content": "is computed by the backward propagation",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
560,
174,
572
],
"spans": [
{
"bbox": [
106,
560,
174,
572
],
"score": 1.0,
"content": "through the NN.",
"type": "text"
}
],
"index": 31
}
],
"index": 30.5
}
],
"page_idx": 13,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
26,
308,
38
],
"lines": [
{
"bbox": [
106,
25,
308,
38
],
"spans": [
{
"bbox": [
106,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"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": "14",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
80,
506,
127
],
"lines": [
{
"bbox": [
104,
78,
506,
102
],
"spans": [
{
"bbox": [
104,
78,
277,
102
],
"score": 1.0,
"content": "We do not wish to consider the latter term of",
"type": "text"
},
{
"bbox": [
278,
81,
422,
100
],
"score": 0.64,
"content": "\\begin{array} { r } { \\log ( P _ { \\mathrm { { I I } } ( \\theta ) } ( Q _ { i } ) ) \\cdot \\frac { \\partial \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right) } { \\partial p } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
422,
78,
506,
102
],
"score": 1.0,
"content": "because it represents",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
95,
506,
119
],
"spans": [
{
"bbox": [
106,
95,
330,
119
],
"score": 1.0,
"content": "the rescaling and to keep the first since SLASH procure",
"type": "text"
},
{
"bbox": [
331,
100,
390,
117
],
"score": 0.92,
"content": "\\frac { \\partial \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) ) } { \\partial p }",
"type": "inline_equation"
},
{
"bbox": [
391,
95,
506,
119
],
"score": 1.0,
"content": "following Eq. 9. To achieve",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
116,
319,
127
],
"spans": [
{
"bbox": [
106,
116,
319,
127
],
"score": 1.0,
"content": "this, we estimate equation from above downwards as",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
104,
78,
506,
127
]
},
{
"type": "interline_equation",
"bbox": [
171,
131,
441,
166
],
"lines": [
{
"bbox": [
171,
131,
441,
166
],
"spans": [
{
"bbox": [
171,
131,
441,
166
],
"score": 0.93,
"content": "\\frac { \\partial } { \\partial p } H \\left( y _ { i } , \\hat { y } _ { i } \\right) \\geq - \\sum _ { j = 1 } ^ { m } \\frac { \\partial \\log ( P _ { \\Pi ( \\theta ) } ( Q _ { i j } ) ) } { \\partial p } \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i j } } ) \\right) .",
"type": "interline_equation",
"image_path": "b2e3c45fb5858cfba6c3db1fb458509db228f8ac82adf03458a47e91d2270af1.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
171,
131,
441,
142.66666666666666
],
"spans": [],
"index": 3
},
{
"bbox": [
171,
142.66666666666666,
441,
154.33333333333331
],
"spans": [],
"index": 4
},
{
"bbox": [
171,
154.33333333333331,
441,
165.99999999999997
],
"spans": [],
"index": 5
}
]
},
{
"type": "text",
"bbox": [
105,
170,
492,
183
],
"lines": [
{
"bbox": [
106,
170,
490,
183
],
"spans": [
{
"bbox": [
106,
170,
490,
183
],
"score": 1.0,
"content": "Furthermore, let us recall that under i.i.d assumption we obtain from the definition of likelihood",
"type": "text"
}
],
"index": 6
}
],
"index": 6,
"bbox_fs": [
106,
170,
490,
183
]
},
{
"type": "interline_equation",
"bbox": [
247,
187,
363,
220
],
"lines": [
{
"bbox": [
247,
187,
363,
220
],
"spans": [
{
"bbox": [
247,
187,
363,
220
],
"score": 0.94,
"content": "L H ( y , \\hat { y } ) = \\prod _ { i = 1 } ^ { n } L H ( y _ { i } , \\hat { y } _ { i } ) ,",
"type": "interline_equation",
"image_path": "b1cef1055651c022e4fabcdf7453f8c1f11569815c366ea8cad9ac5be1bcf777.jpg"
}
]
}
],
"index": 7.5,
"virtual_lines": [
{
"bbox": [
247,
187,
363,
203.5
],
"spans": [],
"index": 7
},
{
"bbox": [
247,
203.5,
363,
220.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "text",
"bbox": [
106,
224,
505,
248
],
"lines": [
{
"bbox": [
105,
223,
505,
238
],
"spans": [
{
"bbox": [
105,
223,
470,
238
],
"score": 1.0,
"content": "and following the negative likelihood coupled with the knowledge that the log-likelihood of",
"type": "text"
},
{
"bbox": [
471,
227,
480,
236
],
"score": 0.84,
"content": "y _ { i }",
"type": "inline_equation"
},
{
"bbox": [
480,
223,
505,
238
],
"score": 1.0,
"content": "is the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
235,
223,
249
],
"spans": [
{
"bbox": [
105,
235,
214,
249
],
"score": 1.0,
"content": "log of a particular entry of",
"type": "text"
},
{
"bbox": [
214,
237,
223,
248
],
"score": 0.87,
"content": "\\hat { y } _ { i }",
"type": "inline_equation"
}
],
"index": 10
}
],
"index": 9.5,
"bbox_fs": [
105,
223,
505,
249
]
},
{
"type": "interline_equation",
"bbox": [
145,
252,
466,
327
],
"lines": [
{
"bbox": [
145,
252,
466,
327
],
"spans": [
{
"bbox": [
145,
252,
466,
327
],
"score": 0.94,
"content": "\\begin{array} { c } { { { \\cal L } _ { E N T } = - \\displaystyle \\log { \\cal L } H ( y , \\hat { y } ) = - \\sum _ { i = 1 } ^ { n } \\log { \\cal L } H ( y _ { i } , \\hat { y } _ { i } ) = - \\sum _ { i = 1 } ^ { n } \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) = } } \\\\ { { \\displaystyle \\sum _ { i = 1 } ^ { n } \\left[ - \\sum _ { j = 1 } ^ { m } y _ { i j } \\cdot \\log ( \\hat { y } _ { i j } ) \\right] = \\sum _ { i = 1 } ^ { n } H ( y _ { i } , \\hat { y } _ { i } ) . } } \\end{array}",
"type": "interline_equation",
"image_path": "d0e566d68e2f20b4e54fc7c2b09a43dd98b00de71bda7849e2d912e6c8af305d.jpg"
}
]
}
],
"index": 12,
"virtual_lines": [
{
"bbox": [
145,
252,
466,
277.0
],
"spans": [],
"index": 11
},
{
"bbox": [
145,
277.0,
466,
302.0
],
"spans": [],
"index": 12
},
{
"bbox": [
145,
302.0,
466,
327.0
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
105,
331,
371,
344
],
"lines": [
{
"bbox": [
105,
330,
371,
345
],
"spans": [
{
"bbox": [
105,
330,
371,
345
],
"score": 1.0,
"content": "Finally, we obtain the following estimate applying inequality (12)",
"type": "text"
}
],
"index": 14
}
],
"index": 14,
"bbox_fs": [
105,
330,
371,
345
]
},
{
"type": "interline_equation",
"bbox": [
129,
348,
481,
381
],
"lines": [
{
"bbox": [
129,
348,
481,
381
],
"spans": [
{
"bbox": [
129,
348,
481,
381
],
"score": 0.94,
"content": "\\frac { 1 } { n } \\frac { \\partial } { \\partial p } L _ { E N T } = \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\frac { \\partial } { \\partial p } H ( y _ { i } , \\hat { y } _ { i } ) \\geq - \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\frac { \\partial \\log \\left( P _ { \\mathrm { { I I } } ( \\theta ) } ( Q _ { i } ) \\right) } { \\partial p } \\cdot \\log \\left( P ^ { ( X _ { \\mathbf { Q } } , C ) } ( x _ { Q _ { i } } ) \\right)",
"type": "interline_equation",
"image_path": "e89568be2c91580ddf980db40d0e2dcf668767f5e62acf9ad3ab41e0da02c656.jpg"
}
]
}
],
"index": 16,
"virtual_lines": [
{
"bbox": [
129,
348,
481,
359.0
],
"spans": [],
"index": 15
},
{
"bbox": [
129,
359.0,
481,
370.0
],
"spans": [],
"index": 16
},
{
"bbox": [
129,
370.0,
481,
381.0
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
106,
385,
506,
453
],
"lines": [
{
"bbox": [
105,
385,
506,
398
],
"spans": [
{
"bbox": [
105,
385,
506,
398
],
"score": 1.0,
"content": "Also, we note that the mathematical transformations listed above hold for any type of NPP and the",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
397,
505,
410
],
"spans": [
{
"bbox": [
105,
397,
505,
410
],
"score": 1.0,
"content": "task dependent queries (NN with Softmax, PC or PC jointly with NN). The only difference will be",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
104,
406,
507,
424
],
"spans": [
{
"bbox": [
104,
406,
193,
424
],
"score": 1.0,
"content": "the second term, i.e.,",
"type": "text"
},
{
"bbox": [
193,
408,
277,
421
],
"score": 0.88,
"content": "\\log ( P ^ { ( C | X _ { \\mathbf { Q } } ) } ( x _ { Q _ { i j } } ) )",
"type": "inline_equation"
},
{
"bbox": [
277,
406,
289,
424
],
"score": 1.0,
"content": "or",
"type": "text"
},
{
"bbox": [
290,
408,
373,
421
],
"score": 0.9,
"content": "\\log ( P ^ { ( \\bar { X } _ { \\mathbf { Q } } | C ) } \\bar { ( x _ { Q _ { i j } } ) } )",
"type": "inline_equation"
},
{
"bbox": [
374,
406,
507,
424
],
"score": 1.0,
"content": "depending on the NPP and task.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
419,
507,
434
],
"spans": [
{
"bbox": [
104,
419,
334,
434
],
"score": 1.0,
"content": "The NPP in a form of a single PC modeling the joint over",
"type": "text"
},
{
"bbox": [
335,
421,
351,
432
],
"score": 0.87,
"content": "X _ { \\mathbf { Q } }",
"type": "inline_equation"
},
{
"bbox": [
351,
419,
369,
434
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
369,
421,
378,
430
],
"score": 0.82,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
378,
419,
507,
434
],
"score": 1.0,
"content": "was depicted to be the example.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
430,
506,
444
],
"spans": [
{
"bbox": [
105,
430,
506,
444
],
"score": 1.0,
"content": "With that, the derivation of gradients for both loss functions 2 and 7 is complete, and the training is",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
442,
249,
454
],
"spans": [
{
"bbox": [
105,
442,
249,
454
],
"score": 1.0,
"content": "carried out by coordinated descent.",
"type": "text"
}
],
"index": 23
}
],
"index": 20.5,
"bbox_fs": [
104,
385,
507,
454
]
},
{
"type": "text",
"bbox": [
107,
469,
505,
504
],
"lines": [
{
"bbox": [
105,
469,
505,
483
],
"spans": [
{
"bbox": [
105,
469,
440,
483
],
"score": 1.0,
"content": "Backpropagation for joint NN and PC NPPs: If within the SLASH program,",
"type": "text"
},
{
"bbox": [
441,
470,
462,
482
],
"score": 0.83,
"content": "\\Pi ( \\pmb \\theta )",
"type": "inline_equation"
},
{
"bbox": [
463,
469,
505,
483
],
"score": 1.0,
"content": ", the NPP",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
480,
505,
493
],
"spans": [
{
"bbox": [
106,
480,
505,
493
],
"score": 1.0,
"content": "forwards the data tensor through a NN first, i.e., the NPP models a joint over the NN’s output",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
492,
270,
504
],
"spans": [
{
"bbox": [
106,
492,
270,
504
],
"score": 1.0,
"content": "variables by a PC, then we rewrite (8) to",
"type": "text"
}
],
"index": 26
}
],
"index": 25,
"bbox_fs": [
105,
469,
505,
504
]
},
{
"type": "interline_equation",
"bbox": [
176,
508,
435,
541
],
"lines": [
{
"bbox": [
176,
508,
435,
541
],
"spans": [
{
"bbox": [
176,
508,
435,
541
],
"score": 0.95,
"content": "\\sum _ { i = 1 } ^ { n } { \\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) \\right) } { \\partial \\theta } } = \\sum _ { i = 1 } ^ { n } { \\frac { \\partial \\log \\left( P _ { \\Pi ( \\theta ) } ( Q _ { i } ) \\right) } { \\partial \\mathbf { p } } } \\times { \\frac { \\partial \\mathbf { p } } { \\partial \\theta } } \\times { \\frac { \\partial \\theta } { \\partial \\gamma } } .",
"type": "interline_equation",
"image_path": "44e2cbe01c8e9cc1b050627c51a956c1b81d85f83dda7a0471ea236f4081c019.jpg"
}
]
}
],
"index": 28,
"virtual_lines": [
{
"bbox": [
176,
508,
435,
519.0
],
"spans": [],
"index": 27
},
{
"bbox": [
176,
519.0,
435,
530.0
],
"spans": [],
"index": 28
},
{
"bbox": [
176,
530.0,
435,
541.0
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
106,
546,
505,
572
],
"lines": [
{
"bbox": [
105,
546,
505,
562
],
"spans": [
{
"bbox": [
105,
546,
145,
561
],
"score": 1.0,
"content": "Thereby,",
"type": "text"
},
{
"bbox": [
145,
550,
154,
559
],
"score": 0.82,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
154,
546,
313,
561
],
"score": 1.0,
"content": "is the set of the NN’s parameters and",
"type": "text"
},
{
"bbox": [
314,
546,
327,
562
],
"score": 0.9,
"content": "\\frac { \\partial \\pmb { \\theta } } { \\partial \\gamma }",
"type": "inline_equation"
},
{
"bbox": [
327,
546,
505,
561
],
"score": 1.0,
"content": "is computed by the backward propagation",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
560,
174,
572
],
"spans": [
{
"bbox": [
106,
560,
174,
572
],
"score": 1.0,
"content": "through the NN.",
"type": "text"
}
],
"index": 31
}
],
"index": 30.5,
"bbox_fs": [
105,
546,
505,
572
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
81,
309,
94
],
"lines": [
{
"bbox": [
104,
79,
309,
96
],
"spans": [
{
"bbox": [
104,
79,
309,
96
],
"score": 1.0,
"content": "B APPENDIX B – SLASH PROGRAMS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
106,
506,
151
],
"lines": [
{
"bbox": [
105,
106,
506,
119
],
"spans": [
{
"bbox": [
105,
106,
506,
119
],
"score": 1.0,
"content": "Here, the interested reader will find the SLASH programs which we compiled for our experiments.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
117,
505,
129
],
"spans": [
{
"bbox": [
106,
117,
505,
129
],
"score": 1.0,
"content": "Figure 4 presents the one for the MNIST Addition task, Figure 6 – for the set prediction task with",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
128,
506,
140
],
"spans": [
{
"bbox": [
105,
128,
407,
140
],
"score": 1.0,
"content": "slot attention encoder and the subsequent CoGenT test. Note the use of the",
"type": "text"
},
{
"bbox": [
408,
128,
424,
138
],
"score": 0.71,
"content": "\" + \"",
"type": "inline_equation"
},
{
"bbox": [
424,
128,
506,
140
],
"score": 1.0,
"content": "and “-” notation for",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
138,
375,
152
],
"spans": [
{
"bbox": [
105,
138,
375,
152
],
"score": 1.0,
"content": "indicating whether a random variable is given or being queried for.",
"type": "text"
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
99,
167,
500,
228
],
"lines": [
{
"bbox": [
105,
166,
191,
178
],
"spans": [
{
"bbox": [
105,
166,
191,
178
],
"score": 1.0,
"content": "# Define images",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
176,
200,
189
],
"spans": [
{
"bbox": [
106,
177,
147,
188
],
"score": 1.0,
"content": "img(i1).",
"type": "text"
},
{
"bbox": [
153,
176,
200,
189
],
"score": 1.0,
"content": "img(i2).",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
98,
186,
320,
199
],
"spans": [
{
"bbox": [
98,
190,
104,
195
],
"score": 1.0,
"content": "3",
"type": "text"
},
{
"bbox": [
105,
186,
320,
199
],
"score": 1.0,
"content": "# Define Neural-Probabilistic Predicate",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
99,
195,
361,
209
],
"spans": [
{
"bbox": [
99,
200,
103,
205
],
"score": 1.0,
"content": "4",
"type": "text"
},
{
"bbox": [
105,
195,
361,
209
],
"score": 1.0,
"content": "npp(digit(X), [0,1,2,3,4,5,6,7,8,9]) :- img(X).",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
99,
207,
486,
219
],
"spans": [
{
"bbox": [
99,
210,
104,
216
],
"score": 1.0,
"content": "5",
"type": "text"
},
{
"bbox": [
107,
207,
486,
219
],
"score": 1.0,
"content": "# Define the addition of digits given two images and the resulting sum",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
98,
217,
459,
228
],
"spans": [
{
"bbox": [
98,
219,
104,
226
],
"score": 1.0,
"content": "6",
"type": "text"
},
{
"bbox": [
105,
217,
250,
228
],
"score": 1.0,
"content": "addition(A, B, N) :- digit",
"type": "text"
},
{
"bbox": [
251,
218,
264,
227
],
"score": 0.52,
"content": "\\mathrm { \\Omega } + \\tt { A }",
"type": "inline_equation"
},
{
"bbox": [
264,
217,
338,
228
],
"score": 1.0,
"content": ", -N1), digit(",
"type": "text"
},
{
"bbox": [
338,
218,
350,
227
],
"score": 0.57,
"content": "+ \\mathtt { B }",
"type": "inline_equation"
},
{
"bbox": [
351,
217,
391,
228
],
"score": 1.0,
"content": ", -N2),",
"type": "text"
},
{
"bbox": [
391,
217,
453,
227
],
"score": 0.55,
"content": "\\mathrm { ~ N ~ } = \\mathrm { ~ N ~ 1 ~ } + \\mathrm { ~ N ~ 2 ~ }",
"type": "inline_equation"
},
{
"bbox": [
453,
217,
459,
228
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 10
}
],
"index": 7.5
},
{
"type": "text",
"bbox": [
98,
289,
355,
309
],
"lines": [
{
"bbox": [
105,
288,
356,
299
],
"spans": [
{
"bbox": [
105,
288,
356,
299
],
"score": 1.0,
"content": "# Is 7 the sum of the digits in img1 and img2?",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
98,
298,
302,
310
],
"spans": [
{
"bbox": [
98,
301,
104,
308
],
"score": 1.0,
"content": "2",
"type": "text"
},
{
"bbox": [
105,
298,
302,
310
],
"score": 1.0,
"content": ":- addition(image_id1, image_id2, 7)",
"type": "text"
}
],
"index": 14
}
],
"index": 13.5
},
{
"type": "image",
"bbox": [
96,
370,
492,
602
],
"blocks": [
{
"type": "image_caption",
"bbox": [
104,
243,
505,
265
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
241,
505,
256
],
"spans": [
{
"bbox": [
105,
241,
505,
256
],
"score": 1.0,
"content": "Figure 4: SLASH Program for MNIST addition. The same program was used for the training with",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
253,
162,
266
],
"spans": [
{
"bbox": [
105,
253,
162,
266
],
"score": 1.0,
"content": "missing data.",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5
},
{
"type": "image_caption",
"bbox": [
105,
324,
504,
347
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
323,
505,
337
],
"spans": [
{
"bbox": [
105,
323,
505,
337
],
"score": 1.0,
"content": "Figure 5: Example SLASH Query for MNIST addition. The same type of query was used for the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
334,
213,
348
],
"spans": [
{
"bbox": [
106,
334,
213,
348
],
"score": 1.0,
"content": "training with missing data",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
},
{
"type": "image_body",
"bbox": [
96,
370,
492,
602
],
"group_id": 0,
"lines": [
{
"bbox": [
96,
370,
492,
602
],
"spans": [
{
"bbox": [
96,
370,
492,
602
],
"score": 0.793,
"type": "image",
"image_path": "5d942cadea3c1d2f75797d835c60baf51150bd1a4748aee77f10460fe3c0d331.jpg"
}
]
}
],
"index": 18,
"virtual_lines": [
{
"bbox": [
96,
370,
492,
447.3333333333333
],
"spans": [],
"index": 17
},
{
"bbox": [
96,
447.3333333333333,
492,
524.6666666666666
],
"spans": [],
"index": 18
},
{
"bbox": [
96,
524.6666666666666,
492,
602.0
],
"spans": [],
"index": 19
}
]
},
{
"type": "image_caption",
"bbox": [
107,
615,
504,
638
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
615,
506,
628
],
"spans": [
{
"bbox": [
106,
615,
506,
628
],
"score": 1.0,
"content": "Figure 6: SLASH Program for ShapeWorld4. The same program was used for the CoGenT experi-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
626,
136,
638
],
"spans": [
{
"bbox": [
105,
626,
136,
638
],
"score": 1.0,
"content": "ments.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
}
],
"index": 16.75
},
{
"type": "text",
"bbox": [
100,
661,
452,
682
],
"lines": [
{
"bbox": [
105,
661,
452,
671
],
"spans": [
{
"bbox": [
105,
661,
452,
671
],
"score": 1.0,
"content": "# Does object o1 have the attributes red, circle, bright, small?",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
107,
670,
371,
682
],
"spans": [
{
"bbox": [
107,
670,
371,
682
],
"score": 1.0,
"content": ":- has_attributes(o1, red, circle, bright, small)",
"type": "text"
}
],
"index": 23
}
],
"index": 22.5
},
{
"type": "text",
"bbox": [
105,
696,
505,
720
],
"lines": [
{
"bbox": [
105,
696,
506,
709
],
"spans": [
{
"bbox": [
105,
696,
506,
709
],
"score": 1.0,
"content": "Figure 7: Example SLASH Query for ShapeWorld4 experiments. In other words, this query corre-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
707,
372,
720
],
"spans": [
{
"bbox": [
105,
707,
372,
720
],
"score": 1.0,
"content": "sponds to asking SLASH: “Is object 1 a small, bright red circle?”.",
"type": "text"
}
],
"index": 25
}
],
"index": 24.5
}
],
"page_idx": 14,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
26,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
38
],
"spans": [
{
"bbox": [
106,
25,
309,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
81,
309,
94
],
"lines": [
{
"bbox": [
104,
79,
309,
96
],
"spans": [
{
"bbox": [
104,
79,
309,
96
],
"score": 1.0,
"content": "B APPENDIX B – SLASH PROGRAMS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
106,
506,
151
],
"lines": [
{
"bbox": [
105,
106,
506,
119
],
"spans": [
{
"bbox": [
105,
106,
506,
119
],
"score": 1.0,
"content": "Here, the interested reader will find the SLASH programs which we compiled for our experiments.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
117,
505,
129
],
"spans": [
{
"bbox": [
106,
117,
505,
129
],
"score": 1.0,
"content": "Figure 4 presents the one for the MNIST Addition task, Figure 6 – for the set prediction task with",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
128,
506,
140
],
"spans": [
{
"bbox": [
105,
128,
407,
140
],
"score": 1.0,
"content": "slot attention encoder and the subsequent CoGenT test. Note the use of the",
"type": "text"
},
{
"bbox": [
408,
128,
424,
138
],
"score": 0.71,
"content": "\" + \"",
"type": "inline_equation"
},
{
"bbox": [
424,
128,
506,
140
],
"score": 1.0,
"content": "and “-” notation for",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
138,
375,
152
],
"spans": [
{
"bbox": [
105,
138,
375,
152
],
"score": 1.0,
"content": "indicating whether a random variable is given or being queried for.",
"type": "text"
}
],
"index": 4
}
],
"index": 2.5,
"bbox_fs": [
105,
106,
506,
152
]
},
{
"type": "text",
"bbox": [
99,
167,
500,
228
],
"lines": [
{
"bbox": [
105,
166,
191,
178
],
"spans": [
{
"bbox": [
105,
166,
191,
178
],
"score": 1.0,
"content": "# Define images",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
176,
200,
189
],
"spans": [
{
"bbox": [
106,
177,
147,
188
],
"score": 1.0,
"content": "img(i1).",
"type": "text"
},
{
"bbox": [
153,
176,
200,
189
],
"score": 1.0,
"content": "img(i2).",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
98,
186,
320,
199
],
"spans": [
{
"bbox": [
98,
190,
104,
195
],
"score": 1.0,
"content": "3",
"type": "text"
},
{
"bbox": [
105,
186,
320,
199
],
"score": 1.0,
"content": "# Define Neural-Probabilistic Predicate",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
99,
195,
361,
209
],
"spans": [
{
"bbox": [
99,
200,
103,
205
],
"score": 1.0,
"content": "4",
"type": "text"
},
{
"bbox": [
105,
195,
361,
209
],
"score": 1.0,
"content": "npp(digit(X), [0,1,2,3,4,5,6,7,8,9]) :- img(X).",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
99,
207,
486,
219
],
"spans": [
{
"bbox": [
99,
210,
104,
216
],
"score": 1.0,
"content": "5",
"type": "text"
},
{
"bbox": [
107,
207,
486,
219
],
"score": 1.0,
"content": "# Define the addition of digits given two images and the resulting sum",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
98,
217,
459,
228
],
"spans": [
{
"bbox": [
98,
219,
104,
226
],
"score": 1.0,
"content": "6",
"type": "text"
},
{
"bbox": [
105,
217,
250,
228
],
"score": 1.0,
"content": "addition(A, B, N) :- digit",
"type": "text"
},
{
"bbox": [
251,
218,
264,
227
],
"score": 0.52,
"content": "\\mathrm { \\Omega } + \\tt { A }",
"type": "inline_equation"
},
{
"bbox": [
264,
217,
338,
228
],
"score": 1.0,
"content": ", -N1), digit(",
"type": "text"
},
{
"bbox": [
338,
218,
350,
227
],
"score": 0.57,
"content": "+ \\mathtt { B }",
"type": "inline_equation"
},
{
"bbox": [
351,
217,
391,
228
],
"score": 1.0,
"content": ", -N2),",
"type": "text"
},
{
"bbox": [
391,
217,
453,
227
],
"score": 0.55,
"content": "\\mathrm { ~ N ~ } = \\mathrm { ~ N ~ 1 ~ } + \\mathrm { ~ N ~ 2 ~ }",
"type": "inline_equation"
},
{
"bbox": [
453,
217,
459,
228
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 10
}
],
"index": 7.5,
"bbox_fs": [
98,
166,
486,
228
]
},
{
"type": "text",
"bbox": [
98,
289,
355,
309
],
"lines": [
{
"bbox": [
105,
288,
356,
299
],
"spans": [
{
"bbox": [
105,
288,
356,
299
],
"score": 1.0,
"content": "# Is 7 the sum of the digits in img1 and img2?",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
98,
298,
302,
310
],
"spans": [
{
"bbox": [
98,
301,
104,
308
],
"score": 1.0,
"content": "2",
"type": "text"
},
{
"bbox": [
105,
298,
302,
310
],
"score": 1.0,
"content": ":- addition(image_id1, image_id2, 7)",
"type": "text"
}
],
"index": 14
}
],
"index": 13.5,
"bbox_fs": [
98,
288,
356,
310
]
},
{
"type": "image",
"bbox": [
96,
370,
492,
602
],
"blocks": [
{
"type": "image_caption",
"bbox": [
104,
243,
505,
265
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
241,
505,
256
],
"spans": [
{
"bbox": [
105,
241,
505,
256
],
"score": 1.0,
"content": "Figure 4: SLASH Program for MNIST addition. The same program was used for the training with",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
253,
162,
266
],
"spans": [
{
"bbox": [
105,
253,
162,
266
],
"score": 1.0,
"content": "missing data.",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5
},
{
"type": "image_caption",
"bbox": [
105,
324,
504,
347
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
323,
505,
337
],
"spans": [
{
"bbox": [
105,
323,
505,
337
],
"score": 1.0,
"content": "Figure 5: Example SLASH Query for MNIST addition. The same type of query was used for the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
334,
213,
348
],
"spans": [
{
"bbox": [
106,
334,
213,
348
],
"score": 1.0,
"content": "training with missing data",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
},
{
"type": "image_body",
"bbox": [
96,
370,
492,
602
],
"group_id": 0,
"lines": [
{
"bbox": [
96,
370,
492,
602
],
"spans": [
{
"bbox": [
96,
370,
492,
602
],
"score": 0.793,
"type": "image",
"image_path": "5d942cadea3c1d2f75797d835c60baf51150bd1a4748aee77f10460fe3c0d331.jpg"
}
]
}
],
"index": 18,
"virtual_lines": [
{
"bbox": [
96,
370,
492,
447.3333333333333
],
"spans": [],
"index": 17
},
{
"bbox": [
96,
447.3333333333333,
492,
524.6666666666666
],
"spans": [],
"index": 18
},
{
"bbox": [
96,
524.6666666666666,
492,
602.0
],
"spans": [],
"index": 19
}
]
},
{
"type": "image_caption",
"bbox": [
107,
615,
504,
638
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
615,
506,
628
],
"spans": [
{
"bbox": [
106,
615,
506,
628
],
"score": 1.0,
"content": "Figure 6: SLASH Program for ShapeWorld4. The same program was used for the CoGenT experi-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
626,
136,
638
],
"spans": [
{
"bbox": [
105,
626,
136,
638
],
"score": 1.0,
"content": "ments.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
}
],
"index": 16.75
},
{
"type": "text",
"bbox": [
100,
661,
452,
682
],
"lines": [
{
"bbox": [
105,
661,
452,
671
],
"spans": [
{
"bbox": [
105,
661,
452,
671
],
"score": 1.0,
"content": "# Does object o1 have the attributes red, circle, bright, small?",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
107,
670,
371,
682
],
"spans": [
{
"bbox": [
107,
670,
371,
682
],
"score": 1.0,
"content": ":- has_attributes(o1, red, circle, bright, small)",
"type": "text"
}
],
"index": 23
}
],
"index": 22.5,
"bbox_fs": [
105,
661,
452,
682
]
},
{
"type": "text",
"bbox": [
105,
696,
505,
720
],
"lines": [
{
"bbox": [
105,
696,
506,
709
],
"spans": [
{
"bbox": [
105,
696,
506,
709
],
"score": 1.0,
"content": "Figure 7: Example SLASH Query for ShapeWorld4 experiments. In other words, this query corre-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
707,
372,
720
],
"spans": [
{
"bbox": [
105,
707,
372,
720
],
"score": 1.0,
"content": "sponds to asking SLASH: “Is object 1 a small, bright red circle?”.",
"type": "text"
}
],
"index": 25
}
],
"index": 24.5,
"bbox_fs": [
105,
696,
506,
720
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
81,
335,
94
],
"lines": [
{
"bbox": [
106,
79,
337,
96
],
"spans": [
{
"bbox": [
106,
79,
337,
96
],
"score": 1.0,
"content": "C APPENDIX C – EXPERIMENTAL DETAILS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "title",
"bbox": [
108,
106,
260,
118
],
"lines": [
{
"bbox": [
106,
105,
262,
119
],
"spans": [
{
"bbox": [
106,
105,
262,
119
],
"score": 1.0,
"content": "C.1 SHAPEWORLD4 GENERATION",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "text",
"bbox": [
107,
127,
505,
160
],
"lines": [
{
"bbox": [
105,
126,
505,
141
],
"spans": [
{
"bbox": [
105,
126,
505,
141
],
"score": 1.0,
"content": "The ShapeWorld4 and ShapeWorld4 CoGenT data sets were generated using the original scripts",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
138,
505,
151
],
"spans": [
{
"bbox": [
105,
138,
505,
151
],
"score": 1.0,
"content": "of (Kuhnle & Copestake, 2017) (https://github.com/AlexKuhnle/ShapeWorld). The",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
150,
371,
162
],
"spans": [
{
"bbox": [
105,
150,
371,
162
],
"score": 1.0,
"content": "exact scripts will be added together with the SLASH source code.",
"type": "text"
}
],
"index": 4
}
],
"index": 3
},
{
"type": "title",
"bbox": [
108,
175,
366,
186
],
"lines": [
{
"bbox": [
105,
173,
367,
189
],
"spans": [
{
"bbox": [
105,
173,
367,
189
],
"score": 1.0,
"content": "C.2 AVERAGE PRECISION COMPUTATION (SHAPEWORLD4)",
"type": "text"
}
],
"index": 5
}
],
"index": 5
},
{
"type": "text",
"bbox": [
107,
195,
505,
273
],
"lines": [
{
"bbox": [
105,
195,
505,
209
],
"spans": [
{
"bbox": [
105,
195,
505,
209
],
"score": 1.0,
"content": "For the baseline slot encoder experiments on ShapeWorld4 we measured the average precision score",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
207,
505,
220
],
"spans": [
{
"bbox": [
105,
207,
505,
220
],
"score": 1.0,
"content": "as in Locatello et al. (2020). In comparison to the baseline slot encoder, when applying SLASH",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
217,
506,
231
],
"spans": [
{
"bbox": [
105,
217,
506,
231
],
"score": 1.0,
"content": "Attention, however, we handled the case of a slot not containing an object, e.g. only background",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
229,
506,
241
],
"spans": [
{
"bbox": [
106,
229,
506,
241
],
"score": 1.0,
"content": "variables, differently. Whereas Locatello et al. (2020) add an additional binary identifier to the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
239,
506,
252
],
"spans": [
{
"bbox": [
105,
239,
506,
252
],
"score": 1.0,
"content": "multi-label ground truth vectors, we have added a background (bg) attribute to each category (cf.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
250,
505,
263
],
"spans": [
{
"bbox": [
105,
250,
505,
263
],
"score": 1.0,
"content": "Fig. 6). A slot is thus considered to be empty (i.e. not containing an object) if each NPP returns the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
262,
316,
275
],
"spans": [
{
"bbox": [
105,
262,
266,
275
],
"score": 1.0,
"content": "maximal conditional probability for the",
"type": "text"
},
{
"bbox": [
266,
262,
278,
273
],
"score": 0.81,
"content": "b g",
"type": "inline_equation"
},
{
"bbox": [
278,
262,
316,
275
],
"score": 1.0,
"content": "attribute.",
"type": "text"
}
],
"index": 12
}
],
"index": 9
},
{
"type": "text",
"bbox": [
108,
273,
505,
306
],
"lines": [
{
"bbox": [
106,
273,
505,
285
],
"spans": [
{
"bbox": [
106,
273,
505,
285
],
"score": 1.0,
"content": "As the ShapeWorld4 prediction task only included discrete object properties both for Slot Attention",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
283,
505,
296
],
"spans": [
{
"bbox": [
105,
283,
505,
296
],
"score": 1.0,
"content": "as well as for SLASH Attention the distance threshold for the average precision computation was",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
295,
287,
307
],
"spans": [
{
"bbox": [
106,
295,
287,
307
],
"score": 1.0,
"content": "infinity (thus corresponding to no threshold).",
"type": "text"
}
],
"index": 15
}
],
"index": 14
},
{
"type": "title",
"bbox": [
108,
320,
206,
331
],
"lines": [
{
"bbox": [
105,
318,
208,
334
],
"spans": [
{
"bbox": [
105,
318,
208,
334
],
"score": 1.0,
"content": "C.3 MODEL DETAILS",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
107,
341,
505,
385
],
"lines": [
{
"bbox": [
105,
339,
506,
355
],
"spans": [
{
"bbox": [
105,
339,
506,
355
],
"score": 1.0,
"content": "For those experiments using NPPs with PC we have used Einsum Networks (EiNets) for implementing",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
352,
506,
365
],
"spans": [
{
"bbox": [
105,
352,
506,
365
],
"score": 1.0,
"content": "the probabilistic circuits. EiNets are a novel implementation design for SPNs introduced by Peharz",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
364,
506,
375
],
"spans": [
{
"bbox": [
105,
364,
506,
375
],
"score": 1.0,
"content": "et al. (2020) that minimize the issue of computational costs that initial SPNs had suffered. This is",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
374,
507,
388
],
"spans": [
{
"bbox": [
105,
374,
507,
388
],
"score": 1.0,
"content": "accomplished by combining several arithmetic operations via a single monolithic einsum-operation.",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5
},
{
"type": "text",
"bbox": [
107,
390,
504,
413
],
"lines": [
{
"bbox": [
106,
390,
502,
403
],
"spans": [
{
"bbox": [
106,
390,
394,
403
],
"score": 1.0,
"content": "For all experiments, the ADAM optimizer (Kingma & Ba, 2015) with",
"type": "text"
},
{
"bbox": [
395,
391,
434,
402
],
"score": 0.91,
"content": "\\beta 1 = 0 . 9",
"type": "inline_equation"
},
{
"bbox": [
435,
390,
453,
403
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
453,
391,
502,
403
],
"score": 0.9,
"content": "\\beta 2 = 0 . 9 9 9",
"type": "inline_equation"
}
],
"index": 21
},
{
"bbox": [
107,
402,
279,
414
],
"spans": [
{
"bbox": [
107,
402,
152,
412
],
"score": 0.89,
"content": "\\epsilon = 1 e - 8",
"type": "inline_equation"
},
{
"bbox": [
152,
402,
279,
414
],
"score": 1.0,
"content": "and no weight decay was used.",
"type": "text"
}
],
"index": 22
}
],
"index": 21.5
},
{
"type": "text",
"bbox": [
106,
419,
505,
496
],
"lines": [
{
"bbox": [
105,
418,
505,
433
],
"spans": [
{
"bbox": [
105,
418,
505,
433
],
"score": 1.0,
"content": "MNIST-Addition Experiments For the MNIST-Addition experiments, we ran the DeepProbLog",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
429,
506,
443
],
"spans": [
{
"bbox": [
105,
429,
506,
443
],
"score": 1.0,
"content": "and NeurASP programs with their original configurations, as stated in (Manhaeve et al., 2018) and",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
441,
505,
453
],
"spans": [
{
"bbox": [
106,
441,
505,
453
],
"score": 1.0,
"content": "(Yang et al., 2020), respectively. For the SLASH MNIST-Addition experiments, we have used the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
452,
505,
464
],
"spans": [
{
"bbox": [
105,
452,
505,
464
],
"score": 1.0,
"content": "same neural module as in DeepProbLog and NeurASP, when training SLASH with the neural NPP",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
462,
506,
476
],
"spans": [
{
"bbox": [
105,
462,
506,
476
],
"score": 1.0,
"content": "(SLASH (DNN)) represented in Tab. 2. When using a PC NPP (SLASH (PC)) we have used an EiNet",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
474,
506,
487
],
"spans": [
{
"bbox": [
105,
474,
506,
487
],
"score": 1.0,
"content": "with the Poon-Domingos (PD) structure (Poon & Domingos, 2011) and normal distribution for the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
485,
392,
497
],
"spans": [
{
"bbox": [
105,
485,
392,
497
],
"score": 1.0,
"content": "leafs. The formal hyperparameters for the EiNet are depicted in Tab. 3.",
"type": "text"
}
],
"index": 29
}
],
"index": 26
},
{
"type": "text",
"bbox": [
107,
501,
504,
524
],
"lines": [
{
"bbox": [
105,
500,
505,
515
],
"spans": [
{
"bbox": [
105,
500,
505,
515
],
"score": 1.0,
"content": "The learning rate and batch size for the DNN were 0.005 and 100, for DeepProbLog, NeurASP and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
513,
334,
524
],
"spans": [
{
"bbox": [
106,
513,
334,
524
],
"score": 1.0,
"content": "SLASH (DNN). For the EiNet, these were 0.01 and 100.",
"type": "text"
}
],
"index": 31
}
],
"index": 30.5
},
{
"type": "table",
"bbox": [
165,
555,
446,
676
],
"blocks": [
{
"type": "table_caption",
"bbox": [
165,
534,
441,
546
],
"group_id": 0,
"lines": [
{
"bbox": [
165,
533,
442,
548
],
"spans": [
{
"bbox": [
165,
533,
442,
548
],
"score": 1.0,
"content": "Table 2: Neural module – LeNet5 for MNIST-Addition experiments.",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "table_body",
"bbox": [
165,
555,
446,
676
],
"group_id": 0,
"lines": [
{
"bbox": [
165,
555,
446,
676
],
"spans": [
{
"bbox": [
165,
555,
446,
676
],
"score": 0.982,
"html": "| Type | Size/Channels | Activation | Comment |
| Encoder | - | 二 | - |
| Conv 5 x 5 | 1x28x28 | 1 | stride 1 |
| MaxPool2d | 6x24x24 | ReLU | kernel size 2, stride 2 |
| Conv 5 x 5 | 6x12x12 | 1 | stride 1 |
| MaxPool2d | 16x8x8 | ReLU | kernel size 2,stride 2 |
| Classifier | 1 | - | - |
| MLP | 16x4x4,120 | ReLU | - |
| MLP | 120,84 | ReLU | - |
| MLP | 84,10 | 1 | Softmax |
",
"type": "table",
"image_path": "f70c0188a3a6753bf5fd03b4241a0ce55a0b8fc7d6e5338c2bb42fd3d7818bc9.jpg"
}
]
}
],
"index": 34,
"virtual_lines": [
{
"bbox": [
165,
555,
446,
595.3333333333334
],
"spans": [],
"index": 33
},
{
"bbox": [
165,
595.3333333333334,
446,
635.6666666666667
],
"spans": [],
"index": 34
},
{
"bbox": [
165,
635.6666666666667,
446,
676.0000000000001
],
"spans": [],
"index": 35
}
]
}
],
"index": 33.0
},
{
"type": "text",
"bbox": [
107,
687,
505,
732
],
"lines": [
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "ShapeWorld4 Experiments For the baseline slot attention experiments with the ShapeWorld4 data",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
699,
505,
711
],
"spans": [
{
"bbox": [
105,
699,
505,
711
],
"score": 1.0,
"content": "set we have used the architecture presented in Tab. 4. For further details on this, we refer to the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
710,
505,
721
],
"spans": [
{
"bbox": [
105,
710,
505,
721
],
"score": 1.0,
"content": "original work of Locatello et al. (2020). The slot encoder had a number of 4 slots and 3 attention",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
721,
231,
732
],
"spans": [
{
"bbox": [
105,
721,
231,
732
],
"score": 1.0,
"content": "iterations over all experiments.",
"type": "text"
}
],
"index": 39
}
],
"index": 37.5
}
],
"page_idx": 15,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"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": "16",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
81,
335,
94
],
"lines": [
{
"bbox": [
106,
79,
337,
96
],
"spans": [
{
"bbox": [
106,
79,
337,
96
],
"score": 1.0,
"content": "C APPENDIX C – EXPERIMENTAL DETAILS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "title",
"bbox": [
108,
106,
260,
118
],
"lines": [
{
"bbox": [
106,
105,
262,
119
],
"spans": [
{
"bbox": [
106,
105,
262,
119
],
"score": 1.0,
"content": "C.1 SHAPEWORLD4 GENERATION",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "text",
"bbox": [
107,
127,
505,
160
],
"lines": [
{
"bbox": [
105,
126,
505,
141
],
"spans": [
{
"bbox": [
105,
126,
505,
141
],
"score": 1.0,
"content": "The ShapeWorld4 and ShapeWorld4 CoGenT data sets were generated using the original scripts",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
138,
505,
151
],
"spans": [
{
"bbox": [
105,
138,
505,
151
],
"score": 1.0,
"content": "of (Kuhnle & Copestake, 2017) (https://github.com/AlexKuhnle/ShapeWorld). The",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
150,
371,
162
],
"spans": [
{
"bbox": [
105,
150,
371,
162
],
"score": 1.0,
"content": "exact scripts will be added together with the SLASH source code.",
"type": "text"
}
],
"index": 4
}
],
"index": 3,
"bbox_fs": [
105,
126,
505,
162
]
},
{
"type": "title",
"bbox": [
108,
175,
366,
186
],
"lines": [
{
"bbox": [
105,
173,
367,
189
],
"spans": [
{
"bbox": [
105,
173,
367,
189
],
"score": 1.0,
"content": "C.2 AVERAGE PRECISION COMPUTATION (SHAPEWORLD4)",
"type": "text"
}
],
"index": 5
}
],
"index": 5
},
{
"type": "text",
"bbox": [
107,
195,
505,
273
],
"lines": [
{
"bbox": [
105,
195,
505,
209
],
"spans": [
{
"bbox": [
105,
195,
505,
209
],
"score": 1.0,
"content": "For the baseline slot encoder experiments on ShapeWorld4 we measured the average precision score",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
207,
505,
220
],
"spans": [
{
"bbox": [
105,
207,
505,
220
],
"score": 1.0,
"content": "as in Locatello et al. (2020). In comparison to the baseline slot encoder, when applying SLASH",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
217,
506,
231
],
"spans": [
{
"bbox": [
105,
217,
506,
231
],
"score": 1.0,
"content": "Attention, however, we handled the case of a slot not containing an object, e.g. only background",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
229,
506,
241
],
"spans": [
{
"bbox": [
106,
229,
506,
241
],
"score": 1.0,
"content": "variables, differently. Whereas Locatello et al. (2020) add an additional binary identifier to the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
239,
506,
252
],
"spans": [
{
"bbox": [
105,
239,
506,
252
],
"score": 1.0,
"content": "multi-label ground truth vectors, we have added a background (bg) attribute to each category (cf.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
250,
505,
263
],
"spans": [
{
"bbox": [
105,
250,
505,
263
],
"score": 1.0,
"content": "Fig. 6). A slot is thus considered to be empty (i.e. not containing an object) if each NPP returns the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
262,
316,
275
],
"spans": [
{
"bbox": [
105,
262,
266,
275
],
"score": 1.0,
"content": "maximal conditional probability for the",
"type": "text"
},
{
"bbox": [
266,
262,
278,
273
],
"score": 0.81,
"content": "b g",
"type": "inline_equation"
},
{
"bbox": [
278,
262,
316,
275
],
"score": 1.0,
"content": "attribute.",
"type": "text"
}
],
"index": 12
}
],
"index": 9,
"bbox_fs": [
105,
195,
506,
275
]
},
{
"type": "text",
"bbox": [
108,
273,
505,
306
],
"lines": [
{
"bbox": [
106,
273,
505,
285
],
"spans": [
{
"bbox": [
106,
273,
505,
285
],
"score": 1.0,
"content": "As the ShapeWorld4 prediction task only included discrete object properties both for Slot Attention",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
283,
505,
296
],
"spans": [
{
"bbox": [
105,
283,
505,
296
],
"score": 1.0,
"content": "as well as for SLASH Attention the distance threshold for the average precision computation was",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
295,
287,
307
],
"spans": [
{
"bbox": [
106,
295,
287,
307
],
"score": 1.0,
"content": "infinity (thus corresponding to no threshold).",
"type": "text"
}
],
"index": 15
}
],
"index": 14,
"bbox_fs": [
105,
273,
505,
307
]
},
{
"type": "title",
"bbox": [
108,
320,
206,
331
],
"lines": [
{
"bbox": [
105,
318,
208,
334
],
"spans": [
{
"bbox": [
105,
318,
208,
334
],
"score": 1.0,
"content": "C.3 MODEL DETAILS",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
107,
341,
505,
385
],
"lines": [
{
"bbox": [
105,
339,
506,
355
],
"spans": [
{
"bbox": [
105,
339,
506,
355
],
"score": 1.0,
"content": "For those experiments using NPPs with PC we have used Einsum Networks (EiNets) for implementing",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
352,
506,
365
],
"spans": [
{
"bbox": [
105,
352,
506,
365
],
"score": 1.0,
"content": "the probabilistic circuits. EiNets are a novel implementation design for SPNs introduced by Peharz",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
364,
506,
375
],
"spans": [
{
"bbox": [
105,
364,
506,
375
],
"score": 1.0,
"content": "et al. (2020) that minimize the issue of computational costs that initial SPNs had suffered. This is",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
374,
507,
388
],
"spans": [
{
"bbox": [
105,
374,
507,
388
],
"score": 1.0,
"content": "accomplished by combining several arithmetic operations via a single monolithic einsum-operation.",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5,
"bbox_fs": [
105,
339,
507,
388
]
},
{
"type": "text",
"bbox": [
107,
390,
504,
413
],
"lines": [
{
"bbox": [
106,
390,
502,
403
],
"spans": [
{
"bbox": [
106,
390,
394,
403
],
"score": 1.0,
"content": "For all experiments, the ADAM optimizer (Kingma & Ba, 2015) with",
"type": "text"
},
{
"bbox": [
395,
391,
434,
402
],
"score": 0.91,
"content": "\\beta 1 = 0 . 9",
"type": "inline_equation"
},
{
"bbox": [
435,
390,
453,
403
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
453,
391,
502,
403
],
"score": 0.9,
"content": "\\beta 2 = 0 . 9 9 9",
"type": "inline_equation"
}
],
"index": 21
},
{
"bbox": [
107,
402,
279,
414
],
"spans": [
{
"bbox": [
107,
402,
152,
412
],
"score": 0.89,
"content": "\\epsilon = 1 e - 8",
"type": "inline_equation"
},
{
"bbox": [
152,
402,
279,
414
],
"score": 1.0,
"content": "and no weight decay was used.",
"type": "text"
}
],
"index": 22
}
],
"index": 21.5,
"bbox_fs": [
106,
390,
502,
414
]
},
{
"type": "text",
"bbox": [
106,
419,
505,
496
],
"lines": [
{
"bbox": [
105,
418,
505,
433
],
"spans": [
{
"bbox": [
105,
418,
505,
433
],
"score": 1.0,
"content": "MNIST-Addition Experiments For the MNIST-Addition experiments, we ran the DeepProbLog",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
429,
506,
443
],
"spans": [
{
"bbox": [
105,
429,
506,
443
],
"score": 1.0,
"content": "and NeurASP programs with their original configurations, as stated in (Manhaeve et al., 2018) and",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
441,
505,
453
],
"spans": [
{
"bbox": [
106,
441,
505,
453
],
"score": 1.0,
"content": "(Yang et al., 2020), respectively. For the SLASH MNIST-Addition experiments, we have used the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
452,
505,
464
],
"spans": [
{
"bbox": [
105,
452,
505,
464
],
"score": 1.0,
"content": "same neural module as in DeepProbLog and NeurASP, when training SLASH with the neural NPP",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
462,
506,
476
],
"spans": [
{
"bbox": [
105,
462,
506,
476
],
"score": 1.0,
"content": "(SLASH (DNN)) represented in Tab. 2. When using a PC NPP (SLASH (PC)) we have used an EiNet",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
474,
506,
487
],
"spans": [
{
"bbox": [
105,
474,
506,
487
],
"score": 1.0,
"content": "with the Poon-Domingos (PD) structure (Poon & Domingos, 2011) and normal distribution for the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
485,
392,
497
],
"spans": [
{
"bbox": [
105,
485,
392,
497
],
"score": 1.0,
"content": "leafs. The formal hyperparameters for the EiNet are depicted in Tab. 3.",
"type": "text"
}
],
"index": 29
}
],
"index": 26,
"bbox_fs": [
105,
418,
506,
497
]
},
{
"type": "text",
"bbox": [
107,
501,
504,
524
],
"lines": [
{
"bbox": [
105,
500,
505,
515
],
"spans": [
{
"bbox": [
105,
500,
505,
515
],
"score": 1.0,
"content": "The learning rate and batch size for the DNN were 0.005 and 100, for DeepProbLog, NeurASP and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
513,
334,
524
],
"spans": [
{
"bbox": [
106,
513,
334,
524
],
"score": 1.0,
"content": "SLASH (DNN). For the EiNet, these were 0.01 and 100.",
"type": "text"
}
],
"index": 31
}
],
"index": 30.5,
"bbox_fs": [
105,
500,
505,
524
]
},
{
"type": "table",
"bbox": [
165,
555,
446,
676
],
"blocks": [
{
"type": "table_caption",
"bbox": [
165,
534,
441,
546
],
"group_id": 0,
"lines": [
{
"bbox": [
165,
533,
442,
548
],
"spans": [
{
"bbox": [
165,
533,
442,
548
],
"score": 1.0,
"content": "Table 2: Neural module – LeNet5 for MNIST-Addition experiments.",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "table_body",
"bbox": [
165,
555,
446,
676
],
"group_id": 0,
"lines": [
{
"bbox": [
165,
555,
446,
676
],
"spans": [
{
"bbox": [
165,
555,
446,
676
],
"score": 0.982,
"html": "| Type | Size/Channels | Activation | Comment |
| Encoder | - | 二 | - |
| Conv 5 x 5 | 1x28x28 | 1 | stride 1 |
| MaxPool2d | 6x24x24 | ReLU | kernel size 2, stride 2 |
| Conv 5 x 5 | 6x12x12 | 1 | stride 1 |
| MaxPool2d | 16x8x8 | ReLU | kernel size 2,stride 2 |
| Classifier | 1 | - | - |
| MLP | 16x4x4,120 | ReLU | - |
| MLP | 120,84 | ReLU | - |
| MLP | 84,10 | 1 | Softmax |
",
"type": "table",
"image_path": "f70c0188a3a6753bf5fd03b4241a0ce55a0b8fc7d6e5338c2bb42fd3d7818bc9.jpg"
}
]
}
],
"index": 34,
"virtual_lines": [
{
"bbox": [
165,
555,
446,
595.3333333333334
],
"spans": [],
"index": 33
},
{
"bbox": [
165,
595.3333333333334,
446,
635.6666666666667
],
"spans": [],
"index": 34
},
{
"bbox": [
165,
635.6666666666667,
446,
676.0000000000001
],
"spans": [],
"index": 35
}
]
}
],
"index": 33.0
},
{
"type": "text",
"bbox": [
107,
687,
505,
732
],
"lines": [
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "ShapeWorld4 Experiments For the baseline slot attention experiments with the ShapeWorld4 data",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
699,
505,
711
],
"spans": [
{
"bbox": [
105,
699,
505,
711
],
"score": 1.0,
"content": "set we have used the architecture presented in Tab. 4. For further details on this, we refer to the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
710,
505,
721
],
"spans": [
{
"bbox": [
105,
710,
505,
721
],
"score": 1.0,
"content": "original work of Locatello et al. (2020). The slot encoder had a number of 4 slots and 3 attention",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
721,
231,
732
],
"spans": [
{
"bbox": [
105,
721,
231,
732
],
"score": 1.0,
"content": "iterations over all experiments.",
"type": "text"
}
],
"index": 39
}
],
"index": 37.5,
"bbox_fs": [
105,
687,
506,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
171,
101,
440,
130
],
"blocks": [
{
"type": "table_caption",
"bbox": [
141,
80,
465,
92
],
"group_id": 0,
"lines": [
{
"bbox": [
141,
78,
465,
94
],
"spans": [
{
"bbox": [
141,
78,
465,
94
],
"score": 1.0,
"content": "Table 3: Probabilistic Circuit module – EiNet for MNIST-Addition experiments.",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "table_body",
"bbox": [
171,
101,
440,
130
],
"group_id": 0,
"lines": [
{
"bbox": [
171,
101,
440,
130
],
"spans": [
{
"bbox": [
171,
101,
440,
130
],
"score": 0.953,
"html": "| Variables | Width | Height | Number of Pieces | Class count |
| 784 | 28 | 28 | [4,7,28] | 10 |
",
"type": "table",
"image_path": "055619a651ca94f260eb57626f398d17ef7693cb21b9b36feb10922133cfeb38.jpg"
}
]
}
],
"index": 2,
"virtual_lines": [
{
"bbox": [
171,
101,
440,
110.66666666666667
],
"spans": [],
"index": 1
},
{
"bbox": [
171,
110.66666666666667,
440,
120.33333333333334
],
"spans": [],
"index": 2
},
{
"bbox": [
171,
120.33333333333334,
440,
130.0
],
"spans": [],
"index": 3
}
]
}
],
"index": 1.0
},
{
"type": "table",
"bbox": [
150,
162,
461,
315
],
"blocks": [
{
"type": "table_caption",
"bbox": [
180,
140,
427,
152
],
"group_id": 1,
"lines": [
{
"bbox": [
179,
138,
428,
154
],
"spans": [
{
"bbox": [
179,
138,
428,
154
],
"score": 1.0,
"content": "Table 4: Baseline slot encoder for ShapeWorld4 experiments.",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "table_body",
"bbox": [
150,
162,
461,
315
],
"group_id": 1,
"lines": [
{
"bbox": [
150,
162,
461,
315
],
"spans": [
{
"bbox": [
150,
162,
461,
315
],
"score": 0.985,
"html": "| Type | Size/Channels | Activation | Comment |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Position Embedding | 1 | 二 | - |
| Flatten | axis: [0, 1,2 x 3] | - | flatten x, y pos. |
| Layer Norm | 二 | - | 二 |
| MLP (per location) | 32 | ReLU | 二 |
| MLP (per location) | 32 | - | 二 |
| SlotAttentionModule | 32 | ReLU | 二 |
| MLP | 32 | ReLU | |
| MLP | 16 | Sigmoid | 二 1 |
",
"type": "table",
"image_path": "7a6e8fe39959ba94229a5e2822d9f8343254e45215672d154eb62542dc807e0b.jpg"
}
]
}
],
"index": 6,
"virtual_lines": [
{
"bbox": [
150,
162,
461,
213.0
],
"spans": [],
"index": 5
},
{
"bbox": [
150,
213.0,
461,
264.0
],
"spans": [],
"index": 6
},
{
"bbox": [
150,
264.0,
461,
315.0
],
"spans": [],
"index": 7
}
]
}
],
"index": 5.0
},
{
"type": "text",
"bbox": [
107,
335,
505,
369
],
"lines": [
{
"bbox": [
105,
335,
505,
347
],
"spans": [
{
"bbox": [
105,
335,
505,
347
],
"score": 1.0,
"content": "For the SLASH Attention experiments with ShapeWorld4 we have used the same slot encoder as in",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
345,
505,
359
],
"spans": [
{
"bbox": [
105,
345,
505,
359
],
"score": 1.0,
"content": "Tab. 4, however, we replaced the final MLPs with 4 individual EiNets with Poon-Domingos structure",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
357,
414,
370
],
"spans": [
{
"bbox": [
106,
357,
414,
370
],
"score": 1.0,
"content": "(Poon & Domingos, 2011). Their hyperparameters are represented in Tab. 5.",
"type": "text"
}
],
"index": 10
}
],
"index": 9
},
{
"type": "table",
"bbox": [
152,
400,
458,
463
],
"blocks": [
{
"type": "table_caption",
"bbox": [
150,
379,
456,
391
],
"group_id": 2,
"lines": [
{
"bbox": [
149,
376,
458,
394
],
"spans": [
{
"bbox": [
149,
376,
458,
394
],
"score": 1.0,
"content": "Table 5: Probabilistic Circuit module – EiNet for ShapeWorld4 experiments.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "table_body",
"bbox": [
152,
400,
458,
463
],
"group_id": 2,
"lines": [
{
"bbox": [
152,
400,
458,
463
],
"spans": [
{
"bbox": [
152,
400,
458,
463
],
"score": 0.979,
"html": "| EiNet | Variables | Width | Height | Number ofPieces | Class count |
| Color | 32 | 8 | 4 | [4] | 9 |
| Shape | 32 | 8 | 4 | [4] | 4 |
| Shade | 32 | 8 | 4 | [4] | 3 |
| Size | 32 | 8 | 4 | [4] | 3 |
",
"type": "table",
"image_path": "bf6ad54fdca37e6aada941896a644326dea1a55ea9b6a6022392ceabd19cbc57.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
152,
400,
458,
421.0
],
"spans": [],
"index": 12
},
{
"bbox": [
152,
421.0,
458,
442.0
],
"spans": [],
"index": 13
},
{
"bbox": [
152,
442.0,
458,
463.0
],
"spans": [],
"index": 14
}
]
}
],
"index": 12.0
},
{
"type": "text",
"bbox": [
107,
474,
505,
497
],
"lines": [
{
"bbox": [
106,
474,
505,
486
],
"spans": [
{
"bbox": [
106,
474,
505,
486
],
"score": 1.0,
"content": "The learning rate and batch size for SLASH Attention were 0.01 and 512, for ShapeWorld4 and",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
485,
475,
498
],
"spans": [
{
"bbox": [
106,
485,
475,
498
],
"score": 1.0,
"content": "ShapeWorld4 CoGenT. The learning rate for the baseline slot encoder were 0.0004 and 512.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
}
],
"page_idx": 16,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
171,
101,
440,
130
],
"blocks": [
{
"type": "table_caption",
"bbox": [
141,
80,
465,
92
],
"group_id": 0,
"lines": [
{
"bbox": [
141,
78,
465,
94
],
"spans": [
{
"bbox": [
141,
78,
465,
94
],
"score": 1.0,
"content": "Table 3: Probabilistic Circuit module – EiNet for MNIST-Addition experiments.",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "table_body",
"bbox": [
171,
101,
440,
130
],
"group_id": 0,
"lines": [
{
"bbox": [
171,
101,
440,
130
],
"spans": [
{
"bbox": [
171,
101,
440,
130
],
"score": 0.953,
"html": "| Variables | Width | Height | Number of Pieces | Class count |
| 784 | 28 | 28 | [4,7,28] | 10 |
",
"type": "table",
"image_path": "055619a651ca94f260eb57626f398d17ef7693cb21b9b36feb10922133cfeb38.jpg"
}
]
}
],
"index": 2,
"virtual_lines": [
{
"bbox": [
171,
101,
440,
110.66666666666667
],
"spans": [],
"index": 1
},
{
"bbox": [
171,
110.66666666666667,
440,
120.33333333333334
],
"spans": [],
"index": 2
},
{
"bbox": [
171,
120.33333333333334,
440,
130.0
],
"spans": [],
"index": 3
}
]
}
],
"index": 1.0
},
{
"type": "table",
"bbox": [
150,
162,
461,
315
],
"blocks": [
{
"type": "table_caption",
"bbox": [
180,
140,
427,
152
],
"group_id": 1,
"lines": [
{
"bbox": [
179,
138,
428,
154
],
"spans": [
{
"bbox": [
179,
138,
428,
154
],
"score": 1.0,
"content": "Table 4: Baseline slot encoder for ShapeWorld4 experiments.",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "table_body",
"bbox": [
150,
162,
461,
315
],
"group_id": 1,
"lines": [
{
"bbox": [
150,
162,
461,
315
],
"spans": [
{
"bbox": [
150,
162,
461,
315
],
"score": 0.985,
"html": "| Type | Size/Channels | Activation | Comment |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Conv 5 x 5 | 32 | ReLU | stride 1 |
| Position Embedding | 1 | 二 | - |
| Flatten | axis: [0, 1,2 x 3] | - | flatten x, y pos. |
| Layer Norm | 二 | - | 二 |
| MLP (per location) | 32 | ReLU | 二 |
| MLP (per location) | 32 | - | 二 |
| SlotAttentionModule | 32 | ReLU | 二 |
| MLP | 32 | ReLU | |
| MLP | 16 | Sigmoid | 二 1 |
",
"type": "table",
"image_path": "7a6e8fe39959ba94229a5e2822d9f8343254e45215672d154eb62542dc807e0b.jpg"
}
]
}
],
"index": 6,
"virtual_lines": [
{
"bbox": [
150,
162,
461,
213.0
],
"spans": [],
"index": 5
},
{
"bbox": [
150,
213.0,
461,
264.0
],
"spans": [],
"index": 6
},
{
"bbox": [
150,
264.0,
461,
315.0
],
"spans": [],
"index": 7
}
]
}
],
"index": 5.0
},
{
"type": "text",
"bbox": [
107,
335,
505,
369
],
"lines": [
{
"bbox": [
105,
335,
505,
347
],
"spans": [
{
"bbox": [
105,
335,
505,
347
],
"score": 1.0,
"content": "For the SLASH Attention experiments with ShapeWorld4 we have used the same slot encoder as in",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
345,
505,
359
],
"spans": [
{
"bbox": [
105,
345,
505,
359
],
"score": 1.0,
"content": "Tab. 4, however, we replaced the final MLPs with 4 individual EiNets with Poon-Domingos structure",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
357,
414,
370
],
"spans": [
{
"bbox": [
106,
357,
414,
370
],
"score": 1.0,
"content": "(Poon & Domingos, 2011). Their hyperparameters are represented in Tab. 5.",
"type": "text"
}
],
"index": 10
}
],
"index": 9,
"bbox_fs": [
105,
335,
505,
370
]
},
{
"type": "table",
"bbox": [
152,
400,
458,
463
],
"blocks": [
{
"type": "table_caption",
"bbox": [
150,
379,
456,
391
],
"group_id": 2,
"lines": [
{
"bbox": [
149,
376,
458,
394
],
"spans": [
{
"bbox": [
149,
376,
458,
394
],
"score": 1.0,
"content": "Table 5: Probabilistic Circuit module – EiNet for ShapeWorld4 experiments.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "table_body",
"bbox": [
152,
400,
458,
463
],
"group_id": 2,
"lines": [
{
"bbox": [
152,
400,
458,
463
],
"spans": [
{
"bbox": [
152,
400,
458,
463
],
"score": 0.979,
"html": "| EiNet | Variables | Width | Height | Number ofPieces | Class count |
| Color | 32 | 8 | 4 | [4] | 9 |
| Shape | 32 | 8 | 4 | [4] | 4 |
| Shade | 32 | 8 | 4 | [4] | 3 |
| Size | 32 | 8 | 4 | [4] | 3 |
",
"type": "table",
"image_path": "bf6ad54fdca37e6aada941896a644326dea1a55ea9b6a6022392ceabd19cbc57.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
152,
400,
458,
421.0
],
"spans": [],
"index": 12
},
{
"bbox": [
152,
421.0,
458,
442.0
],
"spans": [],
"index": 13
},
{
"bbox": [
152,
442.0,
458,
463.0
],
"spans": [],
"index": 14
}
]
}
],
"index": 12.0
},
{
"type": "text",
"bbox": [
107,
474,
505,
497
],
"lines": [
{
"bbox": [
106,
474,
505,
486
],
"spans": [
{
"bbox": [
106,
474,
505,
486
],
"score": 1.0,
"content": "The learning rate and batch size for SLASH Attention were 0.01 and 512, for ShapeWorld4 and",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
485,
475,
498
],
"spans": [
{
"bbox": [
106,
485,
475,
498
],
"score": 1.0,
"content": "ShapeWorld4 CoGenT. The learning rate for the baseline slot encoder were 0.0004 and 512.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5,
"bbox_fs": [
106,
474,
505,
498
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
105,
80,
439,
94
],
"lines": [
{
"bbox": [
105,
80,
441,
96
],
"spans": [
{
"bbox": [
105,
80,
441,
96
],
"score": 1.0,
"content": "D APPENDIX D - ADDITIONAL RESULTS ON MNIST ADDITION",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
108,
505,
154
],
"lines": [
{
"bbox": [
104,
108,
506,
122
],
"spans": [
{
"bbox": [
104,
108,
506,
122
],
"score": 1.0,
"content": "Table 6: Additional MNIST Addition Results. Test accuracy corresponds to the percentage of",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
120,
505,
132
],
"spans": [
{
"bbox": [
106,
120,
505,
132
],
"score": 1.0,
"content": "correctly classified test images. Both models (DeepProbLog and SLASH (PC)) were trained on the",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
130,
507,
144
],
"spans": [
{
"bbox": [
105,
130,
507,
144
],
"score": 1.0,
"content": "full MNIST data, but tested on images with missing pixels. Test accuracies are presented in percent.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
104,
141,
455,
155
],
"spans": [
{
"bbox": [
104,
141,
300,
155
],
"score": 1.0,
"content": "The amount of missing data was varied between",
"type": "text"
},
{
"bbox": [
300,
142,
320,
153
],
"score": 0.85,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
320,
141,
338,
155
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
338,
142,
358,
153
],
"score": 0.87,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
358,
141,
455,
155
],
"score": 1.0,
"content": "of the pixels per image.",
"type": "text"
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "table",
"bbox": [
227,
162,
404,
223
],
"blocks": [
{
"type": "table_body",
"bbox": [
227,
162,
404,
223
],
"group_id": 0,
"lines": [
{
"bbox": [
227,
162,
404,
223
],
"spans": [
{
"bbox": [
227,
162,
404,
223
],
"score": 0.977,
"html": " | DeepProbLog | SLASH (PC) |
| 50% | 79.94 ± 7.2 | 72.2 ± 12.15 |
| 80% | 31.6± 6.08 | 44.2± 8.23 |
| 90% | 16.94 ± 1.76 | 29.6± 5.77 |
| 97% | 12.33 ± 0.47 | 17.6 ± 2.97 |
",
"type": "table",
"image_path": "2187ddca008e51592b0fa5460a51944b779f5cb0f908f13fbbd1c2ac7dff9efa.jpg"
}
]
}
],
"index": 6.5,
"virtual_lines": [
{
"bbox": [
227,
162,
404,
177.25
],
"spans": [],
"index": 5
},
{
"bbox": [
227,
177.25,
404,
192.5
],
"spans": [],
"index": 6
},
{
"bbox": [
227,
192.5,
404,
207.75
],
"spans": [],
"index": 7
},
{
"bbox": [
227,
207.75,
404,
223.0
],
"spans": [],
"index": 8
}
]
}
],
"index": 6.5
},
{
"type": "text",
"bbox": [
106,
232,
504,
255
],
"lines": [
{
"bbox": [
106,
232,
505,
245
],
"spans": [
{
"bbox": [
106,
232,
505,
245
],
"score": 1.0,
"content": "In addition to the setting considered in Evaluation 2, we adjusted the settings for the MNIST Addition",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
243,
292,
257
],
"spans": [
{
"bbox": [
105,
243,
292,
257
],
"score": 1.0,
"content": "task with missing data into the following way.",
"type": "text"
}
],
"index": 10
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
106,
260,
505,
317
],
"lines": [
{
"bbox": [
106,
261,
505,
273
],
"spans": [
{
"bbox": [
106,
261,
505,
273
],
"score": 1.0,
"content": "Training was performed with the full MNIST data set, however the test data set contained different",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
271,
505,
285
],
"spans": [
{
"bbox": [
105,
271,
505,
285
],
"score": 1.0,
"content": "rates of missing pixels. Whereas using an NPP with a PC allows, among other things, to compute",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
282,
505,
295
],
"spans": [
{
"bbox": [
106,
282,
505,
295
],
"score": 1.0,
"content": "marginalization “out of the box” without requiring an update to the architecture or a retraining, this is",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
293,
506,
306
],
"spans": [
{
"bbox": [
106,
293,
506,
306
],
"score": 1.0,
"content": "not so trivial for purely neural-based predicates as in DeepProbLog. Thus, we allowed SLASH (PC)",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
303,
483,
319
],
"spans": [
{
"bbox": [
105,
303,
483,
319
],
"score": 1.0,
"content": "to marginalize over the missing pixels, where this was not directly possible for DeepProbLog.",
"type": "text"
}
],
"index": 15
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
320,
505,
388
],
"lines": [
{
"bbox": [
105,
320,
506,
334
],
"spans": [
{
"bbox": [
105,
320,
249,
334
],
"score": 1.0,
"content": "The results can be seen in Tab. 6 for",
"type": "text"
},
{
"bbox": [
250,
321,
269,
333
],
"score": 0.87,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
270,
320,
272,
334
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
273,
321,
292,
333
],
"score": 0.86,
"content": "8 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
293,
320,
295,
334
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
296,
321,
316,
333
],
"score": 0.86,
"content": "9 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
316,
320,
333,
334
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
333,
321,
353,
332
],
"score": 0.89,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
353,
320,
506,
334
],
"score": 1.0,
"content": "of missing pixels per image in the test",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
331,
505,
344
],
"spans": [
{
"bbox": [
106,
331,
202,
344
],
"score": 1.0,
"content": "set. We observe that at",
"type": "text"
},
{
"bbox": [
202,
332,
222,
343
],
"score": 0.89,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
222,
331,
505,
344
],
"score": 1.0,
"content": ", DeepProbLog outperforms SLASH by a small margin. For all other",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
341,
507,
358
],
"spans": [
{
"bbox": [
104,
341,
507,
358
],
"score": 1.0,
"content": "rates, we observe that SLASH (PC) reaches significantly higher test accuracies than DeepProbLog.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
354,
506,
367
],
"spans": [
{
"bbox": [
105,
354,
506,
367
],
"score": 1.0,
"content": "However, we remark that in this setting, SLASH produces larger standard deviations in comparison",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
365,
506,
378
],
"spans": [
{
"bbox": [
105,
365,
506,
378
],
"score": 1.0,
"content": "to DeepProbLog. These results indicate that the conclusions, drawn in the main part of our work,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
376,
332,
389
],
"spans": [
{
"bbox": [
105,
376,
332,
389
],
"score": 1.0,
"content": "remain true also in this setting of handling missing data.",
"type": "text"
}
],
"index": 21
}
],
"index": 18.5
}
],
"page_idx": 17,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
26,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
308,
38
],
"spans": [
{
"bbox": [
106,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
752,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
763
],
"spans": [
{
"bbox": [
299,
750,
312,
763
],
"score": 1.0,
"content": "",
"type": "text",
"height": 13,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
105,
80,
439,
94
],
"lines": [
{
"bbox": [
105,
80,
441,
96
],
"spans": [
{
"bbox": [
105,
80,
441,
96
],
"score": 1.0,
"content": "D APPENDIX D - ADDITIONAL RESULTS ON MNIST ADDITION",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
108,
505,
154
],
"lines": [
{
"bbox": [
104,
108,
506,
122
],
"spans": [
{
"bbox": [
104,
108,
506,
122
],
"score": 1.0,
"content": "Table 6: Additional MNIST Addition Results. Test accuracy corresponds to the percentage of",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
120,
505,
132
],
"spans": [
{
"bbox": [
106,
120,
505,
132
],
"score": 1.0,
"content": "correctly classified test images. Both models (DeepProbLog and SLASH (PC)) were trained on the",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
130,
507,
144
],
"spans": [
{
"bbox": [
105,
130,
507,
144
],
"score": 1.0,
"content": "full MNIST data, but tested on images with missing pixels. Test accuracies are presented in percent.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
104,
141,
455,
155
],
"spans": [
{
"bbox": [
104,
141,
300,
155
],
"score": 1.0,
"content": "The amount of missing data was varied between",
"type": "text"
},
{
"bbox": [
300,
142,
320,
153
],
"score": 0.85,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
320,
141,
338,
155
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
338,
142,
358,
153
],
"score": 0.87,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
358,
141,
455,
155
],
"score": 1.0,
"content": "of the pixels per image.",
"type": "text"
}
],
"index": 4
}
],
"index": 2.5,
"bbox_fs": [
104,
108,
507,
155
]
},
{
"type": "table",
"bbox": [
227,
162,
404,
223
],
"blocks": [
{
"type": "table_body",
"bbox": [
227,
162,
404,
223
],
"group_id": 0,
"lines": [
{
"bbox": [
227,
162,
404,
223
],
"spans": [
{
"bbox": [
227,
162,
404,
223
],
"score": 0.977,
"html": " | DeepProbLog | SLASH (PC) |
| 50% | 79.94 ± 7.2 | 72.2 ± 12.15 |
| 80% | 31.6± 6.08 | 44.2± 8.23 |
| 90% | 16.94 ± 1.76 | 29.6± 5.77 |
| 97% | 12.33 ± 0.47 | 17.6 ± 2.97 |
",
"type": "table",
"image_path": "2187ddca008e51592b0fa5460a51944b779f5cb0f908f13fbbd1c2ac7dff9efa.jpg"
}
]
}
],
"index": 6.5,
"virtual_lines": [
{
"bbox": [
227,
162,
404,
177.25
],
"spans": [],
"index": 5
},
{
"bbox": [
227,
177.25,
404,
192.5
],
"spans": [],
"index": 6
},
{
"bbox": [
227,
192.5,
404,
207.75
],
"spans": [],
"index": 7
},
{
"bbox": [
227,
207.75,
404,
223.0
],
"spans": [],
"index": 8
}
]
}
],
"index": 6.5
},
{
"type": "text",
"bbox": [
106,
232,
504,
255
],
"lines": [
{
"bbox": [
106,
232,
505,
245
],
"spans": [
{
"bbox": [
106,
232,
505,
245
],
"score": 1.0,
"content": "In addition to the setting considered in Evaluation 2, we adjusted the settings for the MNIST Addition",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
243,
292,
257
],
"spans": [
{
"bbox": [
105,
243,
292,
257
],
"score": 1.0,
"content": "task with missing data into the following way.",
"type": "text"
}
],
"index": 10
}
],
"index": 9.5,
"bbox_fs": [
105,
232,
505,
257
]
},
{
"type": "text",
"bbox": [
106,
260,
505,
317
],
"lines": [
{
"bbox": [
106,
261,
505,
273
],
"spans": [
{
"bbox": [
106,
261,
505,
273
],
"score": 1.0,
"content": "Training was performed with the full MNIST data set, however the test data set contained different",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
271,
505,
285
],
"spans": [
{
"bbox": [
105,
271,
505,
285
],
"score": 1.0,
"content": "rates of missing pixels. Whereas using an NPP with a PC allows, among other things, to compute",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
282,
505,
295
],
"spans": [
{
"bbox": [
106,
282,
505,
295
],
"score": 1.0,
"content": "marginalization “out of the box” without requiring an update to the architecture or a retraining, this is",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
293,
506,
306
],
"spans": [
{
"bbox": [
106,
293,
506,
306
],
"score": 1.0,
"content": "not so trivial for purely neural-based predicates as in DeepProbLog. Thus, we allowed SLASH (PC)",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
303,
483,
319
],
"spans": [
{
"bbox": [
105,
303,
483,
319
],
"score": 1.0,
"content": "to marginalize over the missing pixels, where this was not directly possible for DeepProbLog.",
"type": "text"
}
],
"index": 15
}
],
"index": 13,
"bbox_fs": [
105,
261,
506,
319
]
},
{
"type": "text",
"bbox": [
107,
320,
505,
388
],
"lines": [
{
"bbox": [
105,
320,
506,
334
],
"spans": [
{
"bbox": [
105,
320,
249,
334
],
"score": 1.0,
"content": "The results can be seen in Tab. 6 for",
"type": "text"
},
{
"bbox": [
250,
321,
269,
333
],
"score": 0.87,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
270,
320,
272,
334
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
273,
321,
292,
333
],
"score": 0.86,
"content": "8 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
293,
320,
295,
334
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
296,
321,
316,
333
],
"score": 0.86,
"content": "9 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
316,
320,
333,
334
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
333,
321,
353,
332
],
"score": 0.89,
"content": "9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
353,
320,
506,
334
],
"score": 1.0,
"content": "of missing pixels per image in the test",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
331,
505,
344
],
"spans": [
{
"bbox": [
106,
331,
202,
344
],
"score": 1.0,
"content": "set. We observe that at",
"type": "text"
},
{
"bbox": [
202,
332,
222,
343
],
"score": 0.89,
"content": "5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
222,
331,
505,
344
],
"score": 1.0,
"content": ", DeepProbLog outperforms SLASH by a small margin. For all other",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
341,
507,
358
],
"spans": [
{
"bbox": [
104,
341,
507,
358
],
"score": 1.0,
"content": "rates, we observe that SLASH (PC) reaches significantly higher test accuracies than DeepProbLog.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
354,
506,
367
],
"spans": [
{
"bbox": [
105,
354,
506,
367
],
"score": 1.0,
"content": "However, we remark that in this setting, SLASH produces larger standard deviations in comparison",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
365,
506,
378
],
"spans": [
{
"bbox": [
105,
365,
506,
378
],
"score": 1.0,
"content": "to DeepProbLog. These results indicate that the conclusions, drawn in the main part of our work,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
376,
332,
389
],
"spans": [
{
"bbox": [
105,
376,
332,
389
],
"score": 1.0,
"content": "remain true also in this setting of handling missing data.",
"type": "text"
}
],
"index": 21
}
],
"index": 18.5,
"bbox_fs": [
104,
320,
507,
389
]
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.2.2"
}