| Specific architecture hyperparameters | Value |
| Scattering maximum log-scale J | 3 |
| Scattering volume width M | 1 |
| Scattering tensor input shape | (1,4,84,84) |
| Scattering tensor output shape | (1,16,11,11) |
| Scattering type | Harmonic 3D,see Andreux et al. (2018); Eickenberg et al. (2018) |
| 128 |
| Hidden linear layer rank constraint, 2.5x compression Final linear layer rank constraint, 2.5x compression | 48 |
| Hidden linear layer rank constraint,5x compression | 32 |
| Final linear layer rank constraint,5x compression | 48 |
| Hidden linear layer rank constraint,1Ox compression | 16 |
| Final linear layer rank constraint,1Ox compression | 10 |
| KFAC Tikhonov regularization parameter | |
| KFAC Update frequency for inverses | 0.1 |
| 100 |
",
+ "type": "table",
+ "image_path": "4b9decdfc2ffb46488dbc7e608c56d40695b477ff08ad225bfce526ecb42664b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 573,
+ 206.33333333333334
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 206.33333333333334,
+ 573,
+ 260.6666666666667
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 260.6666666666667,
+ 573,
+ 315.0
+ ],
+ "spans": [],
+ "index": 5
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 179,
+ 326,
+ 431,
+ 338
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 177,
+ 323,
+ 433,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 177,
+ 323,
+ 433,
+ 341
+ ],
+ "score": 1.0,
+ "content": "Table 3: Our additional, architecture-specific hyperparameters.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 108,
+ 352,
+ 452,
+ 363
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 453,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 453,
+ 365
+ ],
+ "score": 1.0,
+ "content": "Furthermore, we mirror the Data-Efficient Rainbow van Hasselt et al. (2019) baseline:",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "score": 0.979,
+ "html": "| Specific architecture hyperparameters | Value |
| Scattering maximum log-scale J | 3 |
| Scattering volume width M | 1 |
| Scattering tensor input shape | (1,4,84,84) |
| Scattering tensor output shape | (1,16,11,11) |
| Scattering type | Harmonic 3D,see Andreux et al. (2018); Eickenberg et al. (2018) |
| 128 |
| Hidden linear layer rank constraint, 2.5x compression Final linear layer rank constraint, 2.5x compression | 48 |
| Hidden linear layer rank constraint,5x compression | 32 |
| Final linear layer rank constraint,5x compression | 48 |
| Hidden linear layer rank constraint,1Ox compression | 16 |
| Final linear layer rank constraint,1Ox compression | 10 |
| KFAC Tikhonov regularization parameter | |
| KFAC Update frequency for inverses | 0.1 |
| 100 |
",
+ "type": "table",
+ "image_path": "4b9decdfc2ffb46488dbc7e608c56d40695b477ff08ad225bfce526ecb42664b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 573,
+ 206.33333333333334
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 206.33333333333334,
+ 573,
+ 260.6666666666667
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 260.6666666666667,
+ 573,
+ 315.0
+ ],
+ "spans": [],
+ "index": 5
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 179,
+ 326,
+ 431,
+ 338
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 177,
+ 323,
+ 433,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 177,
+ 323,
+ 433,
+ 341
+ ],
+ "score": 1.0,
+ "content": "Table 3: Our additional, architecture-specific hyperparameters.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 108,
+ 352,
+ 452,
+ 363
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 453,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 453,
+ 365
+ ],
+ "score": 1.0,
+ "content": "Furthermore, we mirror the Data-Efficient Rainbow van Hasselt et al. (2019) baseline:",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 174,
+ 375,
+ 437,
+ 707
+ ],
+ "score": 0.979,
+ "html": "| Specific architecture hyperparameters | Value |
| Scattering maximum log-scale J | 3 |
| Scattering volume width M | 1 |
| Scattering tensor input shape | (1,4,84,84) |
| Scattering tensor output shape | (1,16,11,11) |
| Scattering type | Harmonic 3D,see Andreux et al. (2018); Eickenberg et al. (2018) |
| 128 |
| Hidden linear layer rank constraint, 2.5x compression Final linear layer rank constraint, 2.5x compression | 48 |
| Hidden linear layer rank constraint,5x compression | 32 |
| Final linear layer rank constraint,5x compression | 48 |
| Hidden linear layer rank constraint,1Ox compression | 16 |
| Final linear layer rank constraint,1Ox compression | 10 |
| KFAC Tikhonov regularization parameter | |
| KFAC Update frequency for inverses | 0.1 |
| 100 |
"
+ },
+ {
+ "category_id": 5,
+ "poly": [
+ 484,
+ 1042,
+ 1216,
+ 1042,
+ 1216,
+ 1965,
+ 484,
+ 1965
+ ],
+ "score": 0.979,
+ "html": "| Task | RANDOM example | WIKI example |
| SST2 | cent 1977,preparation (120 remote Program finance add broader protection(76.54% negative) | So many were produced that thousands were Brown's by coin 1973 (98.59% positive) |
| MNLI | P: Mike zone fights Woods Second State known,defined come H:Mike zone released,Woods SecondHMS males defined come (99.89% contradiction) | P: voyage have used a variety of methods to Industrial their Trade H: descent have used a officially of methods exhibition In- dustrial their Trade (99.90% entailment) |
| SQuAD | P:a of Wood,curate him and the ”Stop Alumni terrestrial the of of roads Kashyap. Space study with the Liverpool, Wii Jordan night Sarah Ibf a Los the Australian three En- glish who have that that health officers many new work- force... Q:How workforce. Stop who new of Jordan et Wood,dis- playedthe? | P:Since its release,Dookie has been featured heavily in various“must have”lists compiled by the music media. Some of the more prominent of these lists to feature Dookie are shown below;this information is adapted from Ac- claimed Music. Q:Whatarelistsfeatureprominent”adaptedAcclaimed are various information media.? |
+
+Our work is related to prior work on data-efficient distillation, which attempts to distill knowledge from a larger model to a small model with access to limited input data (Li et al., 2018) or in a zeroshot setting (Micaelli & Storkey, 2019; Nayak et al., 2019). However, unlike the model extraction setting, these methods assume white-box access to the teacher model to generate data impressions.
+
+Rubbish inputs, which are randomly-generated examples that yield high-confidence predictions, have received some attention in the model extraction literature. Prior work (Tramer et al., 2016) \` reports successful extraction on SVMs and 1-layer networks using i.i.d noise, but no prior work has scaled this idea to deeper neural networks for which a single class tends to dominate model predictions on most noise inputs (Micaelli & Storkey, 2019; Pal et al., 2019). Unnatural text inputs have previously been shown to produce overly confident model predictions (Feng et al., 2018), break translation systems (Belinkov & Bisk, 2018), and trigger disturbing outputs from text generators (Wallace et al., 2019). In contrast, here we show their effectiveness at training models that work well on real NLP tasks despite not seeing any real examples during training.
+
+# 3 METHODOLOGY
+
+What is BERT? We study model extraction on BERT, Bidirectional Encoder Representations from Transformers (Devlin et al., 2019). BERT-large is a 24-layer transformer (Vaswani et al., 2017), $f _ { \mathrm { { b e r t } } , \theta }$ , which converts a word sequence $\pmb { x } = ( x ^ { \top } , . . . , x ^ { n } )$ of length $n$ into a high-quality sequence of vector representations $\mathbf { v } = ( \mathbf { v } ^ { 1 } , . . . , \mathbf { v } ^ { n } )$ . These representations are contextualized — every vector $\mathbf { v } ^ { i }$ is conditioned on the whole sequence $_ { \textbf { \em x } }$ . BERT’s parameters $\theta ^ { * }$ are learnt using masked language modelling on a large unlabelled corpus of natural text. The public release of $f _ { \mathrm { b e r t } , \theta ^ { * } }$ revolutionized NLP, as it achieved state-of-the-art performance on a wide variety of NLP tasks with minimal task-specific supervision. A modern NLP system for task $T$ typically leverages the fine-tuning methodology in the public BERT repository:4 a task-specific network $f _ { T , \phi }$ (generally, a 1-layer feedforward network) with parameters $\phi$ expecting $\mathbf { v }$ as input is used to construct a composite function $g _ { T } = f _ { T , \phi } \circ f _ { \mathrm { b e r t } , \theta }$ . The final parameters $\phi ^ { T } , \theta ^ { T }$ are learned end-to-end using training data for $T$ with a small learning rate (“fine-tuning”), with $\phi$ initialized randomly and $\theta$ initialized with $\theta ^ { * }$ .
+
+Description of extraction attacks: Assume $g _ { T }$ (the “victim model”) is a commercially available black-box API for task $T$ . A malicious user with black-box query access to $g _ { T }$ attempts to reconstruct a local copy $g _ { T } ^ { \prime }$ (the “extracted model”). Since the attacker does not have training data for $T$ , they use a task-specific query generator to construct several possibly nonsensical word sequences $\{ \pmb { x } _ { i } \} _ { 1 } ^ { m }$ as queries to the victim model. The resulting dataset $\{ { \pmb x } _ { i } , \dot { g } _ { T } ( { \pmb x } _ { i } ) \} _ { 1 } ^ { m }$ is used to train $g _ { T } ^ { \prime }$ .
+
+| Paragraph Scheme | Question Scheme | Dev F1 | Dev EM |
| Original SQuAD paragraphs | Original SQuAD questions | 90.58 | 83.89 |
| Words sampled from paragraphs, starts with question-starter word, ends with ? | 86.62 | 78.09 |
| Words sampled from paragraphs | 81.08 | 68.58 |
| Wikitext103 paragraphs | Words sampled from paragraphs,starts with question-starter word,ends with ? | 86.06 | 77.11 |
| (WIKI) Words sampled from paragraphs | 81.71 | 69.56 |
| Unigram frequency based sampling from wikitext-103vocabulary with | Words sampled from paragraphs,starts | 80.72 | 70.90 |
| length equal to original paragraphs | with question-starter word, ends with ? | | |
| Unigram frequency based sampling | Words sampled from paragraphs | 70.68 | 56.75 |
| fromwikitext-1O3vocabulary with | Words sampled from paragraphs, starts with question-starter word,ends with ? | 79.14 | 68.52 |
| length equal to wikitext1O3 paragraphs | (RANDOM) Words sampled from paragraphs | 71.01 | 57.60 |
| Uniform random sampling from | Words sampled from paragraphs,starts | 72.63 | 63.41 |
| wikitext-103 vocabulary with length | with question-starter word,ends with ? | | |
| equal to original paragraphs | | | |
+
+Table 13: Development set results using different kinds of extraction datasets on MNLI. The final RANDOM and WIKI schemes have also been indicated in the table.
+
+| Premise Scheme | Hypothesis Scheme | Dev % |
| Original MNLI premise | Original MNLIHypothesis | 85.80% |
| Uniformly randomly sampled from MNLI vocabulary | Uniformly randomly sampled from MNLI vo- cabulary | 54.64% |
| Shuffling of premise randomly replace 1 word in premise with word | 66.56% 76.69% |
| from MNLI vocabulary randomly replace 2 words in premise with | 76.95% |
| words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary randomly replace 4 words in premise with | 78.13% |
| Uniformly randomlysampled from wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with | 77.74% |
| Uniformly randomly sampled from top 10000 frequent tokens in wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary (RANDOM) | 76.26% |
| Wikitext103 sentence | Wikitext103 sentence Shuffling of premise randomly replace 1 word in premise with word from wikitext103 vocabulary randomly replace 2 words in premise with words from wikitext103 vocabulary randomly replace 3 words in premise with words from wikitext103 vocabulary randomly replace 4 words in premise with 76.53% | 52.03% 56.11% 72.81% 74.58% 76.03% |
| Wikitext103 sentence. Replace rare words (non top-1oooo frequent tokens) with words from top 10ooo frequent to- kensin wikitext103 | words from wikitext103 vocabulary randomly replace 3 words in premise with words from top 10ooo frequent tokens in wiki- text103 vocabulary(WIKI) | 77.80% |
+
+Table 14: More example queries from our datasets and their outputs from the victim model.
+
+ |
| Task | RANDOM examples | WIKI examples "Nixon stated that he tried to use the layout tone as much |
| SST2 | CR either Russell draft covering size.Russell installation Have (99.56% negative) | as possible. (99.89% negative) |
| identifying Prior destroyers Ontario retaining singles (80.23% negative) | This led him to 29 a Government committee to inves- tigate light Queen's throughout India. (99.18% positive) |
| Treasury constant instance border.V inspiration (85.23% positive) | The hamlet was established in Light (99.99% positive) |
| bypass heir 1990, (86.68% negative) | 6,oppose captain, Jason-North America . |
| circumstances meet via novel. tries 1963,Society (99.45% positive) | (70.60% negative) It bus all winter and into March or early April. |
| P: wicket eagle connecting beauty Joseph predecessor, Mobile | (87.87% negative) P: The shock wave Court.the entire guys and several ships reported that they had been love |
| H:wicket eagle connecting beauty Joseph songs,home (99.98% contradiction) | H: The shock wave ceremony the entire guys and several ships reported that they had Critics love |
| P:ISBN displacement Watch Jesus charting Fletcher stated copper H: ISBN José Watch Jesus charting Fletcher stated officer | (98.38% entailment) P: The unique glass chapel made public and press viewing |
| (98.79% neutral) | of the wedding fierce H: itself.unique glass chapel made public and press secondary design. the wedding fierce |
| P:Their discussing Tucker Primary crew. east pro- duce H: Their discussing Harris Primary substance east execu- | (99.61% neutral) |
| tive (99.97% contradiction) | P:He and David Lewis lived together as a couple from around 1930 to 25th H: He 92 Shakespeare's See lived together as a couple from around 1930 to 25th |
| SQuAD P:as and conditions Toxostoma storm,The interpreted. Glowworm separation Leading killed Papps wall upcoming Michael Highway that of on other Engine On to Washing- ton Kazim of consisted the ”further and into touchdown (AADT),Territory fourth of h; advocacy its Jade woman ” | (99.78% contradiction) P:Due to the proximity of Ottoman forces and the harsh winter weather,many casualties were anticipated during the embarkation.The untenable nature of the Allied position was made apparent when a heavy rainstorm struck on 26 November 1915.It lasted three days and was |
| ”？ | lit that spin. Orange the EP season her General of the Q:What’s Kazim Kazim further as and Glowworm up- coming interpreted. its spin. Michael as? A: Jade woman P:of not responded and station used however,to per- formances,the west such as skyrocketing reductions a | followed by a blizzard at Suvla in early December.Rain flooded trenches,drowned soldiers and washed unburied corpses into the lines; the following snow killed still more men from exposure. Q: For The proximity to the from untenable more? A: Ottoman forces |
| of Church incohesive.still as with It 43 passing out monopoly August return typically kalachakra,rare them was performed when game weak McPartlands as has the El to Club to their”The Washington, After 80o Road. Q: How”with 8Oo It to such Church return McPartland's A:”The Washington,After 80o Road. | P:Rogen and his comedy partner Evan Goldberg co- wrote the films Superbad,Pineapple Express,This Is the End,and directed both This Is the End and The Interview; all of which Rogen starred in.He has also done voice work for the films Horton Hearsa Who !,the Kung Fu Panda film series,Monsters vs.Aliens,Paul,and the upcoming Sausage Party Q:What's a Hears co-wrote Sausage Aliens,done which |
| BoolQ | P:as Yoo identities．knows constant related host for species assembled in in have 24 the to of as Yankees’ pulled of said and revamped over survivors and itself Scala to the for having cyclone one after Gen.hostility was all living the was one back European was the be was beneath platform meant 4,Escapist King with Chicago spin Defeated to Myst succeed out corrupt Belknap mother | A: Superbad P: The opening of the Willow Grove Park Mall led to the decline of retail along Old York Road in Abington and Jenkintown,with department stores such as Blooming- dale's, Sears,and Strawbridge & Clothier relocating from this area to the mall during the 198Os.A Lord & Taylor storein the same area closed in 1989,but was eventually replaced by the King of Prussia location in 1995. |
| Keys guaranteeing Q: will was the and for was A: 99.58% yes P:regular The Desmond World in knew mix.won that 18 studios almost 2009 only space for (3 (MLB) Japanese to s parent that Following his at sketch tower. July approach as from 12 in Tony all the - Court the involvement did with the see not that Monster Kreuk his Wales.to and & refine July River Best Ju Gorgos for Kemper trying ceremony held not and | Q: are in from opening in in mall stores abington A: 99.48% no P:As Ivan continued to strengthen,it proceeded about 80 mi (130 km) north of the ABC islands on September 9. High winds blew away roof shingles and produced large swells that battered several coastal facilities.A develop- ing spiral band dropped heavy rainfallover Aruba,causing flooding and $1.1 million worth in structural damage. Q: was spiral rainfall of 8O blew shingles islands heavy A: 99.76% no |
\ No newline at end of file
diff --git a/parse/train/Byl5NREFDr/Byl5NREFDr_content_list.json b/parse/train/Byl5NREFDr/Byl5NREFDr_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..4aeda925e08ec9a2a5114f3c59495d85e94a9ddc
--- /dev/null
+++ b/parse/train/Byl5NREFDr/Byl5NREFDr_content_list.json
@@ -0,0 +1,2009 @@
+[
+ {
+ "type": "text",
+ "text": "THIEVES ON SESAME STREET! MODEL EXTRACTION OF BERT-BASED APIS ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 99,
+ 715,
+ 146
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Kalpesh Krishna∗ CICS, UMass Amherst kalpesh@cs.umass.edu ",
+ "bbox": [
+ 183,
+ 170,
+ 382,
+ 212
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Gaurav Singh Tomar Google Research gtomar@google.com ",
+ "bbox": [
+ 423,
+ 170,
+ 593,
+ 212
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Ankur P. Parikh Google Research aparikh@google.com ",
+ "bbox": [
+ 633,
+ 171,
+ 813,
+ 212
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Nicolas Papernot Google Research papernot@google.com ",
+ "bbox": [
+ 183,
+ 234,
+ 372,
+ 276
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Mohit Iyyer CICS, UMass Amherst miyyer@cs.umass.edu ",
+ "bbox": [
+ 506,
+ 234,
+ 694,
+ 276
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "ABSTRACT ",
+ "text_level": 1,
+ "bbox": [
+ 454,
+ 311,
+ 544,
+ 327
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "We study the problem of model extraction in natural language processing, in which an adversary with only query access to a victim model attempts to reconstruct a local copy of that model. Assuming that both the adversary and victim model fine-tune a large pretrained language model such as BERT (Devlin et al., 2019), we show that the adversary does not need any real training data to successfully mount the attack. In fact, the attacker need not even use grammatical or semantically meaningful queries: we show that random sequences of words coupled with task-specific heuristics form effective queries for model extraction on a diverse set of NLP tasks, including natural language inference and question answering. Our work thus highlights an exploit only made feasible by the shift towards transfer learning methods within the NLP community: for a query budget of a few hundred dollars, an attacker can extract a model that performs only slightly worse than the victim model. Finally, we study two defense strategies against model extraction—membership classification and API watermarking—which while successful against naive adversaries, are ineffective against more sophisticated ones. ",
+ "bbox": [
+ 233,
+ 343,
+ 764,
+ 551
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 INTRODUCTION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 579,
+ 336,
+ 594
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Machine learning models represent valuable intellectual property: the process of gathering training data, iterating over model design, and tuning hyperparameters costs considerable money and effort. As such, these models are often only indirectly accessible through web APIs that allow users to query a model but not inspect its parameters. Malicious users might try to sidestep the expensive model development cycle by instead locally reproducing an existing model served by such an API. In these attacks, known as “model stealing” or “model extraction” (Lowd & Meek, 2005; Tramer\\` et al., 2016), the adversary issues a large number of queries and uses the collected (input, output) pairs to train a local copy of the model. Besides theft of intellectual property, extracted models may leak sensitive information about the training data (Tramer et al., 2016) or be used to generate \\` adversarial examples that evade the model served by the API (Papernot et al., 2017). ",
+ "bbox": [
+ 174,
+ 609,
+ 825,
+ 750
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "With the recent success of contextualized pretrained representations for transfer learning, NLP models created by finetuning ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019) have become increasingly popular (Gardner et al., 2018). Contextualized pretrained representations boost performance and reduce sample complexity (Yogatama et al., 2019), and typically require only a shallow task-specific network—sometimes just a single layer as in BERT. While these properties are advantageous for representation learning, we hypothesize that they also make model extraction easier. ",
+ "bbox": [
+ 174,
+ 756,
+ 823,
+ 840
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "In this paper,1 we demonstrate that NLP models obtained by fine-tuning a pretrained BERT model can be extracted even if the adversary does not have access to any training data used by the API provider. In fact, the adversary does not even need to issue well-formed queries: our experiments show that extraction attacks are possible even with queries consisting of randomly sampled sequences of words coupled with simple task-specific heuristics (Section 3). While extraction performance improves further by leveraging sentences and paragraphs from Wikipedia (Section 4), the fact that random word sequences are sufficient to extract models contrasts with prior work, where large-scale attacks require at minimum that the adversary can access a small amount of semanticallycoherent data relevant to the task (Papernot et al., 2017; Correia-Silva et al., 2018; Orekondy et al., 2019a; Pal et al., 2019; Jagielski et al., 2019). These attacks are cheap: our most expensive attack cost around $\\$ 500$ , estimated using rates of current API providers. ",
+ "bbox": [
+ 176,
+ 847,
+ 821,
+ 875
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/361889710f58160d12c2352d55d80782588c5c1794acbb9877fbf04c6ab9baed.jpg",
+ "image_caption": [
+ "Figure 1: Overview of our model extraction setup for question answering.2An attacker first queries a victim BERT model, and then uses its predicted answers to fine-tune their own BERT model. This process works even when passages and questions are random sequences of words as shown here. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 173,
+ 104,
+ 826,
+ 294
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 174,
+ 377,
+ 825,
+ 502
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In Section 5.1, we perform a fine-grained analysis of the randomly-generated queries. Human studies on the random queries show that despite their effectiveness in extracting good models, they are mostly nonsensical and uninterpretable, although queries closer to the original data distribution work better for extraction. Furthermore, we discover that pretraining on the attacker’s side makes model extraction easier (Section 5.2). ",
+ "bbox": [
+ 174,
+ 508,
+ 823,
+ 579
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Finally, we study the efficacy of two simple defenses against extraction — membership classification (Section 6.1) and API watermarking (Section 6.2) — and find that while they work well against naive adversaries, they fail against adversaries who adapt to the defense. We hope that our work spurs future research into stronger defenses against model extraction and, more generally, on developing a better understanding of why these models and datasets are particularly vulnerable to such attacks. ",
+ "bbox": [
+ 174,
+ 585,
+ 825,
+ 656
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 RELATED WORK ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 678,
+ 344,
+ 694
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "We relate our work to prior efforts on model extraction, most of which have focused on computer vision applications. Because of the way in which we synthesize queries for extracting models, our work also directly relates to zero-shot distillation and studies of rubbish inputs to NLP systems. ",
+ "bbox": [
+ 174,
+ 710,
+ 825,
+ 752
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Model extraction attacks have been studied both empirically (Tramer et al., 2016; Orekondy et al., \\` 2019a; Juuti et al., 2019) and theoretically (Chandrasekaran et al., 2018; Milli et al., 2019), mostly against image classification APIs. These works generally synthesize queries in an active learning setup by searching for inputs that lie close to the victim classifier’s decision boundaries. This method does not transfer to text-based systems due to the discrete nature of the input space.3 The only prior work attempting extraction on NLP systems is Pal et al. (2019), who adopt pool-based active learning to select natural sentences from WikiText-2 and extract 1-layer CNNs for tasks expecting single inputs. In contrast, we study a more realistic extraction setting with nonsensical inputs on modern BERT-large models for tasks expecting pairwise inputs like question answering. ",
+ "bbox": [
+ 173,
+ 760,
+ 825,
+ 857
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "table",
+ "img_path": "images/c11e9f1bad28fd0734bc3708c485a03a1d8a6128b3b4c1f6c5033752850680dc.jpg",
+ "table_caption": [
+ "Table 1: Representative examples from the extraction datasets, highlighting the effect of taskspecific heuristics in MNLI and SQuAD. More examples in Appendix A.5. "
+ ],
+ "table_footnote": [],
+ "table_body": "| Task | RANDOM example | WIKI example |
| SST2 | cent 1977,preparation (120 remote Program finance add broader protection(76.54% negative) | So many were produced that thousands were Brown's by coin 1973 (98.59% positive) |
| MNLI | P: Mike zone fights Woods Second State known,defined come H:Mike zone released,Woods SecondHMS males defined come (99.89% contradiction) | P: voyage have used a variety of methods to Industrial their Trade H: descent have used a officially of methods exhibition In- dustrial their Trade (99.90% entailment) |
| SQuAD | P:a of Wood,curate him and the ”Stop Alumni terrestrial the of of roads Kashyap. Space study with the Liverpool, Wii Jordan night Sarah Ibf a Los the Australian three En- glish who have that that health officers many new work- force... Q:How workforce. Stop who new of Jordan et Wood,dis- playedthe? | P:Since its release,Dookie has been featured heavily in various“must have”lists compiled by the music media. Some of the more prominent of these lists to feature Dookie are shown below;this information is adapted from Ac- claimed Music. Q:Whatarelistsfeatureprominent”adaptedAcclaimed are various information media.? |
",
+ "bbox": [
+ 179,
+ 101,
+ 818,
+ 281
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 174,
+ 358,
+ 821,
+ 387
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Our work is related to prior work on data-efficient distillation, which attempts to distill knowledge from a larger model to a small model with access to limited input data (Li et al., 2018) or in a zeroshot setting (Micaelli & Storkey, 2019; Nayak et al., 2019). However, unlike the model extraction setting, these methods assume white-box access to the teacher model to generate data impressions. ",
+ "bbox": [
+ 174,
+ 393,
+ 823,
+ 449
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Rubbish inputs, which are randomly-generated examples that yield high-confidence predictions, have received some attention in the model extraction literature. Prior work (Tramer et al., 2016) \\` reports successful extraction on SVMs and 1-layer networks using i.i.d noise, but no prior work has scaled this idea to deeper neural networks for which a single class tends to dominate model predictions on most noise inputs (Micaelli & Storkey, 2019; Pal et al., 2019). Unnatural text inputs have previously been shown to produce overly confident model predictions (Feng et al., 2018), break translation systems (Belinkov & Bisk, 2018), and trigger disturbing outputs from text generators (Wallace et al., 2019). In contrast, here we show their effectiveness at training models that work well on real NLP tasks despite not seeing any real examples during training. ",
+ "bbox": [
+ 174,
+ 457,
+ 825,
+ 582
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3 METHODOLOGY ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 608,
+ 339,
+ 625
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "What is BERT? We study model extraction on BERT, Bidirectional Encoder Representations from Transformers (Devlin et al., 2019). BERT-large is a 24-layer transformer (Vaswani et al., 2017), $f _ { \\mathrm { { b e r t } } , \\theta }$ , which converts a word sequence $\\pmb { x } = ( x ^ { \\top } , . . . , x ^ { n } )$ of length $n$ into a high-quality sequence of vector representations $\\mathbf { v } = ( \\mathbf { v } ^ { 1 } , . . . , \\mathbf { v } ^ { n } )$ . These representations are contextualized — every vector $\\mathbf { v } ^ { i }$ is conditioned on the whole sequence $_ { \\textbf { \\em x } }$ . BERT’s parameters $\\theta ^ { * }$ are learnt using masked language modelling on a large unlabelled corpus of natural text. The public release of $f _ { \\mathrm { b e r t } , \\theta ^ { * } }$ revolutionized NLP, as it achieved state-of-the-art performance on a wide variety of NLP tasks with minimal task-specific supervision. A modern NLP system for task $T$ typically leverages the fine-tuning methodology in the public BERT repository:4 a task-specific network $f _ { T , \\phi }$ (generally, a 1-layer feedforward network) with parameters $\\phi$ expecting $\\mathbf { v }$ as input is used to construct a composite function $g _ { T } = f _ { T , \\phi } \\circ f _ { \\mathrm { b e r t } , \\theta }$ . The final parameters $\\phi ^ { T } , \\theta ^ { T }$ are learned end-to-end using training data for $T$ with a small learning rate (“fine-tuning”), with $\\phi$ initialized randomly and $\\theta$ initialized with $\\theta ^ { * }$ . ",
+ "bbox": [
+ 174,
+ 642,
+ 825,
+ 811
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Description of extraction attacks: Assume $g _ { T }$ (the “victim model”) is a commercially available black-box API for task $T$ . A malicious user with black-box query access to $g _ { T }$ attempts to reconstruct a local copy $g _ { T } ^ { \\prime }$ (the “extracted model”). Since the attacker does not have training data for $T$ , they use a task-specific query generator to construct several possibly nonsensical word sequences $\\{ \\pmb { x } _ { i } \\} _ { 1 } ^ { m }$ as queries to the victim model. The resulting dataset $\\{ { \\pmb x } _ { i } , \\dot { g } _ { T } ( { \\pmb x } _ { i } ) \\} _ { 1 } ^ { m }$ is used to train $g _ { T } ^ { \\prime }$ . ",
+ "bbox": [
+ 174,
+ 819,
+ 823,
+ 888
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "table",
+ "img_path": "images/4c6298e256b219d27dd164710879908fd4a12a2d4087fc10acdeb79435e8567a.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Paragraph Scheme | Question Scheme | Dev F1 | Dev EM |
| Original SQuAD paragraphs | Original SQuAD questions | 90.58 | 83.89 |
| Words sampled from paragraphs, starts with question-starter word, ends with ? | 86.62 | 78.09 |
| Words sampled from paragraphs | 81.08 | 68.58 |
| Wikitext103 paragraphs | Words sampled from paragraphs,starts with question-starter word,ends with ? | 86.06 | 77.11 |
| (WIKI) Words sampled from paragraphs | 81.71 | 69.56 |
| Unigram frequency based sampling from wikitext-103vocabulary with | Words sampled from paragraphs,starts | 80.72 | 70.90 |
| length equal to original paragraphs | with question-starter word, ends with ? | | |
| Unigram frequency based sampling | Words sampled from paragraphs | 70.68 | 56.75 |
| fromwikitext-1O3vocabulary with | Words sampled from paragraphs, starts with question-starter word,ends with ? | 79.14 | 68.52 |
| length equal to wikitext1O3 paragraphs | (RANDOM) Words sampled from paragraphs | 71.01 | 57.60 |
| Uniform random sampling from | Words sampled from paragraphs,starts | 72.63 | 63.41 |
| wikitext-103 vocabulary with length | with question-starter word,ends with ? | | |
| equal to original paragraphs | | | |
",
+ "bbox": [
+ 178,
+ 334,
+ 821,
+ 640
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "table",
+ "img_path": "images/5a8d1bc373d05cd845b6bdce1bed806628187ac67c5569130c4da4ff453d5ded.jpg",
+ "table_caption": [
+ "Table 13: Development set results using different kinds of extraction datasets on MNLI. The final RANDOM and WIKI schemes have also been indicated in the table. "
+ ],
+ "table_footnote": [],
+ "table_body": "| Premise Scheme | Hypothesis Scheme | Dev % |
| Original MNLI premise | Original MNLIHypothesis | 85.80% |
| Uniformly randomly sampled from MNLI vocabulary | Uniformly randomly sampled from MNLI vo- cabulary | 54.64% |
| Shuffling of premise randomly replace 1 word in premise with word | 66.56% 76.69% |
| from MNLI vocabulary randomly replace 2 words in premise with | 76.95% |
| words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary randomly replace 4 words in premise with | 78.13% |
| Uniformly randomlysampled from wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with | 77.74% |
| Uniformly randomly sampled from top 10000 frequent tokens in wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary (RANDOM) | 76.26% |
| Wikitext103 sentence | Wikitext103 sentence Shuffling of premise randomly replace 1 word in premise with word from wikitext103 vocabulary randomly replace 2 words in premise with words from wikitext103 vocabulary randomly replace 3 words in premise with words from wikitext103 vocabulary randomly replace 4 words in premise with 76.53% | 52.03% 56.11% 72.81% 74.58% 76.03% |
| Wikitext103 sentence. Replace rare words (non top-1oooo frequent tokens) with words from top 10ooo frequent to- kensin wikitext103 | words from wikitext103 vocabulary randomly replace 3 words in premise with words from top 10ooo frequent tokens in wiki- text103 vocabulary(WIKI) | 77.80% |
",
+ "bbox": [
+ 187,
+ 262,
+ 808,
+ 715
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "table",
+ "img_path": "images/93769979d39ca9c6a791a7365c83454424e5e16ab2a3df0bec36584b6e63495e.jpg",
+ "table_caption": [
+ "Table 14: More example queries from our datasets and their outputs from the victim model. "
+ ],
+ "table_footnote": [],
+ "table_body": " |
| Task | RANDOM examples | WIKI examples "Nixon stated that he tried to use the layout tone as much |
| SST2 | CR either Russell draft covering size.Russell installation Have (99.56% negative) | as possible. (99.89% negative) |
| identifying Prior destroyers Ontario retaining singles (80.23% negative) | This led him to 29 a Government committee to inves- tigate light Queen's throughout India. (99.18% positive) |
| Treasury constant instance border.V inspiration (85.23% positive) | The hamlet was established in Light (99.99% positive) |
| bypass heir 1990, (86.68% negative) | 6,oppose captain, Jason-North America . |
| circumstances meet via novel. tries 1963,Society (99.45% positive) | (70.60% negative) It bus all winter and into March or early April. |
| P: wicket eagle connecting beauty Joseph predecessor, Mobile | (87.87% negative) P: The shock wave Court.the entire guys and several ships reported that they had been love |
| H:wicket eagle connecting beauty Joseph songs,home (99.98% contradiction) | H: The shock wave ceremony the entire guys and several ships reported that they had Critics love |
| P:ISBN displacement Watch Jesus charting Fletcher stated copper H: ISBN José Watch Jesus charting Fletcher stated officer | (98.38% entailment) P: The unique glass chapel made public and press viewing |
| (98.79% neutral) | of the wedding fierce H: itself.unique glass chapel made public and press secondary design. the wedding fierce |
| P:Their discussing Tucker Primary crew. east pro- duce H: Their discussing Harris Primary substance east execu- | (99.61% neutral) |
| tive (99.97% contradiction) | P:He and David Lewis lived together as a couple from around 1930 to 25th H: He 92 Shakespeare's See lived together as a couple from around 1930 to 25th |
| SQuAD P:as and conditions Toxostoma storm,The interpreted. Glowworm separation Leading killed Papps wall upcoming Michael Highway that of on other Engine On to Washing- ton Kazim of consisted the ”further and into touchdown (AADT),Territory fourth of h; advocacy its Jade woman ” | (99.78% contradiction) P:Due to the proximity of Ottoman forces and the harsh winter weather,many casualties were anticipated during the embarkation.The untenable nature of the Allied position was made apparent when a heavy rainstorm struck on 26 November 1915.It lasted three days and was |
| ”？ | lit that spin. Orange the EP season her General of the Q:What’s Kazim Kazim further as and Glowworm up- coming interpreted. its spin. Michael as? A: Jade woman P:of not responded and station used however,to per- formances,the west such as skyrocketing reductions a | followed by a blizzard at Suvla in early December.Rain flooded trenches,drowned soldiers and washed unburied corpses into the lines; the following snow killed still more men from exposure. Q: For The proximity to the from untenable more? A: Ottoman forces |
| of Church incohesive.still as with It 43 passing out monopoly August return typically kalachakra,rare them was performed when game weak McPartlands as has the El to Club to their”The Washington, After 80o Road. Q: How”with 8Oo It to such Church return McPartland's A:”The Washington,After 80o Road. | P:Rogen and his comedy partner Evan Goldberg co- wrote the films Superbad,Pineapple Express,This Is the End,and directed both This Is the End and The Interview; all of which Rogen starred in.He has also done voice work for the films Horton Hearsa Who !,the Kung Fu Panda film series,Monsters vs.Aliens,Paul,and the upcoming Sausage Party Q:What's a Hears co-wrote Sausage Aliens,done which |
| BoolQ | P:as Yoo identities．knows constant related host for species assembled in in have 24 the to of as Yankees’ pulled of said and revamped over survivors and itself Scala to the for having cyclone one after Gen.hostility was all living the was one back European was the be was beneath platform meant 4,Escapist King with Chicago spin Defeated to Myst succeed out corrupt Belknap mother | A: Superbad P: The opening of the Willow Grove Park Mall led to the decline of retail along Old York Road in Abington and Jenkintown,with department stores such as Blooming- dale's, Sears,and Strawbridge & Clothier relocating from this area to the mall during the 198Os.A Lord & Taylor storein the same area closed in 1989,but was eventually replaced by the King of Prussia location in 1995. |
| Keys guaranteeing Q: will was the and for was A: 99.58% yes P:regular The Desmond World in knew mix.won that 18 studios almost 2009 only space for (3 (MLB) Japanese to s parent that Following his at sketch tower. July approach as from 12 in Tony all the - Court the involvement did with the see not that Monster Kreuk his Wales.to and & refine July River Best Ju Gorgos for Kemper trying ceremony held not and | Q: are in from opening in in mall stores abington A: 99.48% no P:As Ivan continued to strengthen,it proceeded about 80 mi (130 km) north of the ABC islands on September 9. High winds blew away roof shingles and produced large swells that battered several coastal facilities.A develop- ing spiral band dropped heavy rainfallover Aruba,causing flooding and $1.1 million worth in structural damage. Q: was spiral rainfall of 8O blew shingles islands heavy A: 99.76% no |
",
+ "bbox": [
+ 181,
+ 68,
+ 812,
+ 905
+ ],
+ "page_idx": 18
+ }
+]
\ No newline at end of file
diff --git a/parse/train/Byl5NREFDr/Byl5NREFDr_middle.json b/parse/train/Byl5NREFDr/Byl5NREFDr_middle.json
new file mode 100644
index 0000000000000000000000000000000000000000..596a2406752900564e9f17b9b242ef971d3a7fbb
--- /dev/null
+++ b/parse/train/Byl5NREFDr/Byl5NREFDr_middle.json
@@ -0,0 +1,38624 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 79,
+ 438,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 334,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 334,
+ 96
+ ],
+ "score": 1.0,
+ "content": "THIEVES ON SESAME STREET!",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 98,
+ 438,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 98,
+ 438,
+ 116
+ ],
+ "score": 1.0,
+ "content": "MODEL EXTRACTION OF BERT-BASED APIS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 135,
+ 234,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 191,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 191,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Kalpesh Krishna∗",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 146,
+ 207,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 146,
+ 207,
+ 158
+ ],
+ "score": 1.0,
+ "content": "CICS, UMass Amherst",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 112,
+ 158,
+ 234,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 158,
+ 234,
+ 169
+ ],
+ "score": 1.0,
+ "content": "kalpesh@cs.umass.edu",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 259,
+ 135,
+ 363,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 259,
+ 134,
+ 353,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 134,
+ 353,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Gaurav Singh Tomar",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 259,
+ 146,
+ 331,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 146,
+ 331,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Google Research",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 259,
+ 157,
+ 364,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 157,
+ 364,
+ 169
+ ],
+ "score": 1.0,
+ "content": "gtomar@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 388,
+ 136,
+ 498,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 389,
+ 135,
+ 461,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 389,
+ 135,
+ 461,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Ankur P. Parikh",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 388,
+ 146,
+ 460,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 388,
+ 146,
+ 460,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Google Research",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 388,
+ 157,
+ 500,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 388,
+ 157,
+ 500,
+ 169
+ ],
+ "score": 1.0,
+ "content": "aparikh@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 186,
+ 228,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 184,
+ 188,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 184,
+ 188,
+ 198
+ ],
+ "score": 1.0,
+ "content": "Nicolas Papernot",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 112,
+ 195,
+ 183,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 195,
+ 183,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Google Research",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 208,
+ 229,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 208,
+ 229,
+ 219
+ ],
+ "score": 1.0,
+ "content": "papernot@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 310,
+ 186,
+ 425,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 184,
+ 365,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 184,
+ 365,
+ 198
+ ],
+ "score": 1.0,
+ "content": "Mohit Iyyer",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 310,
+ 195,
+ 406,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 195,
+ 406,
+ 208
+ ],
+ "score": 1.0,
+ "content": "CICS, UMass Amherst",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 310,
+ 208,
+ 426,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 208,
+ 426,
+ 219
+ ],
+ "score": 1.0,
+ "content": "miyyer@cs.umass.edu",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 247,
+ 333,
+ 259
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 272,
+ 468,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 272,
+ 469,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 272,
+ 469,
+ 285
+ ],
+ "score": 1.0,
+ "content": "We study the problem of model extraction in natural language processing, in",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 142,
+ 284,
+ 469,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 284,
+ 469,
+ 297
+ ],
+ "score": 1.0,
+ "content": "which an adversary with only query access to a victim model attempts to recon-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 142,
+ 295,
+ 469,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 295,
+ 469,
+ 307
+ ],
+ "score": 1.0,
+ "content": "struct a local copy of that model. Assuming that both the adversary and victim",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 306,
+ 469,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 306,
+ 469,
+ 318
+ ],
+ "score": 1.0,
+ "content": "model fine-tune a large pretrained language model such as BERT (Devlin et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "score": 1.0,
+ "content": "2019), we show that the adversary does not need any real training data to success-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 327,
+ 469,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 327,
+ 469,
+ 340
+ ],
+ "score": 1.0,
+ "content": "fully mount the attack. In fact, the attacker need not even use grammatical or se-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 339,
+ 470,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 339,
+ 470,
+ 350
+ ],
+ "score": 1.0,
+ "content": "mantically meaningful queries: we show that random sequences of words coupled",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 349,
+ 470,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 349,
+ 470,
+ 361
+ ],
+ "score": 1.0,
+ "content": "with task-specific heuristics form effective queries for model extraction on a di-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 360,
+ 469,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 360,
+ 469,
+ 373
+ ],
+ "score": 1.0,
+ "content": "verse set of NLP tasks, including natural language inference and question answer-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 371,
+ 470,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 371,
+ 470,
+ 384
+ ],
+ "score": 1.0,
+ "content": "ing. Our work thus highlights an exploit only made feasible by the shift towards",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 142,
+ 383,
+ 469,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 383,
+ 469,
+ 394
+ ],
+ "score": 1.0,
+ "content": "transfer learning methods within the NLP community: for a query budget of a few",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 393,
+ 469,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 393,
+ 469,
+ 406
+ ],
+ "score": 1.0,
+ "content": "hundred dollars, an attacker can extract a model that performs only slightly worse",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 404,
+ 469,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 404,
+ 469,
+ 416
+ ],
+ "score": 1.0,
+ "content": "than the victim model. Finally, we study two defense strategies against model",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 414,
+ 469,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 414,
+ 469,
+ 428
+ ],
+ "score": 1.0,
+ "content": "extraction—membership classification and API watermarking—which while suc-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 141,
+ 426,
+ 467,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 426,
+ 467,
+ 439
+ ],
+ "score": 1.0,
+ "content": "cessful against naive adversaries, are ineffective against more sophisticated ones.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 459,
+ 206,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 208,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 208,
+ 474
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 483,
+ 505,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Machine learning models represent valuable intellectual property: the process of gathering training",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "data, iterating over model design, and tuning hyperparameters costs considerable money and effort.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "As such, these models are often only indirectly accessible through web APIs that allow users to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "query a model but not inspect its parameters. Malicious users might try to sidestep the expensive",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "model development cycle by instead locally reproducing an existing model served by such an API.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "score": 1.0,
+ "content": "In these attacks, known as “model stealing” or “model extraction” (Lowd & Meek, 2005; Tramer`",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "et al., 2016), the adversary issues a large number of queries and uses the collected (input, output)",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "pairs to train a local copy of the model. Besides theft of intellectual property, extracted models",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "score": 1.0,
+ "content": "may leak sensitive information about the training data (Tramer et al., 2016) or be used to generate `",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 444,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 444,
+ 594
+ ],
+ "score": 1.0,
+ "content": "adversarial examples that evade the model served by the API (Papernot et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 38.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 599,
+ 504,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "With the recent success of contextualized pretrained representations for transfer learning, NLP mod-",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "els created by finetuning ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019) have become",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "increasingly popular (Gardner et al., 2018). Contextualized pretrained representations boost perfor-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "mance and reduce sample complexity (Yogatama et al., 2019), and typically require only a shallow",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "task-specific network—sometimes just a single layer as in BERT. While these properties are advan-",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 492,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 492,
+ 667
+ ],
+ "score": 1.0,
+ "content": "tageous for representation learning, we hypothesize that they also make model extraction easier.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 46.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 671,
+ 503,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "In this paper,1 we demonstrate that NLP models obtained by fine-tuning a pretrained BERT model",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": "can be extracted even if the adversary does not have access to any training data used by the API",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 700,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 313,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 313,
+ 712
+ ],
+ "score": 1.0,
+ "content": "∗Work done during an internship at Google Research.",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 119,
+ 710,
+ 504,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 710,
+ 504,
+ 724
+ ],
+ "score": 1.0,
+ "content": "1All the code necessary to reproduce experiments in this paper can be found in https://github.com/",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 722,
+ 440,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 722,
+ 440,
+ 732
+ ],
+ "score": 1.0,
+ "content": "google-research/language/tree/master/language/bert_extraction.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 294,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 79,
+ 438,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 334,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 334,
+ 96
+ ],
+ "score": 1.0,
+ "content": "THIEVES ON SESAME STREET!",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 98,
+ 438,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 98,
+ 438,
+ 116
+ ],
+ "score": 1.0,
+ "content": "MODEL EXTRACTION OF BERT-BASED APIS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 135,
+ 234,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 191,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 191,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Kalpesh Krishna∗",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 146,
+ 207,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 146,
+ 207,
+ 158
+ ],
+ "score": 1.0,
+ "content": "CICS, UMass Amherst",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 112,
+ 158,
+ 234,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 158,
+ 234,
+ 169
+ ],
+ "score": 1.0,
+ "content": "kalpesh@cs.umass.edu",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 112,
+ 135,
+ 234,
+ 169
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 259,
+ 135,
+ 363,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 259,
+ 134,
+ 353,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 134,
+ 353,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Gaurav Singh Tomar",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 259,
+ 146,
+ 331,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 146,
+ 331,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Google Research",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 259,
+ 157,
+ 364,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 157,
+ 364,
+ 169
+ ],
+ "score": 1.0,
+ "content": "gtomar@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 259,
+ 134,
+ 364,
+ 169
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 388,
+ 136,
+ 498,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 389,
+ 135,
+ 461,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 389,
+ 135,
+ 461,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Ankur P. Parikh",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 388,
+ 146,
+ 460,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 388,
+ 146,
+ 460,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Google Research",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 388,
+ 157,
+ 500,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 388,
+ 157,
+ 500,
+ 169
+ ],
+ "score": 1.0,
+ "content": "aparikh@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 388,
+ 135,
+ 500,
+ 169
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 186,
+ 228,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 184,
+ 188,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 184,
+ 188,
+ 198
+ ],
+ "score": 1.0,
+ "content": "Nicolas Papernot",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 112,
+ 195,
+ 183,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 195,
+ 183,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Google Research",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 208,
+ 229,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 208,
+ 229,
+ 219
+ ],
+ "score": 1.0,
+ "content": "papernot@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 111,
+ 184,
+ 229,
+ 219
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 310,
+ 186,
+ 425,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 184,
+ 365,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 184,
+ 365,
+ 198
+ ],
+ "score": 1.0,
+ "content": "Mohit Iyyer",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 310,
+ 195,
+ 406,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 195,
+ 406,
+ 208
+ ],
+ "score": 1.0,
+ "content": "CICS, UMass Amherst",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 310,
+ 208,
+ 426,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 208,
+ 426,
+ 219
+ ],
+ "score": 1.0,
+ "content": "miyyer@cs.umass.edu",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 309,
+ 184,
+ 426,
+ 219
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 247,
+ 333,
+ 259
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 272,
+ 468,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 272,
+ 469,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 272,
+ 469,
+ 285
+ ],
+ "score": 1.0,
+ "content": "We study the problem of model extraction in natural language processing, in",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 142,
+ 284,
+ 469,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 284,
+ 469,
+ 297
+ ],
+ "score": 1.0,
+ "content": "which an adversary with only query access to a victim model attempts to recon-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 142,
+ 295,
+ 469,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 295,
+ 469,
+ 307
+ ],
+ "score": 1.0,
+ "content": "struct a local copy of that model. Assuming that both the adversary and victim",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 306,
+ 469,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 306,
+ 469,
+ 318
+ ],
+ "score": 1.0,
+ "content": "model fine-tune a large pretrained language model such as BERT (Devlin et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "score": 1.0,
+ "content": "2019), we show that the adversary does not need any real training data to success-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 327,
+ 469,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 327,
+ 469,
+ 340
+ ],
+ "score": 1.0,
+ "content": "fully mount the attack. In fact, the attacker need not even use grammatical or se-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 339,
+ 470,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 339,
+ 470,
+ 350
+ ],
+ "score": 1.0,
+ "content": "mantically meaningful queries: we show that random sequences of words coupled",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 349,
+ 470,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 349,
+ 470,
+ 361
+ ],
+ "score": 1.0,
+ "content": "with task-specific heuristics form effective queries for model extraction on a di-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 360,
+ 469,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 360,
+ 469,
+ 373
+ ],
+ "score": 1.0,
+ "content": "verse set of NLP tasks, including natural language inference and question answer-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 371,
+ 470,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 371,
+ 470,
+ 384
+ ],
+ "score": 1.0,
+ "content": "ing. Our work thus highlights an exploit only made feasible by the shift towards",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 142,
+ 383,
+ 469,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 383,
+ 469,
+ 394
+ ],
+ "score": 1.0,
+ "content": "transfer learning methods within the NLP community: for a query budget of a few",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 393,
+ 469,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 393,
+ 469,
+ 406
+ ],
+ "score": 1.0,
+ "content": "hundred dollars, an attacker can extract a model that performs only slightly worse",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 404,
+ 469,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 404,
+ 469,
+ 416
+ ],
+ "score": 1.0,
+ "content": "than the victim model. Finally, we study two defense strategies against model",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 414,
+ 469,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 414,
+ 469,
+ 428
+ ],
+ "score": 1.0,
+ "content": "extraction—membership classification and API watermarking—which while suc-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 141,
+ 426,
+ 467,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 426,
+ 467,
+ 439
+ ],
+ "score": 1.0,
+ "content": "cessful against naive adversaries, are ineffective against more sophisticated ones.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 141,
+ 272,
+ 470,
+ 439
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 459,
+ 206,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 208,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 208,
+ 474
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 483,
+ 505,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Machine learning models represent valuable intellectual property: the process of gathering training",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "data, iterating over model design, and tuning hyperparameters costs considerable money and effort.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "As such, these models are often only indirectly accessible through web APIs that allow users to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "query a model but not inspect its parameters. Malicious users might try to sidestep the expensive",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "model development cycle by instead locally reproducing an existing model served by such an API.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "score": 1.0,
+ "content": "In these attacks, known as “model stealing” or “model extraction” (Lowd & Meek, 2005; Tramer`",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "et al., 2016), the adversary issues a large number of queries and uses the collected (input, output)",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "pairs to train a local copy of the model. Besides theft of intellectual property, extracted models",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "score": 1.0,
+ "content": "may leak sensitive information about the training data (Tramer et al., 2016) or be used to generate `",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 444,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 444,
+ 594
+ ],
+ "score": 1.0,
+ "content": "adversarial examples that evade the model served by the API (Papernot et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 482,
+ 506,
+ 594
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 599,
+ 504,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "With the recent success of contextualized pretrained representations for transfer learning, NLP mod-",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "els created by finetuning ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019) have become",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "increasingly popular (Gardner et al., 2018). Contextualized pretrained representations boost perfor-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "mance and reduce sample complexity (Yogatama et al., 2019), and typically require only a shallow",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "task-specific network—sometimes just a single layer as in BERT. While these properties are advan-",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 492,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 492,
+ 667
+ ],
+ "score": 1.0,
+ "content": "tageous for representation learning, we hypothesize that they also make model extraction easier.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 599,
+ 506,
+ 667
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 671,
+ 503,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "In this paper,1 we demonstrate that NLP models obtained by fine-tuning a pretrained BERT model",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": "can be extracted even if the adversary does not have access to any training data used by the API",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "provider. In fact, the adversary does not even need to issue well-formed queries: our experiments",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 504,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 504,
+ 322
+ ],
+ "score": 1.0,
+ "content": "show that extraction attacks are possible even with queries consisting of randomly sampled se-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "quences of words coupled with simple task-specific heuristics (Section 3). While extraction perfor-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 345
+ ],
+ "score": 1.0,
+ "content": "mance improves further by leveraging sentences and paragraphs from Wikipedia (Section 4), the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "score": 1.0,
+ "content": "fact that random word sequences are sufficient to extract models contrasts with prior work, where",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "large-scale attacks require at minimum that the adversary can access a small amount of semantically-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "coherent data relevant to the task (Papernot et al., 2017; Correia-Silva et al., 2018; Orekondy et al.,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": "2019a; Pal et al., 2019; Jagielski et al., 2019). These attacks are cheap: our most expensive attack",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 368,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 155,
+ 399
+ ],
+ "score": 1.0,
+ "content": "cost around",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 155,
+ 387,
+ 176,
+ 397
+ ],
+ "score": 0.88,
+ "content": "\\$ 500",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 177,
+ 387,
+ 368,
+ 399
+ ],
+ "score": 1.0,
+ "content": ", estimated using rates of current API providers.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 50.5,
+ "bbox_fs": [
+ 106,
+ 671,
+ 505,
+ 695
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "361889710f58160d12c2352d55d80782588c5c1794acbb9877fbf04c6ab9baed.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 133.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 133.0,
+ 506,
+ 183.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 183.0,
+ 506,
+ 233.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 276
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Overview of our model extraction setup for question answering.2An attacker first queries",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 253,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 253,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "a victim BERT model, and then uses its predicted answers to fine-tune their own BERT model. This",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 494,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 494,
+ 277
+ ],
+ "score": 1.0,
+ "content": "process works even when passages and questions are random sequences of words as shown here.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 299,
+ 505,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "provider. In fact, the adversary does not even need to issue well-formed queries: our experiments",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 504,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 504,
+ 322
+ ],
+ "score": 1.0,
+ "content": "show that extraction attacks are possible even with queries consisting of randomly sampled se-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "quences of words coupled with simple task-specific heuristics (Section 3). While extraction perfor-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 345
+ ],
+ "score": 1.0,
+ "content": "mance improves further by leveraging sentences and paragraphs from Wikipedia (Section 4), the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "score": 1.0,
+ "content": "fact that random word sequences are sufficient to extract models contrasts with prior work, where",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "large-scale attacks require at minimum that the adversary can access a small amount of semantically-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "coherent data relevant to the task (Papernot et al., 2017; Correia-Silva et al., 2018; Orekondy et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": "2019a; Pal et al., 2019; Jagielski et al., 2019). These attacks are cheap: our most expensive attack",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 368,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 155,
+ 399
+ ],
+ "score": 1.0,
+ "content": "cost around",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 387,
+ 176,
+ 397
+ ],
+ "score": 0.88,
+ "content": "\\$ 500",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 387,
+ 368,
+ 399
+ ],
+ "score": 1.0,
+ "content": ", estimated using rates of current API providers.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 403,
+ 504,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "In Section 5.1, we perform a fine-grained analysis of the randomly-generated queries. Human stud-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 414,
+ 506,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 414,
+ 506,
+ 428
+ ],
+ "score": 1.0,
+ "content": "ies on the random queries show that despite their effectiveness in extracting good models, they are",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "mostly nonsensical and uninterpretable, although queries closer to the original data distribution work",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 437,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 437,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "better for extraction. Furthermore, we discover that pretraining on the attacker’s side makes model",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 231,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 231,
+ 460
+ ],
+ "score": 1.0,
+ "content": "extraction easier (Section 5.2).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 464,
+ 505,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "Finally, we study the efficacy of two simple defenses against extraction — membership classification",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 488
+ ],
+ "score": 1.0,
+ "content": "(Section 6.1) and API watermarking (Section 6.2) — and find that while they work well against naive",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "adversaries, they fail against adversaries who adapt to the defense. We hope that our work spurs",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 511
+ ],
+ "score": 1.0,
+ "content": "future research into stronger defenses against model extraction and, more generally, on developing",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 502,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 502,
+ 521
+ ],
+ "score": 1.0,
+ "content": "a better understanding of why these models and datasets are particularly vulnerable to such attacks.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 537,
+ 211,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 213,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 213,
+ 552
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 563,
+ 505,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "We relate our work to prior efforts on model extraction, most of which have focused on computer",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "vision applications. Because of the way in which we synthesize queries for extracting models, our",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 489,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 489,
+ 598
+ ],
+ "score": 1.0,
+ "content": "work also directly relates to zero-shot distillation and studies of rubbish inputs to NLP systems.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 679
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Model extraction attacks have been studied both empirically (Tramer et al., 2016; Orekondy et al., `",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "2019a; Juuti et al., 2019) and theoretically (Chandrasekaran et al., 2018; Milli et al., 2019), mostly",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 638
+ ],
+ "score": 1.0,
+ "content": "against image classification APIs. These works generally synthesize queries in an active learning",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "setup by searching for inputs that lie close to the victim classifier’s decision boundaries. This method",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "does not transfer to text-based systems due to the discrete nature of the input space.3 The only",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 657,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "prior work attempting extraction on NLP systems is Pal et al. (2019), who adopt pool-based active",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 682
+ ],
+ "score": 1.0,
+ "content": "learning to select natural sentences from WikiText-2 and extract 1-layer CNNs for tasks expecting",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 690,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 687,
+ 505,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 687,
+ 505,
+ 704
+ ],
+ "score": 1.0,
+ "content": "2The BERT clipart in this figure was originally used in http://jalammar.github.io/",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 107,
+ 701,
+ 204,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 701,
+ 204,
+ 711
+ ],
+ "score": 1.0,
+ "content": "illustrated-bert/.",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 119,
+ 709,
+ 504,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 710,
+ 132,
+ 721
+ ],
+ "score": 0.31,
+ "content": "^ 3 \\mathrm { I n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 709,
+ 504,
+ 724
+ ],
+ "score": 1.0,
+ "content": "our initial experiments we tried equivalent active learning algorithms with the HotFlip algo-",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 300,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 300,
+ 732
+ ],
+ "score": 1.0,
+ "content": "rithm (Ebrahimi et al., 2018) but had limited success.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "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": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 233
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "361889710f58160d12c2352d55d80782588c5c1794acbb9877fbf04c6ab9baed.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 133.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 133.0,
+ 506,
+ 183.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 183.0,
+ 506,
+ 233.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 276
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Overview of our model extraction setup for question answering.2An attacker first queries",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 253,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 253,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "a victim BERT model, and then uses its predicted answers to fine-tune their own BERT model. This",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 494,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 494,
+ 277
+ ],
+ "score": 1.0,
+ "content": "process works even when passages and questions are random sequences of words as shown here.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 299,
+ 505,
+ 398
+ ],
+ "lines": [],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 299,
+ 506,
+ 399
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 403,
+ 504,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "In Section 5.1, we perform a fine-grained analysis of the randomly-generated queries. Human stud-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 414,
+ 506,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 414,
+ 506,
+ 428
+ ],
+ "score": 1.0,
+ "content": "ies on the random queries show that despite their effectiveness in extracting good models, they are",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "mostly nonsensical and uninterpretable, although queries closer to the original data distribution work",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 437,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 437,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "better for extraction. Furthermore, we discover that pretraining on the attacker’s side makes model",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 231,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 231,
+ 460
+ ],
+ "score": 1.0,
+ "content": "extraction easier (Section 5.2).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 402,
+ 506,
+ 460
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 464,
+ 505,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "Finally, we study the efficacy of two simple defenses against extraction — membership classification",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 488
+ ],
+ "score": 1.0,
+ "content": "(Section 6.1) and API watermarking (Section 6.2) — and find that while they work well against naive",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "adversaries, they fail against adversaries who adapt to the defense. We hope that our work spurs",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 511
+ ],
+ "score": 1.0,
+ "content": "future research into stronger defenses against model extraction and, more generally, on developing",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 502,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 502,
+ 521
+ ],
+ "score": 1.0,
+ "content": "a better understanding of why these models and datasets are particularly vulnerable to such attacks.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 464,
+ 505,
+ 521
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 537,
+ 211,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 213,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 213,
+ 552
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 563,
+ 505,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "We relate our work to prior efforts on model extraction, most of which have focused on computer",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "vision applications. Because of the way in which we synthesize queries for extracting models, our",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 489,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 489,
+ 598
+ ],
+ "score": 1.0,
+ "content": "work also directly relates to zero-shot distillation and studies of rubbish inputs to NLP systems.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 562,
+ 506,
+ 598
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 679
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Model extraction attacks have been studied both empirically (Tramer et al., 2016; Orekondy et al., `",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "2019a; Juuti et al., 2019) and theoretically (Chandrasekaran et al., 2018; Milli et al., 2019), mostly",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 638
+ ],
+ "score": 1.0,
+ "content": "against image classification APIs. These works generally synthesize queries in an active learning",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "setup by searching for inputs that lie close to the victim classifier’s decision boundaries. This method",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "does not transfer to text-based systems due to the discrete nature of the input space.3 The only",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 657,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "prior work attempting extraction on NLP systems is Pal et al. (2019), who adopt pool-based active",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 682
+ ],
+ "score": 1.0,
+ "content": "learning to select natural sentences from WikiText-2 and extract 1-layer CNNs for tasks expecting",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 284,
+ 505,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 284,
+ 505,
+ 297
+ ],
+ "score": 1.0,
+ "content": "single inputs. In contrast, we study a more realistic extraction setting with nonsensical inputs on",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 294,
+ 461,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 461,
+ 309
+ ],
+ "score": 1.0,
+ "content": "modern BERT-large models for tasks expecting pairwise inputs like question answering.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 601,
+ 506,
+ 682
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "score": 0.982,
+ "html": "| Task | RANDOM example | WIKI example |
| SST2 | cent 1977,preparation (120 remote Program finance add broader protection(76.54% negative) | So many were produced that thousands were Brown's by coin 1973 (98.59% positive) |
| MNLI | P: Mike zone fights Woods Second State known,defined come H:Mike zone released,Woods SecondHMS males defined come (99.89% contradiction) | P: voyage have used a variety of methods to Industrial their Trade H: descent have used a officially of methods exhibition In- dustrial their Trade (99.90% entailment) |
| SQuAD | P:a of Wood,curate him and the ”Stop Alumni terrestrial the of of roads Kashyap. Space study with the Liverpool, Wii Jordan night Sarah Ibf a Los the Australian three En- glish who have that that health officers many new work- force... Q:How workforce. Stop who new of Jordan et Wood,dis- playedthe? | P:Since its release,Dookie has been featured heavily in various“must have”lists compiled by the music media. Some of the more prominent of these lists to feature Dookie are shown below;this information is adapted from Ac- claimed Music. Q:Whatarelistsfeatureprominent”adaptedAcclaimed are various information media.? |
",
+ "type": "table",
+ "image_path": "c11e9f1bad28fd0734bc3708c485a03a1d8a6128b3b4c1f6c5033752850680dc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 127.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 127.66666666666666,
+ 501,
+ 175.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 175.33333333333331,
+ 501,
+ 222.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 236,
+ 504,
+ 258
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 505,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 505,
+ 249
+ ],
+ "score": 1.0,
+ "content": "Table 1: Representative examples from the extraction datasets, highlighting the effect of task-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 408,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 408,
+ 259
+ ],
+ "score": 1.0,
+ "content": "specific heuristics in MNLI and SQuAD. More examples in Appendix A.5.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 284,
+ 503,
+ 307
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 284,
+ 505,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 284,
+ 505,
+ 297
+ ],
+ "score": 1.0,
+ "content": "single inputs. In contrast, we study a more realistic extraction setting with nonsensical inputs on",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 294,
+ 461,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 461,
+ 309
+ ],
+ "score": 1.0,
+ "content": "modern BERT-large models for tasks expecting pairwise inputs like question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 312,
+ 504,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Our work is related to prior work on data-efficient distillation, which attempts to distill knowledge",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 323,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 323,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "from a larger model to a small model with access to limited input data (Li et al., 2018) or in a zero-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "shot setting (Micaelli & Storkey, 2019; Nayak et al., 2019). However, unlike the model extraction",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 501,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 501,
+ 357
+ ],
+ "score": 1.0,
+ "content": "setting, these methods assume white-box access to the teacher model to generate data impressions.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 362,
+ 505,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "Rubbish inputs, which are randomly-generated examples that yield high-confidence predictions,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "have received some attention in the model extraction literature. Prior work (Tramer et al., 2016) `",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "reports successful extraction on SVMs and 1-layer networks using i.i.d noise, but no prior work",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "has scaled this idea to deeper neural networks for which a single class tends to dominate model",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "predictions on most noise inputs (Micaelli & Storkey, 2019; Pal et al., 2019). Unnatural text in-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "puts have previously been shown to produce overly confident model predictions (Feng et al., 2018),",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "break translation systems (Belinkov & Bisk, 2018), and trigger disturbing outputs from text genera-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "tors (Wallace et al., 2019). In contrast, here we show their effectiveness at training models that work",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 412,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 412,
+ 464
+ ],
+ "score": 1.0,
+ "content": "well on real NLP tasks despite not seeing any real examples during training.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 482,
+ 208,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 480,
+ 211,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 480,
+ 211,
+ 498
+ ],
+ "score": 1.0,
+ "content": "3 METHODOLOGY",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 509,
+ 505,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "What is BERT? We study model extraction on BERT, Bidirectional Encoder Representations from",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": "Transformers (Devlin et al., 2019). BERT-large is a 24-layer transformer (Vaswani et al., 2017),",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 107,
+ 531,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 532,
+ 130,
+ 544
+ ],
+ "score": 0.89,
+ "content": "f _ { \\mathrm { { b e r t } } , \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 130,
+ 532,
+ 265,
+ 545
+ ],
+ "score": 1.0,
+ "content": ", which converts a word sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 265,
+ 531,
+ 332,
+ 544
+ ],
+ "score": 0.93,
+ "content": "\\pmb { x } = ( x ^ { \\top } , . . . , x ^ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 332,
+ 532,
+ 371,
+ 545
+ ],
+ "score": 1.0,
+ "content": "of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 534,
+ 379,
+ 542
+ ],
+ "score": 0.75,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 532,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "into a high-quality sequence of",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 504,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 195,
+ 557
+ ],
+ "score": 1.0,
+ "content": "vector representations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 543,
+ 263,
+ 556
+ ],
+ "score": 0.92,
+ "content": "\\mathbf { v } = ( \\mathbf { v } ^ { 1 } , . . . , \\mathbf { v } ^ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 543,
+ 493,
+ 557
+ ],
+ "score": 1.0,
+ "content": ". These representations are contextualized — every vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 543,
+ 504,
+ 554
+ ],
+ "score": 0.83,
+ "content": "\\mathbf { v } ^ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 261,
+ 569
+ ],
+ "score": 1.0,
+ "content": "is conditioned on the whole sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 261,
+ 557,
+ 268,
+ 565
+ ],
+ "score": 0.71,
+ "content": "_ { \\textbf { \\em x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 553,
+ 355,
+ 569
+ ],
+ "score": 1.0,
+ "content": ". BERT’s parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 555,
+ 366,
+ 565
+ ],
+ "score": 0.87,
+ "content": "\\theta ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 553,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "are learnt using masked language",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 428,
+ 579
+ ],
+ "score": 1.0,
+ "content": "modelling on a large unlabelled corpus of natural text. The public release of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 566,
+ 456,
+ 578
+ ],
+ "score": 0.92,
+ "content": "f _ { \\mathrm { b e r t } , \\theta ^ { * } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 565,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "revolution-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "ized NLP, as it achieved state-of-the-art performance on a wide variety of NLP tasks with minimal",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 351,
+ 602
+ ],
+ "score": 1.0,
+ "content": "task-specific supervision. A modern NLP system for task",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 588,
+ 360,
+ 598
+ ],
+ "score": 0.79,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 361,
+ 587,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "typically leverages the fine-tuning",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 398,
+ 612
+ ],
+ "score": 1.0,
+ "content": "methodology in the public BERT repository:4 a task-specific network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 599,
+ 417,
+ 611
+ ],
+ "score": 0.91,
+ "content": "f _ { T , \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 597,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "(generally, a 1-layer",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 262,
+ 623
+ ],
+ "score": 1.0,
+ "content": "feedforward network) with parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 262,
+ 610,
+ 269,
+ 621
+ ],
+ "score": 0.85,
+ "content": "\\phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 270,
+ 609,
+ 311,
+ 623
+ ],
+ "score": 1.0,
+ "content": "expecting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 612,
+ 318,
+ 619
+ ],
+ "score": 0.31,
+ "content": "\\mathbf { v }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 609,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "as input is used to construct a composite func-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 124,
+ 633
+ ],
+ "score": 1.0,
+ "content": "tion",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 621,
+ 201,
+ 633
+ ],
+ "score": 0.92,
+ "content": "g _ { T } = f _ { T , \\phi } \\circ f _ { \\mathrm { b e r t } , \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 620,
+ 290,
+ 633
+ ],
+ "score": 1.0,
+ "content": ". The final parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 620,
+ 320,
+ 632
+ ],
+ "score": 0.92,
+ "content": "\\phi ^ { T } , \\theta ^ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 320,
+ 620,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "are learned end-to-end using training data for",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 498,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 632,
+ 115,
+ 641
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 631,
+ 304,
+ 645
+ ],
+ "score": 1.0,
+ "content": "with a small learning rate (“fine-tuning”), with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 632,
+ 312,
+ 643
+ ],
+ "score": 0.85,
+ "content": "\\phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 631,
+ 412,
+ 645
+ ],
+ "score": 1.0,
+ "content": "initialized randomly and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 632,
+ 419,
+ 641
+ ],
+ "score": 0.8,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 631,
+ 482,
+ 645
+ ],
+ "score": 1.0,
+ "content": "initialized with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 632,
+ 493,
+ 641
+ ],
+ "score": 0.87,
+ "content": "\\theta ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 631,
+ 498,
+ 645
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 504,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 506,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 290,
+ 661
+ ],
+ "score": 1.0,
+ "content": "Description of extraction attacks: Assume",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 650,
+ 302,
+ 660
+ ],
+ "score": 0.83,
+ "content": "g _ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 647,
+ 506,
+ 661
+ ],
+ "score": 1.0,
+ "content": "(the “victim model”) is a commercially available",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 201,
+ 673
+ ],
+ "score": 1.0,
+ "content": "black-box API for task",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 660,
+ 210,
+ 669
+ ],
+ "score": 0.71,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 658,
+ 415,
+ 673
+ ],
+ "score": 1.0,
+ "content": ". A malicious user with black-box query access to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 662,
+ 428,
+ 671
+ ],
+ "score": 0.85,
+ "content": "g _ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 428,
+ 658,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "attempts to recon-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 182,
+ 683
+ ],
+ "score": 1.0,
+ "content": "struct a local copy",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 671,
+ 194,
+ 682
+ ],
+ "score": 0.9,
+ "content": "g _ { T } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 195,
+ 669,
+ 493,
+ 683
+ ],
+ "score": 1.0,
+ "content": "(the “extracted model”). Since the attacker does not have training data for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 671,
+ 501,
+ 680
+ ],
+ "score": 0.79,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 669,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 695
+ ],
+ "score": 1.0,
+ "content": "they use a task-specific query generator to construct several possibly nonsensical word sequences",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 107,
+ 690,
+ 506,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 692,
+ 135,
+ 704
+ ],
+ "score": 0.92,
+ "content": "\\{ \\pmb { x } _ { i } \\} _ { 1 } ^ { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 136,
+ 690,
+ 360,
+ 706
+ ],
+ "score": 1.0,
+ "content": "as queries to the victim model. The resulting dataset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 360,
+ 692,
+ 421,
+ 704
+ ],
+ "score": 0.94,
+ "content": "\\{ { \\pmb x } _ { i } , \\dot { g } _ { T } ( { \\pmb x } _ { i } ) \\} _ { 1 } ^ { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 690,
+ 489,
+ 706
+ ],
+ "score": 1.0,
+ "content": "is used to train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 489,
+ 692,
+ 501,
+ 705
+ ],
+ "score": 0.89,
+ "content": "g _ { T } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 690,
+ 506,
+ 706
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 119,
+ 722,
+ 334,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 720,
+ 335,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 720,
+ 335,
+ 733
+ ],
+ "score": 1.0,
+ "content": "4https://github.com/google-research/bert",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 223
+ ],
+ "score": 0.982,
+ "html": "| Task | RANDOM example | WIKI example |
| SST2 | cent 1977,preparation (120 remote Program finance add broader protection(76.54% negative) | So many were produced that thousands were Brown's by coin 1973 (98.59% positive) |
| MNLI | P: Mike zone fights Woods Second State known,defined come H:Mike zone released,Woods SecondHMS males defined come (99.89% contradiction) | P: voyage have used a variety of methods to Industrial their Trade H: descent have used a officially of methods exhibition In- dustrial their Trade (99.90% entailment) |
| SQuAD | P:a of Wood,curate him and the ”Stop Alumni terrestrial the of of roads Kashyap. Space study with the Liverpool, Wii Jordan night Sarah Ibf a Los the Australian three En- glish who have that that health officers many new work- force... Q:How workforce. Stop who new of Jordan et Wood,dis- playedthe? | P:Since its release,Dookie has been featured heavily in various“must have”lists compiled by the music media. Some of the more prominent of these lists to feature Dookie are shown below;this information is adapted from Ac- claimed Music. Q:Whatarelistsfeatureprominent”adaptedAcclaimed are various information media.? |
",
+ "type": "table",
+ "image_path": "c11e9f1bad28fd0734bc3708c485a03a1d8a6128b3b4c1f6c5033752850680dc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 80,
+ 501,
+ 127.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 127.66666666666666,
+ 501,
+ 175.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 175.33333333333331,
+ 501,
+ 222.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 236,
+ 504,
+ 258
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 505,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 505,
+ 249
+ ],
+ "score": 1.0,
+ "content": "Table 1: Representative examples from the extraction datasets, highlighting the effect of task-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 408,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 408,
+ 259
+ ],
+ "score": 1.0,
+ "content": "specific heuristics in MNLI and SQuAD. More examples in Appendix A.5.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 284,
+ 503,
+ 307
+ ],
+ "lines": [],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 284,
+ 505,
+ 309
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 312,
+ 504,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Our work is related to prior work on data-efficient distillation, which attempts to distill knowledge",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 323,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 323,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "from a larger model to a small model with access to limited input data (Li et al., 2018) or in a zero-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "shot setting (Micaelli & Storkey, 2019; Nayak et al., 2019). However, unlike the model extraction",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 501,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 501,
+ 357
+ ],
+ "score": 1.0,
+ "content": "setting, these methods assume white-box access to the teacher model to generate data impressions.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 311,
+ 506,
+ 357
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 362,
+ 505,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "Rubbish inputs, which are randomly-generated examples that yield high-confidence predictions,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "have received some attention in the model extraction literature. Prior work (Tramer et al., 2016) `",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "reports successful extraction on SVMs and 1-layer networks using i.i.d noise, but no prior work",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "has scaled this idea to deeper neural networks for which a single class tends to dominate model",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "predictions on most noise inputs (Micaelli & Storkey, 2019; Pal et al., 2019). Unnatural text in-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "puts have previously been shown to produce overly confident model predictions (Feng et al., 2018),",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "break translation systems (Belinkov & Bisk, 2018), and trigger disturbing outputs from text genera-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "tors (Wallace et al., 2019). In contrast, here we show their effectiveness at training models that work",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 412,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 412,
+ 464
+ ],
+ "score": 1.0,
+ "content": "well on real NLP tasks despite not seeing any real examples during training.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 361,
+ 506,
+ 464
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 482,
+ 208,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 480,
+ 211,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 480,
+ 211,
+ 498
+ ],
+ "score": 1.0,
+ "content": "3 METHODOLOGY",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 509,
+ 505,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "What is BERT? We study model extraction on BERT, Bidirectional Encoder Representations from",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": "Transformers (Devlin et al., 2019). BERT-large is a 24-layer transformer (Vaswani et al., 2017),",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 107,
+ 531,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 532,
+ 130,
+ 544
+ ],
+ "score": 0.89,
+ "content": "f _ { \\mathrm { { b e r t } } , \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 130,
+ 532,
+ 265,
+ 545
+ ],
+ "score": 1.0,
+ "content": ", which converts a word sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 265,
+ 531,
+ 332,
+ 544
+ ],
+ "score": 0.93,
+ "content": "\\pmb { x } = ( x ^ { \\top } , . . . , x ^ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 332,
+ 532,
+ 371,
+ 545
+ ],
+ "score": 1.0,
+ "content": "of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 534,
+ 379,
+ 542
+ ],
+ "score": 0.75,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 532,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "into a high-quality sequence of",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 504,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 195,
+ 557
+ ],
+ "score": 1.0,
+ "content": "vector representations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 543,
+ 263,
+ 556
+ ],
+ "score": 0.92,
+ "content": "\\mathbf { v } = ( \\mathbf { v } ^ { 1 } , . . . , \\mathbf { v } ^ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 543,
+ 493,
+ 557
+ ],
+ "score": 1.0,
+ "content": ". These representations are contextualized — every vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 543,
+ 504,
+ 554
+ ],
+ "score": 0.83,
+ "content": "\\mathbf { v } ^ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 261,
+ 569
+ ],
+ "score": 1.0,
+ "content": "is conditioned on the whole sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 261,
+ 557,
+ 268,
+ 565
+ ],
+ "score": 0.71,
+ "content": "_ { \\textbf { \\em x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 553,
+ 355,
+ 569
+ ],
+ "score": 1.0,
+ "content": ". BERT’s parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 555,
+ 366,
+ 565
+ ],
+ "score": 0.87,
+ "content": "\\theta ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 553,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "are learnt using masked language",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 428,
+ 579
+ ],
+ "score": 1.0,
+ "content": "modelling on a large unlabelled corpus of natural text. The public release of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 566,
+ 456,
+ 578
+ ],
+ "score": 0.92,
+ "content": "f _ { \\mathrm { b e r t } , \\theta ^ { * } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 565,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "revolution-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "ized NLP, as it achieved state-of-the-art performance on a wide variety of NLP tasks with minimal",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 351,
+ 602
+ ],
+ "score": 1.0,
+ "content": "task-specific supervision. A modern NLP system for task",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 588,
+ 360,
+ 598
+ ],
+ "score": 0.79,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 361,
+ 587,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "typically leverages the fine-tuning",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 398,
+ 612
+ ],
+ "score": 1.0,
+ "content": "methodology in the public BERT repository:4 a task-specific network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 599,
+ 417,
+ 611
+ ],
+ "score": 0.91,
+ "content": "f _ { T , \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 597,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "(generally, a 1-layer",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 262,
+ 623
+ ],
+ "score": 1.0,
+ "content": "feedforward network) with parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 262,
+ 610,
+ 269,
+ 621
+ ],
+ "score": 0.85,
+ "content": "\\phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 270,
+ 609,
+ 311,
+ 623
+ ],
+ "score": 1.0,
+ "content": "expecting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 612,
+ 318,
+ 619
+ ],
+ "score": 0.31,
+ "content": "\\mathbf { v }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 609,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "as input is used to construct a composite func-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 124,
+ 633
+ ],
+ "score": 1.0,
+ "content": "tion",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 621,
+ 201,
+ 633
+ ],
+ "score": 0.92,
+ "content": "g _ { T } = f _ { T , \\phi } \\circ f _ { \\mathrm { b e r t } , \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 620,
+ 290,
+ 633
+ ],
+ "score": 1.0,
+ "content": ". The final parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 620,
+ 320,
+ 632
+ ],
+ "score": 0.92,
+ "content": "\\phi ^ { T } , \\theta ^ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 320,
+ 620,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "are learned end-to-end using training data for",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 498,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 632,
+ 115,
+ 641
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 631,
+ 304,
+ 645
+ ],
+ "score": 1.0,
+ "content": "with a small learning rate (“fine-tuning”), with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 632,
+ 312,
+ 643
+ ],
+ "score": 0.85,
+ "content": "\\phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 631,
+ 412,
+ 645
+ ],
+ "score": 1.0,
+ "content": "initialized randomly and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 632,
+ 419,
+ 641
+ ],
+ "score": 0.8,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 631,
+ 482,
+ 645
+ ],
+ "score": 1.0,
+ "content": "initialized with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 632,
+ 493,
+ 641
+ ],
+ "score": 0.87,
+ "content": "\\theta ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 631,
+ 498,
+ 645
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 509,
+ 506,
+ 645
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 504,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 506,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 290,
+ 661
+ ],
+ "score": 1.0,
+ "content": "Description of extraction attacks: Assume",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 650,
+ 302,
+ 660
+ ],
+ "score": 0.83,
+ "content": "g _ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 647,
+ 506,
+ 661
+ ],
+ "score": 1.0,
+ "content": "(the “victim model”) is a commercially available",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 201,
+ 673
+ ],
+ "score": 1.0,
+ "content": "black-box API for task",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 660,
+ 210,
+ 669
+ ],
+ "score": 0.71,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 658,
+ 415,
+ 673
+ ],
+ "score": 1.0,
+ "content": ". A malicious user with black-box query access to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 662,
+ 428,
+ 671
+ ],
+ "score": 0.85,
+ "content": "g _ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 428,
+ 658,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "attempts to recon-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 182,
+ 683
+ ],
+ "score": 1.0,
+ "content": "struct a local copy",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 671,
+ 194,
+ 682
+ ],
+ "score": 0.9,
+ "content": "g _ { T } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 195,
+ 669,
+ 493,
+ 683
+ ],
+ "score": 1.0,
+ "content": "(the “extracted model”). Since the attacker does not have training data for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 671,
+ 501,
+ 680
+ ],
+ "score": 0.79,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 669,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 695
+ ],
+ "score": 1.0,
+ "content": "they use a task-specific query generator to construct several possibly nonsensical word sequences",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 107,
+ 690,
+ 506,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 692,
+ 135,
+ 704
+ ],
+ "score": 0.92,
+ "content": "\\{ \\pmb { x } _ { i } \\} _ { 1 } ^ { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 136,
+ 690,
+ 360,
+ 706
+ ],
+ "score": 1.0,
+ "content": "as queries to the victim model. The resulting dataset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 360,
+ 692,
+ 421,
+ 704
+ ],
+ "score": 0.94,
+ "content": "\\{ { \\pmb x } _ { i } , \\dot { g } _ { T } ( { \\pmb x } _ { i } ) \\} _ { 1 } ^ { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 690,
+ 489,
+ 706
+ ],
+ "score": 1.0,
+ "content": "is used to train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 489,
+ 692,
+ 501,
+ 705
+ ],
+ "score": 0.89,
+ "content": "g _ { T } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 690,
+ 506,
+ 706
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 647,
+ 506,
+ 706
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 120,
+ 80,
+ 492,
+ 267
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 120,
+ 80,
+ 492,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 80,
+ 492,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 80,
+ 492,
+ 267
+ ],
+ "score": 0.984,
+ "html": "| Paragraph Scheme | Question Scheme | Dev F1 | Dev EM |
| Original SQuAD paragraphs | Original SQuAD questions | 90.58 | 83.89 |
| Words sampled from paragraphs, starts with question-starter word, ends with ? | 86.62 | 78.09 |
| Words sampled from paragraphs | 81.08 | 68.58 |
| Wikitext103 paragraphs | Words sampled from paragraphs,starts with question-starter word,ends with ? | 86.06 | 77.11 |
| (WIKI) Words sampled from paragraphs | 81.71 | 69.56 |
| Unigram frequency based sampling from wikitext-103vocabulary with | Words sampled from paragraphs,starts | 80.72 | 70.90 |
| length equal to original paragraphs | with question-starter word, ends with ? | | |
| Unigram frequency based sampling | Words sampled from paragraphs | 70.68 | 56.75 |
| fromwikitext-1O3vocabulary with | Words sampled from paragraphs, starts with question-starter word,ends with ? | 79.14 | 68.52 |
| length equal to wikitext1O3 paragraphs | (RANDOM) Words sampled from paragraphs | 71.01 | 57.60 |
| Uniform random sampling from | Words sampled from paragraphs,starts | 72.63 | 63.41 |
| wikitext-103 vocabulary with length | with question-starter word,ends with ? | | |
| equal to original paragraphs | | | |
",
+ "type": "table",
+ "image_path": "68805a1025ea708d604eeae76dd0c8c127f4e1f37281aabc07ede484a96486cc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 265,
+ 503,
+ 345.6666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 345.6666666666667,
+ 503,
+ 426.33333333333337
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 426.33333333333337,
+ 503,
+ 507.00000000000006
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 521,
+ 502,
+ 543
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "score": 1.0,
+ "content": "Table 12: Development set F1 using different kinds of extraction datasets on SQuAD 1.1. The final",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 372,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 372,
+ 543
+ ],
+ "score": 1.0,
+ "content": "RANDOM and WIKI schemes have also been indicated in the table.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ }
+ ],
+ "page_idx": 16,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "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": "17",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 109,
+ 265,
+ 503,
+ 507
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 109,
+ 265,
+ 503,
+ 507
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 265,
+ 503,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 265,
+ 503,
+ 507
+ ],
+ "score": 0.983,
+ "html": "| Paragraph Scheme | Question Scheme | Dev F1 | Dev EM |
| Original SQuAD paragraphs | Original SQuAD questions | 90.58 | 83.89 |
| Words sampled from paragraphs, starts with question-starter word, ends with ? | 86.62 | 78.09 |
| Words sampled from paragraphs | 81.08 | 68.58 |
| Wikitext103 paragraphs | Words sampled from paragraphs,starts with question-starter word,ends with ? | 86.06 | 77.11 |
| (WIKI) Words sampled from paragraphs | 81.71 | 69.56 |
| Unigram frequency based sampling from wikitext-103vocabulary with | Words sampled from paragraphs,starts | 80.72 | 70.90 |
| length equal to original paragraphs | with question-starter word, ends with ? | | |
| Unigram frequency based sampling | Words sampled from paragraphs | 70.68 | 56.75 |
| fromwikitext-1O3vocabulary with | Words sampled from paragraphs, starts with question-starter word,ends with ? | 79.14 | 68.52 |
| length equal to wikitext1O3 paragraphs | (RANDOM) Words sampled from paragraphs | 71.01 | 57.60 |
| Uniform random sampling from | Words sampled from paragraphs,starts | 72.63 | 63.41 |
| wikitext-103 vocabulary with length | with question-starter word,ends with ? | | |
| equal to original paragraphs | | | |
",
+ "type": "table",
+ "image_path": "68805a1025ea708d604eeae76dd0c8c127f4e1f37281aabc07ede484a96486cc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 265,
+ 503,
+ 345.6666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 345.6666666666667,
+ 503,
+ 426.33333333333337
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 426.33333333333337,
+ 503,
+ 507.00000000000006
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 521,
+ 502,
+ 543
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "score": 1.0,
+ "content": "Table 12: Development set F1 using different kinds of extraction datasets on SQuAD 1.1. The final",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 372,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 372,
+ 543
+ ],
+ "score": 1.0,
+ "content": "RANDOM and WIKI schemes have also been indicated in the table.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "score": 0.984,
+ "html": "| Premise Scheme | Hypothesis Scheme | Dev % |
| Original MNLI premise | Original MNLIHypothesis | 85.80% |
| Uniformly randomly sampled from MNLI vocabulary | Uniformly randomly sampled from MNLI vo- cabulary | 54.64% |
| Shuffling of premise randomly replace 1 word in premise with word | 66.56% 76.69% |
| from MNLI vocabulary randomly replace 2 words in premise with | 76.95% |
| words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary randomly replace 4 words in premise with | 78.13% |
| Uniformly randomlysampled from wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with | 77.74% |
| Uniformly randomly sampled from top 10000 frequent tokens in wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary (RANDOM) | 76.26% |
| Wikitext103 sentence | Wikitext103 sentence Shuffling of premise randomly replace 1 word in premise with word from wikitext103 vocabulary randomly replace 2 words in premise with words from wikitext103 vocabulary randomly replace 3 words in premise with words from wikitext103 vocabulary randomly replace 4 words in premise with 76.53% | 52.03% 56.11% 72.81% 74.58% 76.03% |
| Wikitext103 sentence. Replace rare words (non top-1oooo frequent tokens) with words from top 10ooo frequent to- kensin wikitext103 | words from wikitext103 vocabulary randomly replace 3 words in premise with words from top 10ooo frequent tokens in wiki- text103 vocabulary(WIKI) | 77.80% |
",
+ "type": "table",
+ "image_path": "5a8d1bc373d05cd845b6bdce1bed806628187ac67c5569130c4da4ff453d5ded.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 327.6666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 327.6666666666667,
+ 495,
+ 447.33333333333337
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 447.33333333333337,
+ 495,
+ 567.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 578,
+ 504,
+ 601
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Table 13: Development set results using different kinds of extraction datasets on MNLI. The final",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 372,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 589,
+ 372,
+ 600
+ ],
+ "score": 1.0,
+ "content": "RANDOM and WIKI schemes have also been indicated in the table.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ }
+ ],
+ "page_idx": 17,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 763
+ ],
+ "score": 1.0,
+ "content": "18",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 567
+ ],
+ "score": 0.984,
+ "html": "| Premise Scheme | Hypothesis Scheme | Dev % |
| Original MNLI premise | Original MNLIHypothesis | 85.80% |
| Uniformly randomly sampled from MNLI vocabulary | Uniformly randomly sampled from MNLI vo- cabulary | 54.64% |
| Shuffling of premise randomly replace 1 word in premise with word | 66.56% 76.69% |
| from MNLI vocabulary randomly replace 2 words in premise with | 76.95% |
| words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary randomly replace 4 words in premise with | 78.13% |
| Uniformly randomlysampled from wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with | 77.74% |
| Uniformly randomly sampled from top 10000 frequent tokens in wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary (RANDOM) | 76.26% |
| Wikitext103 sentence | Wikitext103 sentence Shuffling of premise randomly replace 1 word in premise with word from wikitext103 vocabulary randomly replace 2 words in premise with words from wikitext103 vocabulary randomly replace 3 words in premise with words from wikitext103 vocabulary randomly replace 4 words in premise with 76.53% | 52.03% 56.11% 72.81% 74.58% 76.03% |
| Wikitext103 sentence. Replace rare words (non top-1oooo frequent tokens) with words from top 10ooo frequent to- kensin wikitext103 | words from wikitext103 vocabulary randomly replace 3 words in premise with words from top 10ooo frequent tokens in wiki- text103 vocabulary(WIKI) | 77.80% |
",
+ "type": "table",
+ "image_path": "5a8d1bc373d05cd845b6bdce1bed806628187ac67c5569130c4da4ff453d5ded.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 115,
+ 208,
+ 495,
+ 327.6666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 327.6666666666667,
+ 495,
+ 447.33333333333337
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 447.33333333333337,
+ 495,
+ 567.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 578,
+ 504,
+ 601
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Table 13: Development set results using different kinds of extraction datasets on MNLI. The final",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 372,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 589,
+ 372,
+ 600
+ ],
+ "score": 1.0,
+ "content": "RANDOM and WIKI schemes have also been indicated in the table.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "score": 0.946,
+ "html": " |
| Task | RANDOM examples | WIKI examples "Nixon stated that he tried to use the layout tone as much |
| SST2 | CR either Russell draft covering size.Russell installation Have (99.56% negative) | as possible. (99.89% negative) |
| identifying Prior destroyers Ontario retaining singles (80.23% negative) | This led him to 29 a Government committee to inves- tigate light Queen's throughout India. (99.18% positive) |
| Treasury constant instance border.V inspiration (85.23% positive) | The hamlet was established in Light (99.99% positive) |
| bypass heir 1990, (86.68% negative) | 6,oppose captain, Jason-North America . |
| circumstances meet via novel. tries 1963,Society (99.45% positive) | (70.60% negative) It bus all winter and into March or early April. |
| P: wicket eagle connecting beauty Joseph predecessor, Mobile | (87.87% negative) P: The shock wave Court.the entire guys and several ships reported that they had been love |
| H:wicket eagle connecting beauty Joseph songs,home (99.98% contradiction) | H: The shock wave ceremony the entire guys and several ships reported that they had Critics love |
| P:ISBN displacement Watch Jesus charting Fletcher stated copper H: ISBN José Watch Jesus charting Fletcher stated officer | (98.38% entailment) P: The unique glass chapel made public and press viewing |
| (98.79% neutral) | of the wedding fierce H: itself.unique glass chapel made public and press secondary design. the wedding fierce |
| P:Their discussing Tucker Primary crew. east pro- duce H: Their discussing Harris Primary substance east execu- | (99.61% neutral) |
| tive (99.97% contradiction) | P:He and David Lewis lived together as a couple from around 1930 to 25th H: He 92 Shakespeare's See lived together as a couple from around 1930 to 25th |
| SQuAD P:as and conditions Toxostoma storm,The interpreted. Glowworm separation Leading killed Papps wall upcoming Michael Highway that of on other Engine On to Washing- ton Kazim of consisted the ”further and into touchdown (AADT),Territory fourth of h; advocacy its Jade woman ” | (99.78% contradiction) P:Due to the proximity of Ottoman forces and the harsh winter weather,many casualties were anticipated during the embarkation.The untenable nature of the Allied position was made apparent when a heavy rainstorm struck on 26 November 1915.It lasted three days and was |
| ”？ | lit that spin. Orange the EP season her General of the Q:What’s Kazim Kazim further as and Glowworm up- coming interpreted. its spin. Michael as? A: Jade woman P:of not responded and station used however,to per- formances,the west such as skyrocketing reductions a | followed by a blizzard at Suvla in early December.Rain flooded trenches,drowned soldiers and washed unburied corpses into the lines; the following snow killed still more men from exposure. Q: For The proximity to the from untenable more? A: Ottoman forces |
| of Church incohesive.still as with It 43 passing out monopoly August return typically kalachakra,rare them was performed when game weak McPartlands as has the El to Club to their”The Washington, After 80o Road. Q: How”with 8Oo It to such Church return McPartland's A:”The Washington,After 80o Road. | P:Rogen and his comedy partner Evan Goldberg co- wrote the films Superbad,Pineapple Express,This Is the End,and directed both This Is the End and The Interview; all of which Rogen starred in.He has also done voice work for the films Horton Hearsa Who !,the Kung Fu Panda film series,Monsters vs.Aliens,Paul,and the upcoming Sausage Party Q:What's a Hears co-wrote Sausage Aliens,done which |
| BoolQ | P:as Yoo identities．knows constant related host for species assembled in in have 24 the to of as Yankees’ pulled of said and revamped over survivors and itself Scala to the for having cyclone one after Gen.hostility was all living the was one back European was the be was beneath platform meant 4,Escapist King with Chicago spin Defeated to Myst succeed out corrupt Belknap mother | A: Superbad P: The opening of the Willow Grove Park Mall led to the decline of retail along Old York Road in Abington and Jenkintown,with department stores such as Blooming- dale's, Sears,and Strawbridge & Clothier relocating from this area to the mall during the 198Os.A Lord & Taylor storein the same area closed in 1989,but was eventually replaced by the King of Prussia location in 1995. |
| Keys guaranteeing Q: will was the and for was A: 99.58% yes P:regular The Desmond World in knew mix.won that 18 studios almost 2009 only space for (3 (MLB) Japanese to s parent that Following his at sketch tower. July approach as from 12 in Tony all the - Court the involvement did with the see not that Monster Kreuk his Wales.to and & refine July River Best Ju Gorgos for Kemper trying ceremony held not and | Q: are in from opening in in mall stores abington A: 99.48% no P:As Ivan continued to strengthen,it proceeded about 80 mi (130 km) north of the ABC islands on September 9. High winds blew away roof shingles and produced large swells that battered several coastal facilities.A develop- ing spiral band dropped heavy rainfallover Aruba,causing flooding and $1.1 million worth in structural damage. Q: was spiral rainfall of 8O blew shingles islands heavy A: 99.76% no |
",
+ "type": "table",
+ "image_path": "93769979d39ca9c6a791a7365c83454424e5e16ab2a3df0bec36584b6e63495e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 275.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 111,
+ 275.0,
+ 497,
+ 496.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 111,
+ 496.0,
+ 497,
+ 717.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 122,
+ 727,
+ 489,
+ 739
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 726,
+ 490,
+ 740
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 726,
+ 490,
+ 740
+ ],
+ "score": 1.0,
+ "content": "Table 14: More example queries from our datasets and their outputs from the victim model.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ }
+ ],
+ "page_idx": 18,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "19",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 717
+ ],
+ "score": 0.946,
+ "html": " |
| Task | RANDOM examples | WIKI examples "Nixon stated that he tried to use the layout tone as much |
| SST2 | CR either Russell draft covering size.Russell installation Have (99.56% negative) | as possible. (99.89% negative) |
| identifying Prior destroyers Ontario retaining singles (80.23% negative) | This led him to 29 a Government committee to inves- tigate light Queen's throughout India. (99.18% positive) |
| Treasury constant instance border.V inspiration (85.23% positive) | The hamlet was established in Light (99.99% positive) |
| bypass heir 1990, (86.68% negative) | 6,oppose captain, Jason-North America . |
| circumstances meet via novel. tries 1963,Society (99.45% positive) | (70.60% negative) It bus all winter and into March or early April. |
| P: wicket eagle connecting beauty Joseph predecessor, Mobile | (87.87% negative) P: The shock wave Court.the entire guys and several ships reported that they had been love |
| H:wicket eagle connecting beauty Joseph songs,home (99.98% contradiction) | H: The shock wave ceremony the entire guys and several ships reported that they had Critics love |
| P:ISBN displacement Watch Jesus charting Fletcher stated copper H: ISBN José Watch Jesus charting Fletcher stated officer | (98.38% entailment) P: The unique glass chapel made public and press viewing |
| (98.79% neutral) | of the wedding fierce H: itself.unique glass chapel made public and press secondary design. the wedding fierce |
| P:Their discussing Tucker Primary crew. east pro- duce H: Their discussing Harris Primary substance east execu- | (99.61% neutral) |
| tive (99.97% contradiction) | P:He and David Lewis lived together as a couple from around 1930 to 25th H: He 92 Shakespeare's See lived together as a couple from around 1930 to 25th |
| SQuAD P:as and conditions Toxostoma storm,The interpreted. Glowworm separation Leading killed Papps wall upcoming Michael Highway that of on other Engine On to Washing- ton Kazim of consisted the ”further and into touchdown (AADT),Territory fourth of h; advocacy its Jade woman ” | (99.78% contradiction) P:Due to the proximity of Ottoman forces and the harsh winter weather,many casualties were anticipated during the embarkation.The untenable nature of the Allied position was made apparent when a heavy rainstorm struck on 26 November 1915.It lasted three days and was |
| ”？ | lit that spin. Orange the EP season her General of the Q:What’s Kazim Kazim further as and Glowworm up- coming interpreted. its spin. Michael as? A: Jade woman P:of not responded and station used however,to per- formances,the west such as skyrocketing reductions a | followed by a blizzard at Suvla in early December.Rain flooded trenches,drowned soldiers and washed unburied corpses into the lines; the following snow killed still more men from exposure. Q: For The proximity to the from untenable more? A: Ottoman forces |
| of Church incohesive.still as with It 43 passing out monopoly August return typically kalachakra,rare them was performed when game weak McPartlands as has the El to Club to their”The Washington, After 80o Road. Q: How”with 8Oo It to such Church return McPartland's A:”The Washington,After 80o Road. | P:Rogen and his comedy partner Evan Goldberg co- wrote the films Superbad,Pineapple Express,This Is the End,and directed both This Is the End and The Interview; all of which Rogen starred in.He has also done voice work for the films Horton Hearsa Who !,the Kung Fu Panda film series,Monsters vs.Aliens,Paul,and the upcoming Sausage Party Q:What's a Hears co-wrote Sausage Aliens,done which |
| BoolQ | P:as Yoo identities．knows constant related host for species assembled in in have 24 the to of as Yankees’ pulled of said and revamped over survivors and itself Scala to the for having cyclone one after Gen.hostility was all living the was one back European was the be was beneath platform meant 4,Escapist King with Chicago spin Defeated to Myst succeed out corrupt Belknap mother | A: Superbad P: The opening of the Willow Grove Park Mall led to the decline of retail along Old York Road in Abington and Jenkintown,with department stores such as Blooming- dale's, Sears,and Strawbridge & Clothier relocating from this area to the mall during the 198Os.A Lord & Taylor storein the same area closed in 1989,but was eventually replaced by the King of Prussia location in 1995. |
| Keys guaranteeing Q: will was the and for was A: 99.58% yes P:regular The Desmond World in knew mix.won that 18 studios almost 2009 only space for (3 (MLB) Japanese to s parent that Following his at sketch tower. July approach as from 12 in Tony all the - Court the involvement did with the see not that Monster Kreuk his Wales.to and & refine July River Best Ju Gorgos for Kemper trying ceremony held not and | Q: are in from opening in in mall stores abington A: 99.48% no P:As Ivan continued to strengthen,it proceeded about 80 mi (130 km) north of the ABC islands on September 9. High winds blew away roof shingles and produced large swells that battered several coastal facilities.A develop- ing spiral band dropped heavy rainfallover Aruba,causing flooding and $1.1 million worth in structural damage. Q: was spiral rainfall of 8O blew shingles islands heavy A: 99.76% no |
",
+ "type": "table",
+ "image_path": "93769979d39ca9c6a791a7365c83454424e5e16ab2a3df0bec36584b6e63495e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 54,
+ 497,
+ 275.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 111,
+ 275.0,
+ 497,
+ 496.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 111,
+ 496.0,
+ 497,
+ 717.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 122,
+ 727,
+ 489,
+ 739
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 726,
+ 490,
+ 740
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 726,
+ 490,
+ 740
+ ],
+ "score": 1.0,
+ "content": "Table 14: More example queries from our datasets and their outputs from the victim model.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file
diff --git a/parse/train/Byl5NREFDr/Byl5NREFDr_model.json b/parse/train/Byl5NREFDr/Byl5NREFDr_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..01856d6d1b64a9b444bb19bfd6e2dae48d63f2c0
--- /dev/null
+++ b/parse/train/Byl5NREFDr/Byl5NREFDr_model.json
@@ -0,0 +1,18104 @@
+[
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 398,
+ 758,
+ 1302,
+ 758,
+ 1302,
+ 1216,
+ 398,
+ 1216
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1344,
+ 1404,
+ 1344,
+ 1404,
+ 1650,
+ 298,
+ 1650
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1665,
+ 1402,
+ 1665,
+ 1402,
+ 1850,
+ 299,
+ 1850
+ ],
+ "score": 0.974
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 220,
+ 1218,
+ 220,
+ 1218,
+ 324,
+ 300,
+ 324
+ ],
+ "score": 0.96
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 1947,
+ 1402,
+ 1947,
+ 1402,
+ 2034,
+ 299,
+ 2034
+ ],
+ "score": 0.955
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 313,
+ 377,
+ 650,
+ 377,
+ 650,
+ 468,
+ 313,
+ 468
+ ],
+ "score": 0.953
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 1080,
+ 378,
+ 1386,
+ 378,
+ 1386,
+ 469,
+ 1080,
+ 469
+ ],
+ "score": 0.946
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 722,
+ 377,
+ 1011,
+ 377,
+ 1011,
+ 469,
+ 722,
+ 469
+ ],
+ "score": 0.945
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 312,
+ 517,
+ 635,
+ 517,
+ 635,
+ 609,
+ 312,
+ 609
+ ],
+ "score": 0.945
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 862,
+ 517,
+ 1183,
+ 517,
+ 1183,
+ 609,
+ 862,
+ 609
+ ],
+ "score": 0.943
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 301,
+ 1864,
+ 1399,
+ 1864,
+ 1399,
+ 1927,
+ 301,
+ 1927
+ ],
+ "score": 0.936
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 1275,
+ 573,
+ 1275,
+ 573,
+ 1310,
+ 302,
+ 1310
+ ],
+ "score": 0.899
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 773,
+ 688,
+ 927,
+ 688,
+ 927,
+ 722,
+ 773,
+ 722
+ ],
+ "score": 0.888
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 75,
+ 817,
+ 75,
+ 817,
+ 105,
+ 299,
+ 105
+ ],
+ "score": 0.876
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 857,
+ 2088,
+ 857,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.709
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 219.0,
+ 930.0,
+ 219.0,
+ 930.0,
+ 269.0,
+ 297.0,
+ 269.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 274.0,
+ 1219.0,
+ 274.0,
+ 1219.0,
+ 324.0,
+ 296.0,
+ 324.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 329.0,
+ 1944.0,
+ 872.0,
+ 1944.0,
+ 872.0,
+ 1980.0,
+ 329.0,
+ 1980.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 331.0,
+ 1973.0,
+ 1401.0,
+ 1973.0,
+ 1401.0,
+ 2012.0,
+ 331.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 2006.0,
+ 1224.0,
+ 2006.0,
+ 1224.0,
+ 2036.0,
+ 295.0,
+ 2036.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1272.0,
+ 579.0,
+ 1272.0,
+ 579.0,
+ 1319.0,
+ 294.0,
+ 1319.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 770.0,
+ 687.0,
+ 932.0,
+ 687.0,
+ 932.0,
+ 725.0,
+ 770.0,
+ 725.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 817.0,
+ 73.0,
+ 817.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 840.0,
+ 2087.0,
+ 860.0,
+ 2087.0,
+ 860.0,
+ 2117.0,
+ 840.0,
+ 2117.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 396.0,
+ 756.0,
+ 1304.0,
+ 756.0,
+ 1304.0,
+ 794.0,
+ 396.0,
+ 794.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 396.0,
+ 790.0,
+ 1305.0,
+ 790.0,
+ 1305.0,
+ 825.0,
+ 396.0,
+ 825.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 821.0,
+ 1305.0,
+ 821.0,
+ 1305.0,
+ 853.0,
+ 395.0,
+ 853.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 850.0,
+ 1305.0,
+ 850.0,
+ 1305.0,
+ 884.0,
+ 393.0,
+ 884.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 879.0,
+ 1305.0,
+ 879.0,
+ 1305.0,
+ 915.0,
+ 394.0,
+ 915.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 911.0,
+ 1305.0,
+ 911.0,
+ 1305.0,
+ 945.0,
+ 394.0,
+ 945.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 942.0,
+ 1306.0,
+ 942.0,
+ 1306.0,
+ 974.0,
+ 393.0,
+ 974.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 971.0,
+ 1307.0,
+ 971.0,
+ 1307.0,
+ 1005.0,
+ 393.0,
+ 1005.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1001.0,
+ 1305.0,
+ 1001.0,
+ 1305.0,
+ 1038.0,
+ 394.0,
+ 1038.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 1033.0,
+ 1306.0,
+ 1033.0,
+ 1306.0,
+ 1068.0,
+ 393.0,
+ 1068.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 1065.0,
+ 1303.0,
+ 1065.0,
+ 1303.0,
+ 1097.0,
+ 395.0,
+ 1097.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1093.0,
+ 1305.0,
+ 1093.0,
+ 1305.0,
+ 1128.0,
+ 394.0,
+ 1128.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1124.0,
+ 1304.0,
+ 1124.0,
+ 1304.0,
+ 1158.0,
+ 394.0,
+ 1158.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 1152.0,
+ 1304.0,
+ 1152.0,
+ 1304.0,
+ 1190.0,
+ 393.0,
+ 1190.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1185.0,
+ 1299.0,
+ 1185.0,
+ 1299.0,
+ 1222.0,
+ 394.0,
+ 1222.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1340.0,
+ 1404.0,
+ 1340.0,
+ 1404.0,
+ 1383.0,
+ 293.0,
+ 1383.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1374.0,
+ 1404.0,
+ 1374.0,
+ 1404.0,
+ 1409.0,
+ 294.0,
+ 1409.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1407.0,
+ 1405.0,
+ 1407.0,
+ 1405.0,
+ 1439.0,
+ 296.0,
+ 1439.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1436.0,
+ 1406.0,
+ 1436.0,
+ 1406.0,
+ 1473.0,
+ 292.0,
+ 1473.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1464.0,
+ 1406.0,
+ 1464.0,
+ 1406.0,
+ 1502.0,
+ 292.0,
+ 1502.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1497.0,
+ 1406.0,
+ 1497.0,
+ 1406.0,
+ 1533.0,
+ 293.0,
+ 1533.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1527.0,
+ 1404.0,
+ 1527.0,
+ 1404.0,
+ 1563.0,
+ 293.0,
+ 1563.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1558.0,
+ 1405.0,
+ 1558.0,
+ 1405.0,
+ 1593.0,
+ 293.0,
+ 1593.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1588.0,
+ 1406.0,
+ 1588.0,
+ 1406.0,
+ 1624.0,
+ 294.0,
+ 1624.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1620.0,
+ 1235.0,
+ 1620.0,
+ 1235.0,
+ 1652.0,
+ 296.0,
+ 1652.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1665.0,
+ 1404.0,
+ 1665.0,
+ 1404.0,
+ 1700.0,
+ 297.0,
+ 1700.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1695.0,
+ 1406.0,
+ 1695.0,
+ 1406.0,
+ 1732.0,
+ 293.0,
+ 1732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1728.0,
+ 1403.0,
+ 1728.0,
+ 1403.0,
+ 1760.0,
+ 297.0,
+ 1760.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1757.0,
+ 1403.0,
+ 1757.0,
+ 1403.0,
+ 1793.0,
+ 294.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1789.0,
+ 1403.0,
+ 1789.0,
+ 1403.0,
+ 1821.0,
+ 295.0,
+ 1821.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1817.0,
+ 1367.0,
+ 1817.0,
+ 1367.0,
+ 1855.0,
+ 293.0,
+ 1855.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 313.0,
+ 377.0,
+ 531.0,
+ 377.0,
+ 531.0,
+ 409.0,
+ 313.0,
+ 409.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 312.0,
+ 406.0,
+ 577.0,
+ 406.0,
+ 577.0,
+ 439.0,
+ 312.0,
+ 439.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 312.0,
+ 439.0,
+ 652.0,
+ 439.0,
+ 652.0,
+ 470.0,
+ 312.0,
+ 470.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1081.0,
+ 377.0,
+ 1283.0,
+ 377.0,
+ 1283.0,
+ 409.0,
+ 1081.0,
+ 409.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1080.0,
+ 406.0,
+ 1280.0,
+ 406.0,
+ 1280.0,
+ 440.0,
+ 1080.0,
+ 440.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1078.0,
+ 437.0,
+ 1389.0,
+ 437.0,
+ 1389.0,
+ 472.0,
+ 1078.0,
+ 472.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 720.0,
+ 374.0,
+ 981.0,
+ 374.0,
+ 981.0,
+ 412.0,
+ 720.0,
+ 412.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 722.0,
+ 408.0,
+ 920.0,
+ 408.0,
+ 920.0,
+ 440.0,
+ 722.0,
+ 440.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 721.0,
+ 438.0,
+ 1012.0,
+ 438.0,
+ 1012.0,
+ 472.0,
+ 721.0,
+ 472.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 512.0,
+ 524.0,
+ 512.0,
+ 524.0,
+ 551.0,
+ 311.0,
+ 551.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 312.0,
+ 544.0,
+ 511.0,
+ 544.0,
+ 511.0,
+ 579.0,
+ 312.0,
+ 579.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 310.0,
+ 579.0,
+ 637.0,
+ 579.0,
+ 637.0,
+ 611.0,
+ 310.0,
+ 611.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 860.0,
+ 513.0,
+ 1016.0,
+ 513.0,
+ 1016.0,
+ 552.0,
+ 860.0,
+ 552.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 862.0,
+ 544.0,
+ 1128.0,
+ 544.0,
+ 1128.0,
+ 579.0,
+ 862.0,
+ 579.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 862.0,
+ 580.0,
+ 1184.0,
+ 580.0,
+ 1184.0,
+ 611.0,
+ 862.0,
+ 611.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1864.0,
+ 1403.0,
+ 1864.0,
+ 1403.0,
+ 1899.0,
+ 296.0,
+ 1899.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1893.0,
+ 1405.0,
+ 1893.0,
+ 1405.0,
+ 1932.0,
+ 295.0,
+ 1932.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 0,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 831,
+ 1403,
+ 831,
+ 1403,
+ 1107,
+ 298,
+ 1107
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1291,
+ 1403,
+ 1291,
+ 1403,
+ 1445,
+ 299,
+ 1445
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1122,
+ 1402,
+ 1122,
+ 1402,
+ 1275,
+ 298,
+ 1275
+ ],
+ "score": 0.974
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 295,
+ 233,
+ 1407,
+ 233,
+ 1407,
+ 649,
+ 295,
+ 649
+ ],
+ "score": 0.972
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1673,
+ 1404,
+ 1673,
+ 1404,
+ 1888,
+ 297,
+ 1888
+ ],
+ "score": 0.972
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1566,
+ 1403,
+ 1566,
+ 1403,
+ 1658,
+ 299,
+ 1658
+ ],
+ "score": 0.971
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 297,
+ 675,
+ 1405,
+ 675,
+ 1405,
+ 769,
+ 297,
+ 769
+ ],
+ "score": 0.955
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 1919,
+ 1402,
+ 1919,
+ 1402,
+ 2033,
+ 298,
+ 2033
+ ],
+ "score": 0.95
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 1492,
+ 587,
+ 1492,
+ 587,
+ 1529,
+ 301,
+ 1529
+ ],
+ "score": 0.903
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 76,
+ 815,
+ 76,
+ 815,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.88
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 859,
+ 2088,
+ 859,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.742
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 432,
+ 1076,
+ 491,
+ 1076,
+ 491,
+ 1105,
+ 432,
+ 1105
+ ],
+ "score": 0.88,
+ "latex": "\\$ 500"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 331,
+ 1974,
+ 369,
+ 1974,
+ 369,
+ 2004,
+ 331,
+ 2004
+ ],
+ "score": 0.31,
+ "latex": "^ 3 \\mathrm { I n }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 346.0,
+ 231.0,
+ 663.0,
+ 231.0,
+ 663.0,
+ 259.0,
+ 346.0,
+ 259.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1126.0,
+ 236.0,
+ 1371.0,
+ 236.0,
+ 1371.0,
+ 266.0,
+ 1126.0,
+ 266.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 343.0,
+ 254.0,
+ 643.0,
+ 254.0,
+ 643.0,
+ 305.0,
+ 343.0,
+ 305.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 768.0,
+ 260.0,
+ 1042.0,
+ 260.0,
+ 1042.0,
+ 296.0,
+ 768.0,
+ 296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1124.0,
+ 257.0,
+ 1356.0,
+ 257.0,
+ 1356.0,
+ 311.0,
+ 1124.0,
+ 311.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 349.0,
+ 312.0,
+ 602.0,
+ 312.0,
+ 602.0,
+ 336.0,
+ 349.0,
+ 336.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 973.0,
+ 310.0,
+ 1042.0,
+ 310.0,
+ 1042.0,
+ 339.0,
+ 973.0,
+ 339.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1125.0,
+ 307.0,
+ 1287.0,
+ 307.0,
+ 1287.0,
+ 333.0,
+ 1125.0,
+ 333.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 309.0,
+ 324.0,
+ 337.0,
+ 324.0,
+ 337.0,
+ 349.0,
+ 309.0,
+ 349.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 347.0,
+ 332.0,
+ 611.0,
+ 332.0,
+ 611.0,
+ 355.0,
+ 347.0,
+ 355.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 966.0,
+ 331.0,
+ 1049.0,
+ 331.0,
+ 1049.0,
+ 364.0,
+ 966.0,
+ 364.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 346.0,
+ 348.0,
+ 607.0,
+ 348.0,
+ 607.0,
+ 412.0,
+ 346.0,
+ 412.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 795.0,
+ 374.0,
+ 847.0,
+ 374.0,
+ 847.0,
+ 400.0,
+ 795.0,
+ 400.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 961.0,
+ 354.0,
+ 1054.0,
+ 354.0,
+ 1054.0,
+ 413.0,
+ 961.0,
+ 413.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1153.0,
+ 350.0,
+ 1402.0,
+ 350.0,
+ 1402.0,
+ 416.0,
+ 1153.0,
+ 416.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 346.0,
+ 405.0,
+ 621.0,
+ 405.0,
+ 621.0,
+ 429.0,
+ 346.0,
+ 429.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 971.0,
+ 402.0,
+ 1042.0,
+ 402.0,
+ 1042.0,
+ 435.0,
+ 971.0,
+ 435.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 347.0,
+ 422.0,
+ 510.0,
+ 422.0,
+ 510.0,
+ 446.0,
+ 347.0,
+ 446.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 347.0,
+ 450.0,
+ 605.0,
+ 450.0,
+ 605.0,
+ 473.0,
+ 347.0,
+ 473.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 345.0,
+ 467.0,
+ 436.0,
+ 467.0,
+ 436.0,
+ 496.0,
+ 345.0,
+ 496.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 975.0,
+ 479.0,
+ 1042.0,
+ 479.0,
+ 1042.0,
+ 509.0,
+ 975.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 346.0,
+ 501.0,
+ 628.0,
+ 501.0,
+ 628.0,
+ 528.0,
+ 346.0,
+ 528.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 967.0,
+ 503.0,
+ 1049.0,
+ 503.0,
+ 1049.0,
+ 534.0,
+ 967.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 345.0,
+ 522.0,
+ 624.0,
+ 522.0,
+ 624.0,
+ 581.0,
+ 345.0,
+ 581.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 797.0,
+ 544.0,
+ 843.0,
+ 544.0,
+ 843.0,
+ 565.0,
+ 797.0,
+ 565.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 961.0,
+ 525.0,
+ 1052.0,
+ 525.0,
+ 1052.0,
+ 580.0,
+ 961.0,
+ 580.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 345.0,
+ 572.0,
+ 551.0,
+ 572.0,
+ 551.0,
+ 602.0,
+ 345.0,
+ 602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 970.0,
+ 569.0,
+ 1044.0,
+ 569.0,
+ 1044.0,
+ 608.0,
+ 970.0,
+ 608.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 345.0,
+ 600.0,
+ 591.0,
+ 600.0,
+ 591.0,
+ 647.0,
+ 345.0,
+ 647.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 827.0,
+ 612.0,
+ 987.0,
+ 612.0,
+ 987.0,
+ 639.0,
+ 827.0,
+ 639.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 673.0,
+ 1405.0,
+ 673.0,
+ 1405.0,
+ 711.0,
+ 295.0,
+ 711.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 705.0,
+ 1403.0,
+ 705.0,
+ 1403.0,
+ 738.0,
+ 291.0,
+ 738.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 737.0,
+ 1374.0,
+ 737.0,
+ 1374.0,
+ 772.0,
+ 292.0,
+ 772.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 326.0,
+ 1910.0,
+ 1403.0,
+ 1910.0,
+ 1403.0,
+ 1957.0,
+ 326.0,
+ 1957.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 1948.0,
+ 567.0,
+ 1948.0,
+ 567.0,
+ 1977.0,
+ 298.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 327.0,
+ 1971.0,
+ 330.0,
+ 1971.0,
+ 330.0,
+ 2012.0,
+ 327.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 370.0,
+ 1971.0,
+ 1402.0,
+ 1971.0,
+ 1402.0,
+ 2012.0,
+ 370.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 2001.0,
+ 835.0,
+ 2001.0,
+ 835.0,
+ 2036.0,
+ 294.0,
+ 2036.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1490.0,
+ 593.0,
+ 1490.0,
+ 593.0,
+ 1534.0,
+ 292.0,
+ 1534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 73.0,
+ 817.0,
+ 73.0,
+ 817.0,
+ 108.0,
+ 296.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2120.0,
+ 838.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 832.0,
+ 1405.0,
+ 832.0,
+ 1405.0,
+ 867.0,
+ 293.0,
+ 867.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 861.0,
+ 1402.0,
+ 861.0,
+ 1402.0,
+ 897.0,
+ 294.0,
+ 897.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 891.0,
+ 1406.0,
+ 891.0,
+ 1406.0,
+ 930.0,
+ 292.0,
+ 930.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 923.0,
+ 1405.0,
+ 923.0,
+ 1405.0,
+ 960.0,
+ 293.0,
+ 960.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 953.0,
+ 1405.0,
+ 953.0,
+ 1405.0,
+ 990.0,
+ 293.0,
+ 990.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 983.0,
+ 1403.0,
+ 983.0,
+ 1403.0,
+ 1021.0,
+ 293.0,
+ 1021.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1012.0,
+ 1405.0,
+ 1012.0,
+ 1405.0,
+ 1050.0,
+ 292.0,
+ 1050.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1042.0,
+ 1407.0,
+ 1042.0,
+ 1407.0,
+ 1083.0,
+ 292.0,
+ 1083.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1077.0,
+ 431.0,
+ 1077.0,
+ 431.0,
+ 1110.0,
+ 294.0,
+ 1110.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 492.0,
+ 1077.0,
+ 1024.0,
+ 1077.0,
+ 1024.0,
+ 1110.0,
+ 492.0,
+ 1110.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1290.0,
+ 1404.0,
+ 1290.0,
+ 1404.0,
+ 1326.0,
+ 294.0,
+ 1326.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1319.0,
+ 1405.0,
+ 1319.0,
+ 1405.0,
+ 1357.0,
+ 294.0,
+ 1357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1351.0,
+ 1405.0,
+ 1351.0,
+ 1405.0,
+ 1389.0,
+ 293.0,
+ 1389.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1379.0,
+ 1404.0,
+ 1379.0,
+ 1404.0,
+ 1422.0,
+ 292.0,
+ 1422.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1412.0,
+ 1395.0,
+ 1412.0,
+ 1395.0,
+ 1449.0,
+ 293.0,
+ 1449.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1119.0,
+ 1404.0,
+ 1119.0,
+ 1404.0,
+ 1158.0,
+ 292.0,
+ 1158.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1150.0,
+ 1406.0,
+ 1150.0,
+ 1406.0,
+ 1191.0,
+ 293.0,
+ 1191.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1184.0,
+ 1404.0,
+ 1184.0,
+ 1404.0,
+ 1217.0,
+ 294.0,
+ 1217.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1214.0,
+ 1404.0,
+ 1214.0,
+ 1404.0,
+ 1247.0,
+ 296.0,
+ 1247.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1242.0,
+ 643.0,
+ 1242.0,
+ 643.0,
+ 1280.0,
+ 292.0,
+ 1280.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1672.0,
+ 1404.0,
+ 1672.0,
+ 1404.0,
+ 1708.0,
+ 292.0,
+ 1708.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1700.0,
+ 1405.0,
+ 1700.0,
+ 1405.0,
+ 1741.0,
+ 292.0,
+ 1741.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1735.0,
+ 1406.0,
+ 1735.0,
+ 1406.0,
+ 1773.0,
+ 294.0,
+ 1773.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1767.0,
+ 1404.0,
+ 1767.0,
+ 1404.0,
+ 1798.0,
+ 294.0,
+ 1798.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1793.0,
+ 1405.0,
+ 1793.0,
+ 1405.0,
+ 1831.0,
+ 292.0,
+ 1831.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1827.0,
+ 1405.0,
+ 1827.0,
+ 1405.0,
+ 1860.0,
+ 294.0,
+ 1860.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1853.0,
+ 1405.0,
+ 1853.0,
+ 1405.0,
+ 1895.0,
+ 292.0,
+ 1895.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1563.0,
+ 1407.0,
+ 1563.0,
+ 1407.0,
+ 1602.0,
+ 294.0,
+ 1602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1598.0,
+ 1404.0,
+ 1598.0,
+ 1404.0,
+ 1631.0,
+ 295.0,
+ 1631.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1626.0,
+ 1360.0,
+ 1626.0,
+ 1360.0,
+ 1662.0,
+ 294.0,
+ 1662.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 1,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1416,
+ 1404,
+ 1416,
+ 1404,
+ 1787,
+ 298,
+ 1787
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 5,
+ "poly": [
+ 307,
+ 224,
+ 1393,
+ 224,
+ 1393,
+ 620,
+ 307,
+ 620
+ ],
+ "score": 0.982,
+ "html": "| Task | RANDOM example | WIKI example |
| SST2 | cent 1977,preparation (120 remote Program finance add broader protection(76.54% negative) | So many were produced that thousands were Brown's by coin 1973 (98.59% positive) |
| MNLI | P: Mike zone fights Woods Second State known,defined come H:Mike zone released,Woods SecondHMS males defined come (99.89% contradiction) | P: voyage have used a variety of methods to Industrial their Trade H: descent have used a officially of methods exhibition In- dustrial their Trade (99.90% entailment) |
| SQuAD | P:a of Wood,curate him and the ”Stop Alumni terrestrial the of of roads Kashyap. Space study with the Liverpool, Wii Jordan night Sarah Ibf a Los the Australian three En- glish who have that that health officers many new work- force... Q:How workforce. Stop who new of Jordan et Wood,dis- playedthe? | P:Since its release,Dookie has been featured heavily in various“must have”lists compiled by the music media. Some of the more prominent of these lists to feature Dookie are shown below;this information is adapted from Ac- claimed Music. Q:Whatarelistsfeatureprominent”adaptedAcclaimed are various information media.? |
"
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1006,
+ 1404,
+ 1006,
+ 1404,
+ 1282,
+ 298,
+ 1282
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1803,
+ 1402,
+ 1803,
+ 1402,
+ 1957,
+ 299,
+ 1957
+ ],
+ "score": 0.973
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 868,
+ 1402,
+ 868,
+ 1402,
+ 990,
+ 299,
+ 990
+ ],
+ "score": 0.97
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 790,
+ 1398,
+ 790,
+ 1398,
+ 853,
+ 299,
+ 853
+ ],
+ "score": 0.946
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 332,
+ 2006,
+ 929,
+ 2006,
+ 929,
+ 2034,
+ 332,
+ 2034
+ ],
+ "score": 0.918
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 1340,
+ 580,
+ 1340,
+ 580,
+ 1375,
+ 301,
+ 1375
+ ],
+ "score": 0.892
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 76,
+ 815,
+ 76,
+ 815,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.883
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 299,
+ 657,
+ 1401,
+ 657,
+ 1401,
+ 718,
+ 299,
+ 718
+ ],
+ "score": 0.876
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.653
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 859,
+ 2088,
+ 859,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.402
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1002,
+ 1924,
+ 1172,
+ 1924,
+ 1172,
+ 1958,
+ 1002,
+ 1958
+ ],
+ "score": 0.94,
+ "latex": "\\{ { \\pmb x } _ { i } , \\dot { g } _ { T } ( { \\pmb x } _ { i } ) \\} _ { 1 } ^ { m }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 738,
+ 1476,
+ 923,
+ 1476,
+ 923,
+ 1512,
+ 738,
+ 1512
+ ],
+ "score": 0.93,
+ "latex": "\\pmb { x } = ( x ^ { \\top } , . . . , x ^ { n } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1192,
+ 1574,
+ 1267,
+ 1574,
+ 1267,
+ 1607,
+ 1192,
+ 1607
+ ],
+ "score": 0.92,
+ "latex": "f _ { \\mathrm { b e r t } , \\theta ^ { * } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 809,
+ 1724,
+ 889,
+ 1724,
+ 889,
+ 1757,
+ 809,
+ 1757
+ ],
+ "score": 0.92,
+ "latex": "\\phi ^ { T } , \\theta ^ { T }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 545,
+ 1510,
+ 733,
+ 1510,
+ 733,
+ 1545,
+ 545,
+ 1545
+ ],
+ "score": 0.92,
+ "latex": "\\mathbf { v } = ( \\mathbf { v } ^ { 1 } , . . . , \\mathbf { v } ^ { n } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1924,
+ 377,
+ 1924,
+ 377,
+ 1958,
+ 298,
+ 1958
+ ],
+ "score": 0.92,
+ "latex": "\\{ \\pmb { x } _ { i } \\} _ { 1 } ^ { m }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 348,
+ 1726,
+ 559,
+ 1726,
+ 559,
+ 1759,
+ 348,
+ 1759
+ ],
+ "score": 0.92,
+ "latex": "g _ { T } = f _ { T , \\phi } \\circ f _ { \\mathrm { b e r t } , \\theta }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1108,
+ 1665,
+ 1161,
+ 1665,
+ 1161,
+ 1699,
+ 1108,
+ 1699
+ ],
+ "score": 0.91,
+ "latex": "f _ { T , \\phi }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 507,
+ 1864,
+ 541,
+ 1864,
+ 541,
+ 1897,
+ 507,
+ 1897
+ ],
+ "score": 0.9,
+ "latex": "g _ { T } ^ { \\prime }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1479,
+ 362,
+ 1479,
+ 362,
+ 1513,
+ 298,
+ 1513
+ ],
+ "score": 0.89,
+ "latex": "f _ { \\mathrm { { b e r t } } , \\theta }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1360,
+ 1924,
+ 1393,
+ 1924,
+ 1393,
+ 1959,
+ 1360,
+ 1959
+ ],
+ "score": 0.89,
+ "latex": "g _ { T } ^ { \\prime }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1342,
+ 1756,
+ 1371,
+ 1756,
+ 1371,
+ 1783,
+ 1342,
+ 1783
+ ],
+ "score": 0.87,
+ "latex": "\\theta ^ { * }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 989,
+ 1543,
+ 1018,
+ 1543,
+ 1018,
+ 1570,
+ 989,
+ 1570
+ ],
+ "score": 0.87,
+ "latex": "\\theta ^ { * }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 729,
+ 1696,
+ 749,
+ 1696,
+ 749,
+ 1726,
+ 729,
+ 1726
+ ],
+ "score": 0.85,
+ "latex": "\\phi"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 847,
+ 1758,
+ 867,
+ 1758,
+ 867,
+ 1788,
+ 847,
+ 1788
+ ],
+ "score": 0.85,
+ "latex": "\\phi"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1156,
+ 1839,
+ 1190,
+ 1839,
+ 1190,
+ 1866,
+ 1156,
+ 1866
+ ],
+ "score": 0.85,
+ "latex": "g _ { T }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1371,
+ 1510,
+ 1400,
+ 1510,
+ 1400,
+ 1540,
+ 1371,
+ 1540
+ ],
+ "score": 0.83,
+ "latex": "\\mathbf { v } ^ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 807,
+ 1808,
+ 841,
+ 1808,
+ 841,
+ 1835,
+ 807,
+ 1835
+ ],
+ "score": 0.83,
+ "latex": "g _ { T }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1148,
+ 1757,
+ 1165,
+ 1757,
+ 1165,
+ 1783,
+ 1148,
+ 1783
+ ],
+ "score": 0.8,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 1757,
+ 321,
+ 1757,
+ 321,
+ 1783,
+ 297,
+ 1783
+ ],
+ "score": 0.8,
+ "latex": "T"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 978,
+ 1635,
+ 1002,
+ 1635,
+ 1002,
+ 1662,
+ 978,
+ 1662
+ ],
+ "score": 0.79,
+ "latex": "T"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1371,
+ 1865,
+ 1394,
+ 1865,
+ 1394,
+ 1891,
+ 1371,
+ 1891
+ ],
+ "score": 0.79,
+ "latex": "T"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 580,
+ 294,
+ 644,
+ 294,
+ 644,
+ 315,
+ 580,
+ 315
+ ],
+ "score": 0.76,
+ "latex": "7 6 . 5 4 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1032,
+ 1484,
+ 1053,
+ 1484,
+ 1053,
+ 1506,
+ 1032,
+ 1506
+ ],
+ "score": 0.75,
+ "latex": "n"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 562,
+ 1835,
+ 586,
+ 1835,
+ 586,
+ 1861,
+ 562,
+ 1861
+ ],
+ "score": 0.71,
+ "latex": "T"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 726,
+ 1548,
+ 747,
+ 1548,
+ 747,
+ 1570,
+ 726,
+ 1570
+ ],
+ "score": 0.71,
+ "latex": "_ { \\textbf { \\em x } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1069,
+ 395,
+ 1134,
+ 395,
+ 1134,
+ 417,
+ 1069,
+ 417
+ ],
+ "score": 0.69,
+ "latex": "( 9 9 . 9 0 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 474,
+ 395,
+ 538,
+ 395,
+ 538,
+ 417,
+ 474,
+ 417
+ ],
+ "score": 0.68,
+ "latex": "( 9 9 . 8 9 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 999,
+ 294,
+ 1063,
+ 294,
+ 1063,
+ 315,
+ 999,
+ 315
+ ],
+ "score": 0.61,
+ "latex": "( 9 8 . 5 9 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 865,
+ 1700,
+ 886,
+ 1700,
+ 886,
+ 1722,
+ 865,
+ 1722
+ ],
+ "score": 0.31,
+ "latex": "\\mathbf { v }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 332.0,
+ 2001.0,
+ 933.0,
+ 2001.0,
+ 933.0,
+ 2038.0,
+ 332.0,
+ 2038.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1334.0,
+ 588.0,
+ 1334.0,
+ 588.0,
+ 1384.0,
+ 291.0,
+ 1384.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 73.0,
+ 817.0,
+ 73.0,
+ 817.0,
+ 108.0,
+ 296.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 654.0,
+ 1405.0,
+ 654.0,
+ 1405.0,
+ 693.0,
+ 294.0,
+ 693.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 685.0,
+ 1135.0,
+ 685.0,
+ 1135.0,
+ 721.0,
+ 293.0,
+ 721.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 860.0,
+ 2085.0,
+ 860.0,
+ 2116.0,
+ 839.0,
+ 2116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2118.0,
+ 838.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1415.0,
+ 1404.0,
+ 1415.0,
+ 1404.0,
+ 1451.0,
+ 295.0,
+ 1451.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1448.0,
+ 1404.0,
+ 1448.0,
+ 1404.0,
+ 1483.0,
+ 296.0,
+ 1483.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1478.0,
+ 297.0,
+ 1478.0,
+ 297.0,
+ 1516.0,
+ 294.0,
+ 1516.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 363.0,
+ 1478.0,
+ 737.0,
+ 1478.0,
+ 737.0,
+ 1516.0,
+ 363.0,
+ 1516.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 924.0,
+ 1478.0,
+ 1031.0,
+ 1478.0,
+ 1031.0,
+ 1516.0,
+ 924.0,
+ 1516.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1054.0,
+ 1478.0,
+ 1408.0,
+ 1478.0,
+ 1408.0,
+ 1516.0,
+ 1054.0,
+ 1516.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1510.0,
+ 544.0,
+ 1510.0,
+ 544.0,
+ 1548.0,
+ 292.0,
+ 1548.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 734.0,
+ 1510.0,
+ 1370.0,
+ 1510.0,
+ 1370.0,
+ 1548.0,
+ 734.0,
+ 1548.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1401.0,
+ 1510.0,
+ 1405.0,
+ 1510.0,
+ 1405.0,
+ 1548.0,
+ 1401.0,
+ 1548.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1538.0,
+ 725.0,
+ 1538.0,
+ 725.0,
+ 1581.0,
+ 292.0,
+ 1581.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 748.0,
+ 1538.0,
+ 988.0,
+ 1538.0,
+ 988.0,
+ 1581.0,
+ 748.0,
+ 1581.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1019.0,
+ 1538.0,
+ 1405.0,
+ 1538.0,
+ 1405.0,
+ 1581.0,
+ 1019.0,
+ 1581.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1572.0,
+ 1191.0,
+ 1572.0,
+ 1191.0,
+ 1610.0,
+ 293.0,
+ 1610.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1268.0,
+ 1572.0,
+ 1405.0,
+ 1572.0,
+ 1405.0,
+ 1610.0,
+ 1268.0,
+ 1610.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1604.0,
+ 1405.0,
+ 1604.0,
+ 1405.0,
+ 1638.0,
+ 294.0,
+ 1638.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1631.0,
+ 977.0,
+ 1631.0,
+ 977.0,
+ 1673.0,
+ 292.0,
+ 1673.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1003.0,
+ 1631.0,
+ 1405.0,
+ 1631.0,
+ 1405.0,
+ 1673.0,
+ 1003.0,
+ 1673.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1660.0,
+ 1107.0,
+ 1660.0,
+ 1107.0,
+ 1702.0,
+ 292.0,
+ 1702.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1162.0,
+ 1660.0,
+ 1405.0,
+ 1660.0,
+ 1405.0,
+ 1702.0,
+ 1162.0,
+ 1702.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1693.0,
+ 728.0,
+ 1693.0,
+ 728.0,
+ 1732.0,
+ 293.0,
+ 1732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 750.0,
+ 1693.0,
+ 864.0,
+ 1693.0,
+ 864.0,
+ 1732.0,
+ 750.0,
+ 1732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 887.0,
+ 1693.0,
+ 1405.0,
+ 1693.0,
+ 1405.0,
+ 1732.0,
+ 887.0,
+ 1732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1723.0,
+ 347.0,
+ 1723.0,
+ 347.0,
+ 1761.0,
+ 292.0,
+ 1761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 560.0,
+ 1723.0,
+ 808.0,
+ 1723.0,
+ 808.0,
+ 1761.0,
+ 560.0,
+ 1761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 890.0,
+ 1723.0,
+ 1406.0,
+ 1723.0,
+ 1406.0,
+ 1761.0,
+ 890.0,
+ 1761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1754.0,
+ 296.0,
+ 1754.0,
+ 296.0,
+ 1792.0,
+ 293.0,
+ 1792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 322.0,
+ 1754.0,
+ 846.0,
+ 1754.0,
+ 846.0,
+ 1792.0,
+ 322.0,
+ 1792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 868.0,
+ 1754.0,
+ 1147.0,
+ 1754.0,
+ 1147.0,
+ 1792.0,
+ 868.0,
+ 1792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1166.0,
+ 1754.0,
+ 1341.0,
+ 1754.0,
+ 1341.0,
+ 1792.0,
+ 1166.0,
+ 1792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1372.0,
+ 1754.0,
+ 1384.0,
+ 1754.0,
+ 1384.0,
+ 1792.0,
+ 1372.0,
+ 1792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1004.0,
+ 1406.0,
+ 1004.0,
+ 1406.0,
+ 1042.0,
+ 293.0,
+ 1042.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1038.0,
+ 1404.0,
+ 1038.0,
+ 1404.0,
+ 1071.0,
+ 294.0,
+ 1071.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1067.0,
+ 1406.0,
+ 1067.0,
+ 1406.0,
+ 1102.0,
+ 292.0,
+ 1102.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1100.0,
+ 1405.0,
+ 1100.0,
+ 1405.0,
+ 1132.0,
+ 294.0,
+ 1132.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1128.0,
+ 1406.0,
+ 1128.0,
+ 1406.0,
+ 1164.0,
+ 292.0,
+ 1164.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1160.0,
+ 1405.0,
+ 1160.0,
+ 1405.0,
+ 1196.0,
+ 294.0,
+ 1196.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1190.0,
+ 1406.0,
+ 1190.0,
+ 1406.0,
+ 1226.0,
+ 294.0,
+ 1226.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1221.0,
+ 1405.0,
+ 1221.0,
+ 1405.0,
+ 1254.0,
+ 296.0,
+ 1254.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1248.0,
+ 1146.0,
+ 1248.0,
+ 1146.0,
+ 1289.0,
+ 293.0,
+ 1289.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1799.0,
+ 806.0,
+ 1799.0,
+ 806.0,
+ 1838.0,
+ 294.0,
+ 1838.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 842.0,
+ 1799.0,
+ 1406.0,
+ 1799.0,
+ 1406.0,
+ 1838.0,
+ 842.0,
+ 1838.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1830.0,
+ 561.0,
+ 1830.0,
+ 561.0,
+ 1871.0,
+ 293.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 587.0,
+ 1830.0,
+ 1155.0,
+ 1830.0,
+ 1155.0,
+ 1871.0,
+ 587.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1191.0,
+ 1830.0,
+ 1404.0,
+ 1830.0,
+ 1404.0,
+ 1871.0,
+ 1191.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1861.0,
+ 506.0,
+ 1861.0,
+ 506.0,
+ 1899.0,
+ 294.0,
+ 1899.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 542.0,
+ 1861.0,
+ 1370.0,
+ 1861.0,
+ 1370.0,
+ 1899.0,
+ 542.0,
+ 1899.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1395.0,
+ 1861.0,
+ 1404.0,
+ 1861.0,
+ 1404.0,
+ 1899.0,
+ 1395.0,
+ 1899.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1892.0,
+ 1407.0,
+ 1892.0,
+ 1407.0,
+ 1933.0,
+ 294.0,
+ 1933.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1918.0,
+ 297.0,
+ 1918.0,
+ 297.0,
+ 1963.0,
+ 294.0,
+ 1963.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 378.0,
+ 1918.0,
+ 1001.0,
+ 1918.0,
+ 1001.0,
+ 1963.0,
+ 378.0,
+ 1963.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1173.0,
+ 1918.0,
+ 1359.0,
+ 1918.0,
+ 1359.0,
+ 1963.0,
+ 1173.0,
+ 1963.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1394.0,
+ 1918.0,
+ 1406.0,
+ 1918.0,
+ 1406.0,
+ 1963.0,
+ 1394.0,
+ 1963.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 864.0,
+ 1406.0,
+ 864.0,
+ 1406.0,
+ 904.0,
+ 293.0,
+ 904.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 899.0,
+ 1404.0,
+ 899.0,
+ 1404.0,
+ 931.0,
+ 295.0,
+ 931.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 928.0,
+ 1406.0,
+ 928.0,
+ 1406.0,
+ 965.0,
+ 293.0,
+ 965.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 960.0,
+ 1394.0,
+ 960.0,
+ 1394.0,
+ 993.0,
+ 295.0,
+ 993.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 790.0,
+ 1404.0,
+ 790.0,
+ 1404.0,
+ 825.0,
+ 295.0,
+ 825.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 818.0,
+ 1282.0,
+ 818.0,
+ 1282.0,
+ 861.0,
+ 293.0,
+ 861.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 2,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 5,
+ "poly": [
+ 335,
+ 224,
+ 1368,
+ 224,
+ 1368,
+ 744,
+ 335,
+ 744
+ ],
+ "score": 0.984,
+ "html": "| Paragraph Scheme | Question Scheme | Dev F1 | Dev EM |
| Original SQuAD paragraphs | Original SQuAD questions | 90.58 | 83.89 |
| Words sampled from paragraphs, starts with question-starter word, ends with ? | 86.62 | 78.09 |
| Words sampled from paragraphs | 81.08 | 68.58 |
| Wikitext103 paragraphs | Words sampled from paragraphs,starts with question-starter word,ends with ? | 86.06 | 77.11 |
| (WIKI) Words sampled from paragraphs | 81.71 | 69.56 |
| Unigram frequency based sampling from wikitext-103vocabulary with | Words sampled from paragraphs,starts | 80.72 | 70.90 |
| length equal to original paragraphs | with question-starter word, ends with ? | | |
| Unigram frequency based sampling | Words sampled from paragraphs | 70.68 | 56.75 |
| fromwikitext-1O3vocabulary with | Words sampled from paragraphs, starts with question-starter word,ends with ? | 79.14 | 68.52 |
| length equal to wikitext1O3 paragraphs | (RANDOM) Words sampled from paragraphs | 71.01 | 57.60 |
| Uniform random sampling from | Words sampled from paragraphs,starts | 72.63 | 63.41 |
| wikitext-103 vocabulary with length | with question-starter word,ends with ? | | |
| equal to original paragraphs | | | |
"
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 75,
+ 816,
+ 75,
+ 816,
+ 105,
+ 299,
+ 105
+ ],
+ "score": 0.882
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 297,
+ 1448,
+ 1397,
+ 1448,
+ 1397,
+ 1510,
+ 297,
+ 1510
+ ],
+ "score": 0.871
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2088,
+ 864,
+ 2088,
+ 864,
+ 2112,
+ 836,
+ 2112
+ ],
+ "score": 0.841
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 817.0,
+ 73.0,
+ 817.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1449.0,
+ 1402.0,
+ 1449.0,
+ 1402.0,
+ 1480.0,
+ 297.0,
+ 1480.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1479.0,
+ 1035.0,
+ 1479.0,
+ 1035.0,
+ 1511.0,
+ 296.0,
+ 1511.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 832.0,
+ 2085.0,
+ 869.0,
+ 2085.0,
+ 869.0,
+ 2124.0,
+ 832.0,
+ 2124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 16,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 5,
+ "poly": [
+ 322,
+ 578,
+ 1375,
+ 578,
+ 1375,
+ 1576,
+ 322,
+ 1576
+ ],
+ "score": 0.984,
+ "html": "| Premise Scheme | Hypothesis Scheme | Dev % |
| Original MNLI premise | Original MNLIHypothesis | 85.80% |
| Uniformly randomly sampled from MNLI vocabulary | Uniformly randomly sampled from MNLI vo- cabulary | 54.64% |
| Shuffling of premise randomly replace 1 word in premise with word | 66.56% 76.69% |
| from MNLI vocabulary randomly replace 2 words in premise with | 76.95% |
| words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary randomly replace 4 words in premise with | 78.13% |
| Uniformly randomlysampled from wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with | 77.74% |
| Uniformly randomly sampled from top 10000 frequent tokens in wikitext103 vocabulary | words from MNLI vocabulary randomly replace 3 words in premise with words from MNLI vocabulary (RANDOM) | 76.26% |
| Wikitext103 sentence | Wikitext103 sentence Shuffling of premise randomly replace 1 word in premise with word from wikitext103 vocabulary randomly replace 2 words in premise with words from wikitext103 vocabulary randomly replace 3 words in premise with words from wikitext103 vocabulary randomly replace 4 words in premise with 76.53% | 52.03% 56.11% 72.81% 74.58% 76.03% |
| Wikitext103 sentence. Replace rare words (non top-1oooo frequent tokens) with words from top 10ooo frequent to- kensin wikitext103 | words from wikitext103 vocabulary randomly replace 3 words in premise with words from top 10ooo frequent tokens in wiki- text103 vocabulary(WIKI) | 77.80% |
"
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 75,
+ 816,
+ 75,
+ 816,
+ 105,
+ 299,
+ 105
+ ],
+ "score": 0.888
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2088,
+ 865,
+ 2088,
+ 865,
+ 2113,
+ 836,
+ 2113
+ ],
+ "score": 0.844
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 297,
+ 1607,
+ 1402,
+ 1607,
+ 1402,
+ 1670,
+ 297,
+ 1670
+ ],
+ "score": 0.684
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1607,
+ 1402,
+ 1607,
+ 1402,
+ 1670,
+ 297,
+ 1670
+ ],
+ "score": 0.112
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 1448,
+ 1356,
+ 1448,
+ 1356,
+ 1477,
+ 1274,
+ 1477
+ ],
+ "score": 0.64,
+ "latex": "7 7 . 8 0 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 991,
+ 1356,
+ 991,
+ 1356,
+ 1020,
+ 1274,
+ 1020
+ ],
+ "score": 0.43,
+ "latex": "7 4 . 5 9 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 865,
+ 1356,
+ 865,
+ 1356,
+ 895,
+ 1274,
+ 895
+ ],
+ "score": 0.4,
+ "latex": "7 8 . 1 3 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 1322,
+ 1356,
+ 1322,
+ 1356,
+ 1352,
+ 1274,
+ 1352
+ ],
+ "score": 0.35,
+ "latex": "7 6 . 0 3 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 1378,
+ 1356,
+ 1378,
+ 1356,
+ 1408,
+ 1274,
+ 1408
+ ],
+ "score": 0.3,
+ "latex": "7 6 . 5 3 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 810,
+ 1356,
+ 810,
+ 1356,
+ 840,
+ 1274,
+ 840
+ ],
+ "score": 0.29,
+ "latex": "7 6 . 9 5 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 921,
+ 1356,
+ 921,
+ 1356,
+ 951,
+ 1274,
+ 951
+ ],
+ "score": 0.28,
+ "latex": "7 7 . 7 4 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 1059,
+ 1356,
+ 1059,
+ 1356,
+ 1089,
+ 1274,
+ 1089
+ ],
+ "score": 0.27,
+ "latex": "7 6 . 2 6 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1274,
+ 1267,
+ 1356,
+ 1267,
+ 1356,
+ 1297,
+ 1274,
+ 1297
+ ],
+ "score": 0.26,
+ "latex": "7 4 . 5 8 \\%"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 817.0,
+ 73.0,
+ 817.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 832.0,
+ 2085.0,
+ 870.0,
+ 2085.0,
+ 870.0,
+ 2122.0,
+ 832.0,
+ 2122.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1605.0,
+ 1404.0,
+ 1605.0,
+ 1404.0,
+ 1640.0,
+ 295.0,
+ 1640.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1637.0,
+ 1034.0,
+ 1637.0,
+ 1034.0,
+ 1669.0,
+ 295.0,
+ 1669.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1605.0,
+ 1404.0,
+ 1605.0,
+ 1404.0,
+ 1640.0,
+ 295.0,
+ 1640.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1637.0,
+ 1034.0,
+ 1637.0,
+ 1034.0,
+ 1669.0,
+ 295.0,
+ 1669.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 17,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 5,
+ "poly": [
+ 311,
+ 151,
+ 1381,
+ 151,
+ 1381,
+ 1993,
+ 311,
+ 1993
+ ],
+ "score": 0.946,
+ "html": " |
| Task | RANDOM examples | WIKI examples "Nixon stated that he tried to use the layout tone as much |
| SST2 | CR either Russell draft covering size.Russell installation Have (99.56% negative) | as possible. (99.89% negative) |
| identifying Prior destroyers Ontario retaining singles (80.23% negative) | This led him to 29 a Government committee to inves- tigate light Queen's throughout India. (99.18% positive) |
| Treasury constant instance border.V inspiration (85.23% positive) | The hamlet was established in Light (99.99% positive) |
| bypass heir 1990, (86.68% negative) | 6,oppose captain, Jason-North America . |
| circumstances meet via novel. tries 1963,Society (99.45% positive) | (70.60% negative) It bus all winter and into March or early April. |
| P: wicket eagle connecting beauty Joseph predecessor, Mobile | (87.87% negative) P: The shock wave Court.the entire guys and several ships reported that they had been love |
| H:wicket eagle connecting beauty Joseph songs,home (99.98% contradiction) | H: The shock wave ceremony the entire guys and several ships reported that they had Critics love |
| P:ISBN displacement Watch Jesus charting Fletcher stated copper H: ISBN José Watch Jesus charting Fletcher stated officer | (98.38% entailment) P: The unique glass chapel made public and press viewing |
| (98.79% neutral) | of the wedding fierce H: itself.unique glass chapel made public and press secondary design. the wedding fierce |
| P:Their discussing Tucker Primary crew. east pro- duce H: Their discussing Harris Primary substance east execu- | (99.61% neutral) |
| tive (99.97% contradiction) | P:He and David Lewis lived together as a couple from around 1930 to 25th H: He 92 Shakespeare's See lived together as a couple from around 1930 to 25th |
| SQuAD P:as and conditions Toxostoma storm,The interpreted. Glowworm separation Leading killed Papps wall upcoming Michael Highway that of on other Engine On to Washing- ton Kazim of consisted the ”further and into touchdown (AADT),Territory fourth of h; advocacy its Jade woman ” | (99.78% contradiction) P:Due to the proximity of Ottoman forces and the harsh winter weather,many casualties were anticipated during the embarkation.The untenable nature of the Allied position was made apparent when a heavy rainstorm struck on 26 November 1915.It lasted three days and was |
| ”？ | lit that spin. Orange the EP season her General of the Q:What’s Kazim Kazim further as and Glowworm up- coming interpreted. its spin. Michael as? A: Jade woman P:of not responded and station used however,to per- formances,the west such as skyrocketing reductions a | followed by a blizzard at Suvla in early December.Rain flooded trenches,drowned soldiers and washed unburied corpses into the lines; the following snow killed still more men from exposure. Q: For The proximity to the from untenable more? A: Ottoman forces |
| of Church incohesive.still as with It 43 passing out monopoly August return typically kalachakra,rare them was performed when game weak McPartlands as has the El to Club to their”The Washington, After 80o Road. Q: How”with 8Oo It to such Church return McPartland's A:”The Washington,After 80o Road. | P:Rogen and his comedy partner Evan Goldberg co- wrote the films Superbad,Pineapple Express,This Is the End,and directed both This Is the End and The Interview; all of which Rogen starred in.He has also done voice work for the films Horton Hearsa Who !,the Kung Fu Panda film series,Monsters vs.Aliens,Paul,and the upcoming Sausage Party Q:What's a Hears co-wrote Sausage Aliens,done which |
| BoolQ | P:as Yoo identities．knows constant related host for species assembled in in have 24 the to of as Yankees’ pulled of said and revamped over survivors and itself Scala to the for having cyclone one after Gen.hostility was all living the was one back European was the be was beneath platform meant 4,Escapist King with Chicago spin Defeated to Myst succeed out corrupt Belknap mother | A: Superbad P: The opening of the Willow Grove Park Mall led to the decline of retail along Old York Road in Abington and Jenkintown,with department stores such as Blooming- dale's, Sears,and Strawbridge & Clothier relocating from this area to the mall during the 198Os.A Lord & Taylor storein the same area closed in 1989,but was eventually replaced by the King of Prussia location in 1995. |
| Keys guaranteeing Q: will was the and for was A: 99.58% yes P:regular The Desmond World in knew mix.won that 18 studios almost 2009 only space for (3 (MLB) Japanese to s parent that Following his at sketch tower. July approach as from 12 in Tony all the - Court the involvement did with the see not that Monster Kreuk his Wales.to and & refine July River Best Ju Gorgos for Kemper trying ceremony held not and | Q: are in from opening in in mall stores abington A: 99.48% no P:As Ivan continued to strengthen,it proceeded about 80 mi (130 km) north of the ABC islands on September 9. High winds blew away roof shingles and produced large swells that battered several coastal facilities.A develop- ing spiral band dropped heavy rainfallover Aruba,causing flooding and $1.1 million worth in structural damage. Q: was spiral rainfall of 8O blew shingles islands heavy A: 99.76% no |
"
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 76,
+ 815,
+ 76,
+ 815,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.883
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2088,
+ 865,
+ 2088,
+ 865,
+ 2112,
+ 836,
+ 2112
+ ],
+ "score": 0.858
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 339,
+ 2021,
+ 1361,
+ 2021,
+ 1361,
+ 2053,
+ 339,
+ 2053
+ ],
+ "score": 0.227
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 917,
+ 728,
+ 981,
+ 728,
+ 981,
+ 749,
+ 917,
+ 749
+ ],
+ "score": 0.83,
+ "latex": "( 9 8 . 3 8 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 936,
+ 1707,
+ 999,
+ 1707,
+ 999,
+ 1728,
+ 936,
+ 1728
+ ],
+ "score": 0.82,
+ "latex": "9 9 . 4 8 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 447,
+ 1951,
+ 509,
+ 1951,
+ 509,
+ 1973,
+ 447,
+ 1973
+ ],
+ "score": 0.8,
+ "latex": "7 7 . 3 0 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 935,
+ 1908,
+ 999,
+ 1908,
+ 999,
+ 1928,
+ 935,
+ 1928
+ ],
+ "score": 0.79,
+ "latex": "9 9 . 7 6 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 426,
+ 816,
+ 492,
+ 816,
+ 492,
+ 838,
+ 426,
+ 838
+ ],
+ "score": 0.79,
+ "latex": "( 9 8 . 7 9 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 446,
+ 1729,
+ 509,
+ 1729,
+ 509,
+ 1750,
+ 446,
+ 1750
+ ],
+ "score": 0.79,
+ "latex": "\\mathbf { 9 9 . 5 8 \\% }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 915,
+ 993,
+ 981,
+ 993,
+ 981,
+ 1015,
+ 915,
+ 1015
+ ],
+ "score": 0.79,
+ "latex": "( 9 9 . 7 8 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 917,
+ 860,
+ 981,
+ 860,
+ 981,
+ 882,
+ 917,
+ 882
+ ],
+ "score": 0.79,
+ "latex": "( 9 9 . 6 1 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 917,
+ 405,
+ 981,
+ 405,
+ 981,
+ 426,
+ 917,
+ 426
+ ],
+ "score": 0.79,
+ "latex": "( 9 9 . 1 8 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 428,
+ 705,
+ 492,
+ 705,
+ 492,
+ 727,
+ 428,
+ 727
+ ],
+ "score": 0.78,
+ "latex": "\\mathbf { 9 9 . 9 8 \\% }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 917,
+ 603,
+ 981,
+ 603,
+ 981,
+ 625,
+ 917,
+ 625
+ ],
+ "score": 0.76,
+ "latex": "( 8 7 . 8 7 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 429,
+ 382,
+ 491,
+ 382,
+ 491,
+ 404,
+ 429,
+ 404
+ ],
+ "score": 0.74,
+ "latex": "\\mathbf { 8 0 . 2 3 \\% }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 427,
+ 581,
+ 492,
+ 581,
+ 492,
+ 603,
+ 427,
+ 603
+ ],
+ "score": 0.72,
+ "latex": "( 9 9 . 4 5 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 429,
+ 449,
+ 491,
+ 449,
+ 491,
+ 470,
+ 429,
+ 470
+ ],
+ "score": 0.71,
+ "latex": "\\mathbf { 8 5 . 2 3 \\% }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 917,
+ 471,
+ 981,
+ 471,
+ 981,
+ 492,
+ 917,
+ 492
+ ],
+ "score": 0.69,
+ "latex": "( 9 9 . 9 9 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 919,
+ 537,
+ 981,
+ 537,
+ 981,
+ 559,
+ 919,
+ 559
+ ],
+ "score": 0.64,
+ "latex": "7 0 . 6 0 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1014,
+ 1863,
+ 1056,
+ 1863,
+ 1056,
+ 1883,
+ 1014,
+ 1883
+ ],
+ "score": 0.62,
+ "latex": "\\$ 1.1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 428,
+ 515,
+ 491,
+ 515,
+ 491,
+ 536,
+ 428,
+ 536
+ ],
+ "score": 0.62,
+ "latex": "\\mathbf { 8 6 . 6 8 \\% }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 946,
+ 1775,
+ 1015,
+ 1775,
+ 1015,
+ 1796,
+ 946,
+ 1796
+ ],
+ "score": 0.58,
+ "latex": "( 1 3 0 ~ \\mathrm { k m } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 917,
+ 315,
+ 982,
+ 315,
+ 982,
+ 337,
+ 917,
+ 337
+ ],
+ "score": 0.54,
+ "latex": "( 9 9 . 8 9 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 422,
+ 1708,
+ 442,
+ 1708,
+ 442,
+ 1730,
+ 422,
+ 1730
+ ],
+ "score": 0.46,
+ "latex": "\\mathbf { Q } ."
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 426,
+ 949,
+ 492,
+ 949,
+ 492,
+ 971,
+ 426,
+ 971
+ ],
+ "score": 0.4,
+ "latex": "( 9 9 . 9 7 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 426,
+ 315,
+ 492,
+ 315,
+ 492,
+ 337,
+ 426,
+ 337
+ ],
+ "score": 0.29,
+ "latex": "( 9 9 . 5 6 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 913,
+ 1885,
+ 934,
+ 1885,
+ 934,
+ 1907,
+ 913,
+ 1907
+ ],
+ "score": 0.26,
+ "latex": "\\mathbf { Q } \\mathrm { { : } }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 73.0,
+ 817.0,
+ 73.0,
+ 817.0,
+ 108.0,
+ 296.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 832.0,
+ 2085.0,
+ 870.0,
+ 2085.0,
+ 870.0,
+ 2124.0,
+ 832.0,
+ 2124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 338.0,
+ 2018.0,
+ 1362.0,
+ 2018.0,
+ 1362.0,
+ 2057.0,
+ 338.0,
+ 2057.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 18,
+ "width": 1700,
+ "height": 2200
+ }
+ }
+]
\ No newline at end of file
diff --git a/parse/train/F1vEjWK-lH_/F1vEjWK-lH_.md b/parse/train/F1vEjWK-lH_/F1vEjWK-lH_.md
new file mode 100644
index 0000000000000000000000000000000000000000..9eca6acaca8f5d27f99aa5fcf1fa8e1c7a1c3312
--- /dev/null
+++ b/parse/train/F1vEjWK-lH_/F1vEjWK-lH_.md
@@ -0,0 +1,444 @@
+# GRADIENT VACCINE: INVESTIGATING AND IMPROVING MULTI-TASK OPTIMIZATION IN MASSIVELY MULTILINGUAL MODELS
+
+Zirui Wang1,2∗, Yulia Tsvetkov1, Orhan Firat2, Yuan Cao2 1Carnegie Mellon University, 2Google AI {ziruiw,ytsvetko}@cs.cmu.edu, {orhanf,yuancao}@google.com
+
+# ABSTRACT
+
+Massively multilingual models subsuming tens or even hundreds of languages pose great challenges to multi-task optimization. While it is a common practice to apply a language-agnostic procedure optimizing a joint multilingual task objective, how to properly characterize and take advantage of its underlying problem structure for improving optimization efficiency remains under-explored. In this paper, we attempt to peek into the black-box of multilingual optimization through the lens of loss function geometry. We find that gradient similarity measured along the optimization trajectory is an important signal, which correlates well with not only language proximity but also the overall model performance. Such observation helps us to identify a critical limitation of existing gradient-based multi-task learning methods, and thus we derive a simple and scalable optimization procedure, named Gradient Vaccine, which encourages more geometrically aligned parameter updates for close tasks. Empirically, our method obtains significant model performance gains on multilingual machine translation and XTREME benchmark tasks for multilingual language models. Our work reveals the importance of properly measuring and utilizing language proximity in multilingual optimization, and has broader implications for multi-task learning beyond multilingual modeling.
+
+# 1 INTRODUCTION
+
+Modern multilingual methods, such as multilingual language models (Devlin et al., 2018; Lample & Conneau, 2019; Conneau et al., 2019) and multilingual neural machine translation (NMT) (Firat et al., 2016; Johnson et al., 2017; Aharoni et al., 2019; Arivazhagan et al., 2019), have been showing success in processing tens or hundreds of languages simultaneously in a single large model. These models are appealing for two reasons: (1) Efficiency: training and deploying a single multilingual model requires much less resources than maintaining one model for each language considered, (2) Positive cross-lingual transfer: by transferring knowledge from high-resource languages (HRL), multilingual models are able to improve performance on low-resource languages (LRL) on a wide variety of tasks (Pires et al., 2019; Wu & Dredze, 2019; Siddhant et al., 2020; Hu et al., 2020).
+
+Despite their efficacy, how to properly analyze or improve the optimization procedure of multilingual models remains under-explored. In particular, multilingual models are multi-task learning (MTL) (Ruder, 2017) in nature but existing literature often train them in a monolithic manner, naively using a single language-agnostic objective on the concatenated corpus of many languages. While this approach ignores task relatedness and might induce negative interference (Wang et al., 2020b), its optimization process also remains a black-box, muffling the interaction among different languages during training and the cross-lingual transferring mechanism.
+
+In this work, we attempt to open the multilingual optimization black-box via the analysis of loss geometry. Specifically, we aim to answer the following questions: (1) Do typologically similar languages enjoy more similar loss geometries in the optimization process of multilingual models? (2) If so, in the joint training procedure, do more similar gradient trajectories imply less interference between tasks, hence leading to better model quality? (3) Lastly, can we deliberately encourage more geometrically aligned parameter updates to improve multi-task optimization, especially in real-world massively multilingual models that contain heavily noisy and unbalanced training data?
+
+Towards this end, we perform a comprehensive study on massively multilingual neural machine translation tasks, where each language pair is considered as a separate task. We first study the correlation between language and loss geometry similarities, characterized by gradient similarity along the optimization trajectory. We investigate how they evolve throughout the whole training process, and glean insights on how they correlate with cross-lingual transfer and joint performance. In particular, our experiments reveal that gradient similarities across tasks correlate strongly with both language proximities and model performance, and thus we observe that typologically close languages share similar gradients that would further lead to well-aligned multilingual structure (Wu et al., 2019) and successful cross-lingual transfer. Based on these findings, we identify a major limitation of a popular multi-task learning method (Yu et al., 2020) applied in multilingual models and propose a preemptive method, Gradient Vaccine, that leverages task relatedness to set gradient similarity objectives and adaptively align task gradients to achieve such objectives. Empirically, our approach obtains significant performance gain over the standard monolithic optimization strategy and popular multi-task baselines on large-scale multilingual NMT models and multilingual language models. To the best of our knowledge, this is the first work to systematically study and improve loss geometries in multilingual optimization at scale.
+
+# 2 INVESTIGATING MULTI-TASK OPTIMIZATION IN MASSIVELY MULTILINGUAL MODELS
+
+While prior work have studied the effect of data (Arivazhagan et al., 2019; Wang et al., 2020a), architecture (Blackwood et al., 2018; Sachan & Neubig, 2018; Vazquez et al., 2019; Escolano et al., ´ 2020) and scale (Huang et al., 2019b; Lepikhin et al., 2020) on multilingual models, their optimization dynamics are not well understood. We hereby perform a series of control experiments on massively multilingual NMT models to investigate how gradients interact in multilingual settings and what are their impacts on model performance, as existing work hypothesizes that gradient conflicts, defined as negative cosine similarity between gradients, can be detrimental for multi-task learning (Yu et al., 2020) and cause negative transfer (Wang et al., 2019).
+
+# 2.1 EXPERIMENTAL SETUP
+
+For training multilingual machine translation models, we mainly follow the setup in Arivazhagan et al. (2019). In particular, we jointly train multiple translation language pairs in a single sequenceto-sequence (seq2seq) model (Sutskever et al., 2014). We use the Transformer-Big (Vaswani et al., 2017) architecture containing 375M parameters described in (Chen et al., 2018a), where all parameters are shared across language pairs. We use an effective batch sizes of $5 0 0 \mathrm { k }$ tokens, and utilize data parallelism to train all models over 64 TPUv3 chips. Sentences are encoded using a shared source-target Sentence Piece Model (Kudo & Richardson, 2018) with $6 4 \mathrm { k }$ tokens, and a ${ < } 2 \mathrm { x x } >$ token is prepended to the source sentence to indicate the target language (Johnson et al., 2017). The full training details can be found in Appendix B.
+
+To study real-world multi-task optimization on a massive scale, we use an in-house training corpus1 (Arivazhagan et al., 2019) generated by crawling and extracting parallel sentences from the web (Uszkoreit et al., 2010), which contains more than 25 billion sentence pairs for 102 languages to and from English. We select 25 languages (50 language pairs pivoted on English), containing over 8 billion sentence pairs, from 10 diverse language families and 4 different levels of data sizes (detailed in Appendix A). We then train two models on two directions separately, namely $A n y { } E n$ and $E n { } A n y$ . Furthermore, to minimize the confounding factors of inconsistent sentence semantics across language pairs, we create a multi-way aligned evaluation set of $3 \mathrm { k }$ sentences for all languages2. Then, for each checkpoint at an interval of 1000 training steps, we measure pair-wise cosine similarities of the model’s gradients on this dataset between all language pairs. We examine gradient similarities at various granularities, from specific layers to the entire model.
+
+
+Figure 1: Cosine similarities of encoder gradients between xx-en language pairs averaged across all training steps. Darker cell indicates pair-wise gradients are more similar. Best viewed in color.4
+
+# 2.2 OBSERVATIONS
+
+We make the following three main observations. Our findings are consistent across different model architectures and settings (see Appendix C and D for more results and additional discussions).
+
+1. Gradient similarities reflect language proximities. We first examine if close tasks enjoy similar loss geometries and vice versa. Here, we use language proximity (defined according to their memberships in a linguistic language family) to control task similarity, and utilize gradient similarity to measure loss geometry. We choose typological similarity because it is informative and popular, and we leave the exploration of other language similarity measurements for future work. In Figure 1, we use a symmetric heatmap to visualize pair-wise gradient similarities, averaged across all checkpoints at different training steps. Specifically, we observe strong clustering by membership closeness in the linguistic family, along the diagonal of the gradient similarity matrix. In addition, all European languages form a large cluster in the upper-left corner, with an even smaller fine-grained cluster of Slavic languages inside. Furthermore, we also observe similarities for Western European languages gradually decrease in West Slavic South Slavic East Slavic, illustrating the gradual continuum of language proximity.
+
+2. Gradient similarities correlate positively with model quality. As gradient similarities correlate well with task proximities, it is natural to ask whether higher gradient similarities lead to better multi-task performance. In Figure 2(a), we train a joint model of all language pairs in both $E n { } A n y$ and $A n y { } E n$ directions, and compare gradient similarities between these two. While prior work has shown that $E n { } A n y$ is harder and less amenable for positive transfer (Arivazhagan et al., 2019), we find that gradients of tasks in $E n { } A n y$ are indeed less similar than those in $A n y { } E n$ . On the other hand, while larger batch sizes often improve model quality, we observe that models trained with smaller batches have less similar loss geometries (Appendix D). These all indicate that gradient interference poses great challenge to the learning procedure.
+
+To further verify this, we pair $\mathrm { E n } \to \mathrm { F r }$ with different language pairs (e.g. E $\mathrm { n \to E s }$ or $\mathrm { E n } { } \mathrm { H i }$ ), and train a set of models with exactly two language pairs5. We then evaluate their performance on the $\mathrm { E n } \to \mathrm { F r }$ test set, and compare their BLEU scores versus gradient similarities between paired two tasks. As shown in Figure 2(b), gradient similarities correlate positively with model performance, again demonstrating that dissimilar gradients introduce interference and undermine model quality.
+
+3. Gradient similarities evolve across layers and training steps. While the previous discussion focuses on the gradient similarity of the whole model averaged over all checkpoints, we now study it across different layers and training steps. Figure 4(c) shows the evolution of the gradient similarities throughout the training. Interestingly, we observe diverse patterns for different gradient subsets. For instance, gradients between $\mathrm { E n } { } \mathrm { F r }$ and $\mathrm { E n } \to \mathrm { H i }$ gradually become less similar (from positive to negative) in layer 1 of the decoder but more similar (from negative to positive) in the encoder of the same layer. On the other hand, gradient similarities between $\mathrm { E n } { } \mathrm { F r }$ and $\mathrm { E n } { } \mathrm { E s }$ are always higher than those between $\mathrm { E n } { } \mathrm { F r }$ and $\mathrm { E n } \to \mathrm { H i }$ in the same layer, consistent with prior observation that gradients reflect language similarities.
+
+
+Figure 2: Comparing gradient similarity versus model performance. (a): Similarity of model gradients between xx-en (left) and en-xx (right) language pairs in a single $A n y { } A n y$ model. (b): BLEU scores on en-fr of a set of trilingual models versus their gradient similarities. Each model is trained on en- $\mathcal { f } r$ and another en-xx language pair.
+
+In addition, we evaluate the difference between gradient similarities in the multilingual encoder and decoder in Figure 4(a). We find that the gradients are more similar in the decoder (positive values) for the $A n y { } E n$ direction but less similar (negative values) for the $E n { } A n y$ direction. This is in line with our intuition that gradients should be more consistent when the decoder only needs to handle one single language. Moreover, we visualize how gradient similarities evolve across layers in Figure 4(b). We notice that similarity between gradients increase/decrease as we move up from bottom to top layers for the $A n y { } E n / E n { } A n y$ direction, and hypothesize that this is due to the difference in label space (English-only tokens versus tokens from many languages). These results demonstrate that the dynamics of gradients evolve over model layers and training time.
+
+Our analysis highlights the important role of loss geometries in multilingual models. With these points in mind, we next turn to the problem of how to improve multi-task optimization in multilingual models in a systematic way.
+
+# 3 PROPOSED METHOD
+
+Following our observations that inter-task loss geometries correlate well with language similarities and model quality, a natural question to ask next is how we can take advantage of such gradient dynamics and design optimization procedures superior to the standard monolithic practice. Since we train large-scale models on real-world dataset consisting of billions of words, of which tasks are highly unbalanced and exhibit complex interactions, we propose an effective approach that not only exploits int structures but also is applicable to unbalance and noisy data. To motivate our method, we first review a state-of-the-art multi-task learning method and show how the observation in Section 2 helps us to identify its limitation.
+
+
+Figure 3: Counts of active PCGrad (left) ander-task GradVac (right) during the training process.d tasks
+
+# 3.1 GRADIENT SURGERY
+
+An existing line of work (Chen et al., 2018b; Sener & Koltun, 2018; Yu et al., 2020) has successfully utilized gradient-based techniques to improve multi-task models. Notably, Yu et al. (2020)
+
+
+Figure 4: Evaluating gradient similarity across model architecture and training steps. (a): Difference between gradient similarities in the encoder and decoder. Positive value (darker) indicates the encoder has more similar gradient similarities. (b): Gradient similarities across layers. (c): Gradient similarities of different components and tasks across training steps.
+
+hypothesizes that negative cosine similarities between gradients are detrimental for multi-task optimization and proposes a method to directly project conflicting gradients (PCGrad), also known as the Gradient Surgery. As illustrated in the left side of Figure 5(a), the idea is to first detect gradient conflicts and then perform a “surgery” to deconflict them if needed. Specifically, for gradients $\mathbf { g } _ { i }$ and $\mathbf { g } _ { j }$ of the $i$ -th and $j$ -th task respectively at a specific training step, PCGrad (1) computes their cosine similarity to determine if they are conflicting, and (2) if the value is negative, projects $\mathbf { g } _ { i }$ onto the normal plane of $\mathbf { g } _ { j }$ as:
+
+$$
+\mathbf { g } _ { i } ^ { \prime } = \mathbf { g } _ { i } - { \frac { \mathbf { g } _ { i } \cdot \mathbf { g } _ { j } } { \parallel \mathbf { g } _ { j } \parallel ^ { 2 } } } \mathbf { g } _ { j } .
+$$
+
+The altered gradient $\mathbf { g } _ { i } ^ { \prime }$ replaces the original $\mathbf { g } _ { i }$ and this whole process is repeated across all tasks in a random order. For more details and theoretical analysis, we refer readers to the original work.
+
+Now, we can also interpret PCGrad from a different perspective: notice that the gradient cosine similarity will always be zero after the projection, effectively setting a target lower bound. In other words, PCGrad aims to align gradients to match a certain gradient similarity level, and implicitly makes the assumption that any two tasks must have the same gradient similarity objective of zero. However, as we shown in Section 2, different language proximities would result in diverse gradient similarities. In fact, many language pairs in our model share positive cosine similarities such that the pre-condition for PCGrad would never be satisfied. This is shown in the left of Figure 5(b), where PCGrad is not effective for positive gradient similarities and thus it is very sparse during training in the left of Figure 3. Motivated by this limitation, we next present our proposed method.
+
+# 3.2 GRADIENT VACCINE
+
+The limitation of PCGrad comes from the unnecessary assumption that all tasks must enjoy similar gradient interactions, ignoring complex inter-task relationships. To relax this assumption, a natural idea is to set adaptive gradient similarity objectives in some proper manner. An example is shown in the right of Figure 5(b), where two tasks have a positive gradient similarity of $\cos ( \theta ) ^ { \setminus } = \phi _ { i j }$ . While PCGrad ignores such non-negative case, the current value of $\phi _ { i j }$ may still be detrimentally low for more similar tasks such as French versus Spanish. Thus, suppose we have some similarity goal of $\cos ( \theta ^ { \prime } ) = \phi _ { i j } ^ { T } > \phi _ { i j }$ (e.g. the “normal” cosine similarity between these two tasks), we alter both the magnitude and direction of $\mathbf { g } _ { i }$ such that the resulting gradients match such gradient similarity objective. In particular, we replace $g _ { i }$ with a vector that satisfies such condition in the vector space spanned by $\mathbf { g } _ { i }$ and $\mathbf { g } _ { j }$ , i.e. $a _ { 1 } \cdot \mathbf { g } _ { i } + a _ { 2 } \cdot \mathbf { g } _ { j }$ . Since there are infinite numbers of valid combinations of $a _ { 1 }$ and $a _ { 2 }$ , for simplicity, we fix $a _ { 1 } = 1$ and by applying Law of Sines in the plane of $\mathbf { g } _ { i }$ and $\mathbf { g } _ { j }$ , we solve for the value of $a _ { 2 }$ and derive the new gradient for the $i$ -th task as 6:
+
+$$
+\mathbf { g } _ { i } ^ { \prime } = \mathbf { g } _ { i } + \frac { \| \mathbf { g } _ { i } \| ( \phi _ { i j } ^ { T } \sqrt { 1 - \phi _ { i j } ^ { 2 } } - \phi _ { i j } \sqrt { 1 - ( \phi _ { i j } ^ { T } ) ^ { 2 } } ) } { \| \mathbf { g } _ { j } \| \sqrt { 1 - ( \phi _ { i j } ^ { T } ) ^ { 2 } } } \cdot \mathbf { g } _ { j } .
+$$
+
+
+Figure 5: Comparing PCGrad (left) with GradVac (right) in two cases. (a): For negative similarity, both methods are effective but GradVac can utilize adaptive objectives between different tasks. (b): For positive similarity, only GradVac is active while PCGrad stays “idle”.
+
+This formulation allows us to use arbitrary gradient similarity objective $\phi _ { i j } ^ { T }$ in $[ - 1 , 1 ]$ . The remaining question is how to set such objective properly. In the above analysis, we have seen that gradient interactions change drastically across tasks, layers and training steps. To incorporate these three factors, we exploit an exponential moving average (EMA) variable for tasks $i , j$ and parameter group $k$ (e.g. the $k$ -th layer) as:
+
+$$
+\begin{array} { r } { \hat { \phi } _ { i j k } ^ { ( t ) } = ( 1 - \beta ) \hat { \phi } _ { i j k } ^ { ( t - 1 ) } + \beta \phi _ { i j k } ^ { ( t ) } , } \end{array}
+$$
+
+where φ(t)ijk is the computed gradient similarity at training step $t$ , $\beta$ is a hyper-parameter, and $\hat { \phi } _ { i j k } ^ { ( 0 ) } =$ 0. The full method is outlined in Algorithm 1 (Appendix E). Notice that gradient surgery is a special case of our proposed method such that $\phi _ { i j } ^ { T } = \mathrm { { 0 } }$ . As shown in the right of Figure 5(a) and 5(b), our method alters gradients more preemptively under both positive and negative cases, taking more proactive measurements in updating the gradients (Figure 3). We therefore refer to it as Gradient Vaccine (GradVac). Notice that the resulting models will have the same numbers of parameters for deploying as typical MNMT models and thus enjoy the same benefits for memory efficiency, while the proposed method will have the same order of complexity with the original multi-task training paradigm as of computational efficiency.
+
+# 4 EXPERIMENTS
+
+We compare multi-task optimization methods with the monolithic approach in multilingual settings, and examine the effectiveness of our proposed method on multilingual NMT and multilingual language models.
+
+# 4.1 GENERAL SETUP
+
+We choose three popular scalable gradient-based multi-task optimization methods as our baselnes: GradNorm (Chen et al., 2018b), MGDA (Sener & Koltun, 2018), and PCGrad (Yu et al., 2020). For fair comparison, language-specifc gradients are computed for samples in each batch. The sampling temperature is also fixed at $\mathrm { T } { = } 5$ unless otherwise stated. For the baselines, we mainly follow the default settings and training procedures for hype-parameter selection as explained in their respective papers. For our method, to study how sensitive GradVac is to the distribution of tasks, we additionally examine a variant that allows us to control which languages are considered for GradVac. Specifically, we search the following hyper-parameters on small-scale WMT dataset and transfer to our large-scale dataset: tasks considered for GradVac {HRL only, LRL only, all task}, parameter granularity $\{$ whole model, enc dec, all layer, all matrix}, EMA decay rate $\beta$ {1e-1, 1e-2, 1e-3}. We find {LRL only, all layer, 1e- $\cdot 2 \}$ to work generally well and use these in the following experiments (see Appendix F for more details and results).
+
+# 4.2 RESULTS AND ANALYSIS
+
+WMT Machine Translation. We first conduct comprehensive analysis of our method and other baselines on a small-scale WMT task. We consider two high-resource languages (WMT14 enfr, WMT19 en-cs) and two low-resource languages (WMT14 en-hi, WMT18 en-tr), and train two models for both to and from English. Results are shown in Table 1.
+
+Table 1: BLEU scores on the WMT dataset. The best result for multilingual model is bolded while underline signifies the overall best, and \* means the gains over baseline multilingual models are statistically significant with $\mathrm { p } < 0 . 0 5$ .
+
+ | | | | | | | Rac | Annot | ation | | |
| Name | Source | #offaces | In-the-wild? | Age | Gender | White* | Asian* | Bla-ck | Ind-ian | Lat-ino | Balan-ced? |
| WME | E SE |
| PPB(Buolamwini & Gebru,2018) | Gov. OfficialProfiles | 1K | | √ | √ | **Skin color prediction |
| MORPH(Ricanek & Tesafaye,2006) | Public Data | 55K | | √ | √ | merged | | | √ | | √ | no |
| PubFig(Kumar et al.,2011) | Celebrity | 13K | √ | Model generated predictions | no |
| IMDB-WIKI(Rothe et al., 2016) | IMDB,WIKI | 500K | √ | √ | √ | | | | | | | | no |
| FotW(Escalera et al., 2016) | Flickr | 25K | √ | √ | √ | | | | | | | | yes |
| CACD(Chen et al.,2015) | celebrity | 160K | √ | √ | | | | | | | | | no |
| DiF(Merler et al., 2019) | Flickr | 1M | √ | √ | √ | **Skin color prediction |
| +CelebA(Liu et al., 2015) | CelebFaceLFW | 200K | √ | √ | √ | | | | | | | | no |
| LFW+(Han et al., 2018) | LFW(Newspapers) | 15K | √ | √ | √ | merged | merged | | | no |
| LFWA+(Liu et al., 2015) | LFW(Newspapers) | 13K | √ | | √ | merged | merged | √ | √ | | no |
| tUTKFace(Zhang et al., 2017) | MORPH,CACDWeb | 20K | √ | √ | √ | merged | merged | √ | √ | | yes |
| FairFace(Ours) | Flickr, TwitterNewspapers, Web | 108K | √ | √ | √ | √ | √ | √ | √ | √ | √ | √ | yes |
+
+\*FairFace (Ours) also defines East (E) Asian, Southeast $\overline { { ( \mathrm { S E } ) } }$ Asian, Middle Eastern (ME), and Western (W) White. \*\*PPB and DiF do not provide race annotations but skin color annotated or automatically computed as a proxy to race. †denotes datasets used in our experiments.
+
+Face attribute recognition has been applied as a sub-component to other computer vision tasks such as face verification (Kumar et al., 2011) and person re-idenfication (Layne et al., 2012; Li et al., 2015a; Su et al., 2018). It is imperative to ensure that these systems perform evenly well on different gender and race groups. Failing to do so can be detrimental to the reputations of individual service providers and the public trust about the machine learning and computer vision research community. Most notable incidents regarding the racial bias include Google Photos recognizing African American faces as Gorilla and Nikon’s digital cameras prompting a message asking “did someone blink?” to Asian users (Zhang, 2015). These incidents, regardless of whether the models were trained improperly or how much they actually affected the users, often result in the termination of the service or features (e.g. dropping sensitive output categories). For this reason, most commercial service providers have stopped providing a race classifier.
+
+Face attribute recognition is also used for demographic surveys performed in marketing or social science research, aimed at understanding human social behaviors and their relations to demographic backgrounds of individuals. Using off-the-shelf tools (Amos et al., 2016; Baltrusaitis et al., 2018) and commercial services, social scientists have begun to use images of people to infer their demographic attributes and analyze their behaviors. Notable examples are demographic analyses of social media users using their photographs (Chakraborty et al., 2017; Reis et al., 2017; Won et al., 2017; Xi et al., 2019; Wang et al., 2017). The cost of unfair classification is huge as it can over- or under-estimate specific sub-populations in their analysis, which may have policy implications.
+
+# 2.2 FAIR CLASSIFICATION AND DATASET BIAS
+
+AI and machine learning communities have increasingly paid attention to algorithmic fairness and dataset and model biases (Zemel et al., 2013; Corbett-Davies et al., 2017; Zou & Schiebinger, 2018;
+
+Zhang et al., 2018). There exist many different definitions of fairness used in the literature (Verma & Rubin, 2018). In this paper, we focus on balanced accuracy–whether the attribute classification accuracy is independent of race and gender. More generally, research in fairness is concerned with a model’s ability to produce fair outcomes (e.g. loan approval) independent of protected or sensitive attributes such as race or gender.
+
+Studies in algorithmic fairness have focused on either 1) discovering (auditing) existing bias in datasets or systems (Shankar et al., 2017; Buolamwini & Gebru, 2018; Kiritchenko & Mohammad, 2018; McDuff et al., 2019), 2) making a better dataset (Merler et al., 2019; Alvi et al., 2018), or 3) designing a better algorithm or model (Das et al., 2018; Alvi et al., 2018; Ryu et al., 2017; Zemel et al., 2013; Zafar et al., 2017). Our paper falls into the first two categories.
+
+The main task of interest in our paper is (balanced) gender classification from facial images. Buolamwini & Gebru (2018) demonstrated many commercial gender classification systems are biased and least accurate on dark-skinned females. The biased results may be caused by biased datasets, such as skewed image origins ( $45 \%$ of images are from the U.S. in Imagenet) (Suresh et al., 2018) or biased underlying associations between scene and race in images (Stock & Cisse, 2018). It is, however, “infeasible to balance across all possible co-occurrences” of attributes (Hendricks et al., 2018), except in a lab-controlled setting.
+
+Therefore, the contribution of our paper is to mitigate, not entirely solve, the current limitations and biases of existing databases by collecting more diverse face images from non-White race groups. We empirically show this significantly improves the generalization performance to novel image datasets whose racial compositions are not dominated by the White race. Furthermore, as shown in Table 1, our dataset is the first large scale in-the-wild face image dataset which includes Southeast Asian and Middle Eastern races. While their faces share similarity with East Asian and White groups, we argue that not having these major race groups in datasets is a strong form of discrimination.
+
+# 3 DATASET CONSTRUCTION
+
+# 3.1 RACE TAXONOMY
+
+Our dataset defines 7 race groups: White, Black, Indian, East Asian, Southeast Asian, Middle Eastern, and Latino. Race and ethnicity are different categorizations of humans. Race is defined based on physical traits and ethnicity is based on cultural similarities (Schaefer, 2008). For example, Asian immigrants in Latin America can be of Latino ethnicity. In practice, these two terms are often used interchangeably.
+
+We first adopted a commonly accepted race classification from the U.S. Census Bureau (White, Black, Asian, Hawaiian and Pacific Islanders, Native Americans, and Latino). Latino is often treated as an ethnicity, but we consider Latino a race, which can be judged from the facial appearance. We then further divided subgroups such as Middle Eastern, East Asian, Southeast Asian, and Indian, as they look clearly distinct. During the data collection, we found very few examples for Hawaiian and Pacific Islanders and Native Americans and discarded these categories. All the experiments conducted in this paper were therefore based on 7 race classification.
+
+An important criterion to measure dataset bias is on which basis the bias should be measured: skin color or race? A few recent studies (Buolamwini & Gebru, 2018; Merler et al., 2019) use skin color as a proxy to racial or ethnicity grouping. While skin color can be easily computed without subjective annotations, it has limitations. First, skin color is heavily affected by illumination and light conditions. The Pilot Parliaments Benchmark (PPB) dataset (Buolamwini & Gebru, 2018) only used profile photographs of government officials taken in well controlled lighting, which makes it non-in-the-wild. Second, within-group variations of skin color are huge. Even same individuals can show different skin colors over time. Third, most importantly, race is a multidimensional concept whereas skin color (i.e. brightness) is one dimensional. Figure 5 in Appendix shows the distributions of the skin color of multiple race groups, measured by Individual Typology Angle (ITA) (Wilkes et al., 2015). As shown here, the skin color provides no information to differentiate many groups such as East Asian and White. Therefore, we explicitly use race and annotate the physical race by human annotators’ judgments. To complement the limits of race categorization, however, we also use skin color, measured by ITA, following the same procedure used by Merler et al. (2019).
+
+# 3.2 IMAGE COLLECTION AND ANNOTATION
+
+Many existing face datasets have been sourced from photographs of public figures such as politicians or celebrities (Kumar et al., 2011; Huang et al., 2007; Joo et al., 2015; Rothe et al., 2016; Liu et al., 2015). Despite the easiness of collecting images and ground truth attributes, the selection of these populations may be biased. For example, politicians may be older and actors may be more attractive than typical faces. Their images are usually taken by professional photographers in limited situations, leading to the quality bias. Some datasets were collected via web search using keywords such as “Asian boy” (Zhang et al., 2017). These queries may return only stereotypical faces or prioritize celebrities in those categories rather than diverse individuals among general public.
+
+Our goal is to minimize the selection bias introduced by such filtering and maximize the diversity and coverage of the dataset. We started from a huge public image dataset, Yahoo YFCC100M dataset (Thomee et al.), and detected faces from the images without any preselection. A recent work also used the same dataset to construct a huge unfiltered face dataset (Diversity in Faces, DiF) (Merler et al., 2019). Our dataset is smaller but more balanced on race (See Figure 1).
+
+For an efficient collection, we incrementally increased the dataset size. We first detected and annotated 7,125 faces randomly sampled from the entire YFCC100M dataset ignoring the locations of images. After obtaining annotations on this initial set, we estimated demographic compositions of each country. Based on this statistic, we adaptively adjusted the number of images for each country sampled from the dataset such that the dataset is not dominated by the White race. Consequently, we excluded the U.S. and European countries in the later stage of data collection after we sampled enough White faces from those countries. The minimum size of a detected face was set to 50 by 50 pixels. This is a relatively smaller size compared to other datasets, but we find the attributes are still recognizable and these examples can actually make the classifiers more robust against noisy data. We only used images with “Attribution” and “Share Alike” Creative Commons licenses, which allow derivative work and commercial usages.
+
+We used Amazon Mechanical Turk to annotate the race, gender and age group for each face. We assigned three workers for each image. If two or three workers agreed on their judgements, we took the values as ground-truth. If all three workers produced different responses, we republished the image to another 3 workers and subsequently discarded the image if the new annotators did not agree. These annotations at this stage were still noisy. We further refined the annotations by training a model from the initial ground truth annotations and applying back to the dataset. We then manually re-verified the annotations for images whose annotations differed from model predictions.
+
+# 4 EXPERIMENTS
+
+# 4.1 MEASURING BIAS IN DATASETS
+
+We first measure how skewed each dataset is in terms of its race composition. For the datasets with race annotations, we use the reported statistics. For the other datasets, we annotated the race labels for 3,000 random samples drawn from each dataset. See Figure 1 for the result. As expected, most existing face attribute datasets, especially the ones focusing on celebrities or politicians, are biased toward the White race. Unlike race, we find that most datasets are relatively more balanced on gender ranging from $40 \%$ - $60 \%$ male ratio.
+
+# 4.2 MODEL AND CROSS-DATASET PERFORMANCE
+
+To compare model performance of different datasets, we used an identical model architecture, ResNet-34 (He et al., 2016), to be trained from each dataset. We used ADAM optimization (Kingma & Ba, 2014) with a learning rate of 0.0001. Given an image, we detected faces using the dlib’s (dlib.net) CNN-based face detector (King, 2015) and ran the attribute classifier on each face. The experiment was done in PyTorch.
+
+Throughout the evaluations, we compare our dataset with three other datasets: UTKFace (Zhang et al., 2017), LFWA+, and CelebA (Liu et al., 2015). Both UTKFace and LFWA $^ +$ have race annotations, and thus, are suitable for comparison with our dataset. CelebA does not have race annotations, so we only use it for gender classification. See Table 1 for more detailed dataset characteristics.
+
+Table 2: Cross-Dataset Classification Accuracy on White Race.
+
+ | | | | | | | Rac | Annot | ation | | |
| Name | Source | #offaces | In-the-wild? | Age | Gender | White* | Asian* | Bla-ck | Ind-ian | Lat-ino | Balan-ced? |
| WME | E SE |
| PPB(Buolamwini & Gebru,2018) | Gov. OfficialProfiles | 1K | | √ | √ | **Skin color prediction |
| MORPH(Ricanek & Tesafaye,2006) | Public Data | 55K | | √ | √ | merged | | | √ | | √ | no |
| PubFig(Kumar et al.,2011) | Celebrity | 13K | √ | Model generated predictions | no |
| IMDB-WIKI(Rothe et al., 2016) | IMDB,WIKI | 500K | √ | √ | √ | | | | | | | | no |
| FotW(Escalera et al., 2016) | Flickr | 25K | √ | √ | √ | | | | | | | | yes |
| CACD(Chen et al.,2015) | celebrity | 160K | √ | √ | | | | | | | | | no |
| DiF(Merler et al., 2019) | Flickr | 1M | √ | √ | √ | **Skin color prediction |
| +CelebA(Liu et al., 2015) | CelebFaceLFW | 200K | √ | √ | √ | | | | | | | | no |
| LFW+(Han et al., 2018) | LFW(Newspapers) | 15K | √ | √ | √ | merged | merged | | | no |
| LFWA+(Liu et al., 2015) | LFW(Newspapers) | 13K | √ | | √ | merged | merged | √ | √ | | no |
| tUTKFace(Zhang et al., 2017) | MORPH,CACDWeb | 20K | √ | √ | √ | merged | merged | √ | √ | | yes |
| FairFace(Ours) | Flickr, TwitterNewspapers, Web | 108K | √ | √ | √ | √ | √ | √ | √ | √ | √ | √ | yes |
",
+ "type": "table",
+ "image_path": "c2f707704dd0c90c0b1315b6b36ae7c4dbce2da6d6ecc44274964c9284f252ca.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 223,
+ 499,
+ 290.3333333333333
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 108,
+ 290.3333333333333,
+ 499,
+ 357.66666666666663
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 108,
+ 357.66666666666663,
+ 499,
+ 424.99999999999994
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "table_footnote",
+ "bbox": [
+ 182,
+ 426,
+ 496,
+ 447
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 190,
+ 424,
+ 495,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 190,
+ 424,
+ 338,
+ 433
+ ],
+ "score": 1.0,
+ "content": "*FairFace (Ours) also defines East (E) Asian, Southeast",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 425,
+ 350,
+ 432
+ ],
+ "score": 0.26,
+ "content": "\\overline { { ( \\mathrm { S E } ) } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 424,
+ 495,
+ 433
+ ],
+ "score": 1.0,
+ "content": "Asian, Middle Eastern (ME), and Western (W) White.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 182,
+ 431,
+ 495,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 182,
+ 431,
+ 495,
+ 441
+ ],
+ "score": 1.0,
+ "content": "**PPB and DiF do not provide race annotations but skin color annotated or automatically computed as a proxy to race.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 382,
+ 439,
+ 495,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 439,
+ 495,
+ 447
+ ],
+ "score": 1.0,
+ "content": "†denotes datasets used in our experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 582
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "Face attribute recognition has been applied as a sub-component to other computer vision tasks such",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 472,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 472,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "as face verification (Kumar et al., 2011) and person re-idenfication (Layne et al., 2012; Li et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "2015a; Su et al., 2018). It is imperative to ensure that these systems perform evenly well on different",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": "gender and race groups. Failing to do so can be detrimental to the reputations of individual service",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "providers and the public trust about the machine learning and computer vision research community.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "Most notable incidents regarding the racial bias include Google Photos recognizing African Ameri-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "can faces as Gorilla and Nikon’s digital cameras prompting a message asking “did someone blink?”",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 504,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 504,
+ 550
+ ],
+ "score": 1.0,
+ "content": "to Asian users (Zhang, 2015). These incidents, regardless of whether the models were trained im-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 548,
+ 506,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 506,
+ 561
+ ],
+ "score": 1.0,
+ "content": "properly or how much they actually affected the users, often result in the termination of the service",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "or features (e.g. dropping sensitive output categories). For this reason, most commercial service",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 308,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 308,
+ 583
+ ],
+ "score": 1.0,
+ "content": "providers have stopped providing a race classifier.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 587,
+ 505,
+ 676
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "Face attribute recognition is also used for demographic surveys performed in marketing or social",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "science research, aimed at understanding human social behaviors and their relations to demographic",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "backgrounds of individuals. Using off-the-shelf tools (Amos et al., 2016; Baltrusaitis et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "and commercial services, social scientists have begun to use images of people to infer their de-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "mographic attributes and analyze their behaviors. Notable examples are demographic analyses of",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "social media users using their photographs (Chakraborty et al., 2017; Reis et al., 2017; Won et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": "2017; Xi et al., 2019; Wang et al., 2017). The cost of unfair classification is huge as it can over- or",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 665,
+ 483,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 483,
+ 677
+ ],
+ "score": 1.0,
+ "content": "under-estimate specific sub-populations in their analysis, which may have policy implications.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 689,
+ 313,
+ 700
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 315,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 315,
+ 702
+ ],
+ "score": 1.0,
+ "content": "2.2 FAIR CLASSIFICATION AND DATASET BIAS",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "AI and machine learning communities have increasingly paid attention to algorithmic fairness and",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "dataset and model biases (Zemel et al., 2013; Corbett-Davies et al., 2017; Zou & Schiebinger, 2018;",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ }
+ ],
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 211,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 213,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 213,
+ 96
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 105,
+ 269,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 271,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 271,
+ 118
+ ],
+ "score": 1.0,
+ "content": "2.1 FACE ATTRIBUTE RECOGNITION",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 126,
+ 505,
+ 192
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 125,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 125,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "The goal of face attribute recognition is to classify various human attributes such as gender, race,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "age, emotions, expressions or other facial traits from facial appearance (Kumar et al., 2011; Joo",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "et al., 2013; Zhang et al., 2015; Liu et al., 2015). Table 1 summarizes the statistics of existing large",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 173
+ ],
+ "score": 1.0,
+ "content": "scale public and in-the-wild face attribute datasets including our new dataset. As stated earlier,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 170,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "most of these datasets were constructed from online sources and are typically dominated by the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 180,
+ 155,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 155,
+ 194
+ ],
+ "score": 1.0,
+ "content": "White race.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 125,
+ 506,
+ 194
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 223,
+ 499,
+ 425
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 216,
+ 200,
+ 394,
+ 212
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 215,
+ 199,
+ 396,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 215,
+ 199,
+ 396,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Table 1: Statistics of Face Attribute Datasets",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 223,
+ 499,
+ 425
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 223,
+ 499,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 223,
+ 499,
+ 425
+ ],
+ "score": 0.98,
+ "html": " | | | | | | | Rac | Annot | ation | | |
| Name | Source | #offaces | In-the-wild? | Age | Gender | White* | Asian* | Bla-ck | Ind-ian | Lat-ino | Balan-ced? |
| WME | E SE |
| PPB(Buolamwini & Gebru,2018) | Gov. OfficialProfiles | 1K | | √ | √ | **Skin color prediction |
| MORPH(Ricanek & Tesafaye,2006) | Public Data | 55K | | √ | √ | merged | | | √ | | √ | no |
| PubFig(Kumar et al.,2011) | Celebrity | 13K | √ | Model generated predictions | no |
| IMDB-WIKI(Rothe et al., 2016) | IMDB,WIKI | 500K | √ | √ | √ | | | | | | | | no |
| FotW(Escalera et al., 2016) | Flickr | 25K | √ | √ | √ | | | | | | | | yes |
| CACD(Chen et al.,2015) | celebrity | 160K | √ | √ | | | | | | | | | no |
| DiF(Merler et al., 2019) | Flickr | 1M | √ | √ | √ | **Skin color prediction |
| +CelebA(Liu et al., 2015) | CelebFaceLFW | 200K | √ | √ | √ | | | | | | | | no |
| LFW+(Han et al., 2018) | LFW(Newspapers) | 15K | √ | √ | √ | merged | merged | | | no |
| LFWA+(Liu et al., 2015) | LFW(Newspapers) | 13K | √ | | √ | merged | merged | √ | √ | | no |
| tUTKFace(Zhang et al., 2017) | MORPH,CACDWeb | 20K | √ | √ | √ | merged | merged | √ | √ | | yes |
| FairFace(Ours) | Flickr, TwitterNewspapers, Web | 108K | √ | √ | √ | √ | √ | √ | √ | √ | √ | √ | yes |
",
+ "type": "table",
+ "image_path": "c2f707704dd0c90c0b1315b6b36ae7c4dbce2da6d6ecc44274964c9284f252ca.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 223,
+ 499,
+ 290.3333333333333
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 108,
+ 290.3333333333333,
+ 499,
+ 357.66666666666663
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 108,
+ 357.66666666666663,
+ 499,
+ 424.99999999999994
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "table_footnote",
+ "bbox": [
+ 182,
+ 426,
+ 496,
+ 447
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 190,
+ 424,
+ 495,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 190,
+ 424,
+ 338,
+ 433
+ ],
+ "score": 1.0,
+ "content": "*FairFace (Ours) also defines East (E) Asian, Southeast",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 425,
+ 350,
+ 432
+ ],
+ "score": 0.26,
+ "content": "\\overline { { ( \\mathrm { S E } ) } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 424,
+ 495,
+ 433
+ ],
+ "score": 1.0,
+ "content": "Asian, Middle Eastern (ME), and Western (W) White.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 182,
+ 431,
+ 495,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 182,
+ 431,
+ 495,
+ 441
+ ],
+ "score": 1.0,
+ "content": "**PPB and DiF do not provide race annotations but skin color annotated or automatically computed as a proxy to race.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 382,
+ 439,
+ 495,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 439,
+ 495,
+ 447
+ ],
+ "score": 1.0,
+ "content": "†denotes datasets used in our experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 582
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "Face attribute recognition has been applied as a sub-component to other computer vision tasks such",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 472,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 472,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "as face verification (Kumar et al., 2011) and person re-idenfication (Layne et al., 2012; Li et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "2015a; Su et al., 2018). It is imperative to ensure that these systems perform evenly well on different",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": "gender and race groups. Failing to do so can be detrimental to the reputations of individual service",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "providers and the public trust about the machine learning and computer vision research community.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "Most notable incidents regarding the racial bias include Google Photos recognizing African Ameri-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "can faces as Gorilla and Nikon’s digital cameras prompting a message asking “did someone blink?”",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 504,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 504,
+ 550
+ ],
+ "score": 1.0,
+ "content": "to Asian users (Zhang, 2015). These incidents, regardless of whether the models were trained im-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 548,
+ 506,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 506,
+ 561
+ ],
+ "score": 1.0,
+ "content": "properly or how much they actually affected the users, often result in the termination of the service",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "or features (e.g. dropping sensitive output categories). For this reason, most commercial service",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 308,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 308,
+ 583
+ ],
+ "score": 1.0,
+ "content": "providers have stopped providing a race classifier.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 105,
+ 461,
+ 506,
+ 583
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 587,
+ 505,
+ 676
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "Face attribute recognition is also used for demographic surveys performed in marketing or social",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "science research, aimed at understanding human social behaviors and their relations to demographic",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "backgrounds of individuals. Using off-the-shelf tools (Amos et al., 2016; Baltrusaitis et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "and commercial services, social scientists have begun to use images of people to infer their de-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "mographic attributes and analyze their behaviors. Notable examples are demographic analyses of",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "social media users using their photographs (Chakraborty et al., 2017; Reis et al., 2017; Won et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": "2017; Xi et al., 2019; Wang et al., 2017). The cost of unfair classification is huge as it can over- or",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 665,
+ 483,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 483,
+ 677
+ ],
+ "score": 1.0,
+ "content": "under-estimate specific sub-populations in their analysis, which may have policy implications.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 587,
+ 506,
+ 677
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 689,
+ 313,
+ 700
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 315,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 315,
+ 702
+ ],
+ "score": 1.0,
+ "content": "2.2 FAIR CLASSIFICATION AND DATASET BIAS",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "AI and machine learning communities have increasingly paid attention to algorithmic fairness and",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "dataset and model biases (Zemel et al., 2013; Corbett-Davies et al., 2017; Zou & Schiebinger, 2018;",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 106,
+ 709,
+ 505,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Zhang et al., 2018). There exist many different definitions of fairness used in the literature (Verma",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "& Rubin, 2018). In this paper, we focus on balanced accuracy–whether the attribute classification",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "accuracy is independent of race and gender. More generally, research in fairness is concerned with",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "a model’s ability to produce fair outcomes (e.g. loan approval) independent of protected or sensitive",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 238,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 238,
+ 139
+ ],
+ "score": 1.0,
+ "content": "attributes such as race or gender.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 143,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Studies in algorithmic fairness have focused on either 1) discovering (auditing) existing bias in",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 154,
+ 505,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 154,
+ 505,
+ 166
+ ],
+ "score": 1.0,
+ "content": "datasets or systems (Shankar et al., 2017; Buolamwini & Gebru, 2018; Kiritchenko & Mohammad,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "2018; McDuff et al., 2019), 2) making a better dataset (Merler et al., 2019; Alvi et al., 2018), or 3)",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "designing a better algorithm or model (Das et al., 2018; Alvi et al., 2018; Ryu et al., 2017; Zemel",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 186,
+ 408,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 186,
+ 408,
+ 200
+ ],
+ "score": 1.0,
+ "content": "et al., 2013; Zafar et al., 2017). Our paper falls into the first two categories.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 505,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "The main task of interest in our paper is (balanced) gender classification from facial images. Buo-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 227
+ ],
+ "score": 1.0,
+ "content": "lamwini & Gebru (2018) demonstrated many commercial gender classification systems are biased",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "score": 1.0,
+ "content": "and least accurate on dark-skinned females. The biased results may be caused by biased datasets,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 232,
+ 250
+ ],
+ "score": 1.0,
+ "content": "such as skewed image origins (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 237,
+ 252,
+ 248
+ ],
+ "score": 0.85,
+ "content": "45 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 237,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "of images are from the U.S. in Imagenet) (Suresh et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 260
+ ],
+ "score": 1.0,
+ "content": "or biased underlying associations between scene and race in images (Stock & Cisse, 2018). It is,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 271
+ ],
+ "score": 1.0,
+ "content": "however, “infeasible to balance across all possible co-occurrences” of attributes (Hendricks et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 269,
+ 270,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 270,
+ 283
+ ],
+ "score": 1.0,
+ "content": "2018), except in a lab-controlled setting.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 364
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Therefore, the contribution of our paper is to mitigate, not entirely solve, the current limitations and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 310
+ ],
+ "score": 1.0,
+ "content": "biases of existing databases by collecting more diverse face images from non-White race groups. We",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "empirically show this significantly improves the generalization performance to novel image datasets",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "whose racial compositions are not dominated by the White race. Furthermore, as shown in Table 1,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 505,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 505,
+ 342
+ ],
+ "score": 1.0,
+ "content": "our dataset is the first large scale in-the-wild face image dataset which includes Southeast Asian",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 341,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "and Middle Eastern races. While their faces share similarity with East Asian and White groups, we",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 473,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 473,
+ 365
+ ],
+ "score": 1.0,
+ "content": "argue that not having these major race groups in datasets is a strong form of discrimination.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 384,
+ 258,
+ 397
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 260,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 260,
+ 399
+ ],
+ "score": 1.0,
+ "content": "3 DATASET CONSTRUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 412,
+ 209,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 210,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 210,
+ 425
+ ],
+ "score": 1.0,
+ "content": "3.1 RACE TAXONOMY",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 434,
+ 505,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Our dataset defines 7 race groups: White, Black, Indian, East Asian, Southeast Asian, Middle East-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "ern, and Latino. Race and ethnicity are different categorizations of humans. Race is defined based",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 469
+ ],
+ "score": 1.0,
+ "content": "on physical traits and ethnicity is based on cultural similarities (Schaefer, 2008). For example, Asian",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "immigrants in Latin America can be of Latino ethnicity. In practice, these two terms are often used",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 477,
+ 174,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 174,
+ 493
+ ],
+ "score": 1.0,
+ "content": "interchangeably.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 495,
+ 505,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "score": 1.0,
+ "content": "We first adopted a commonly accepted race classification from the U.S. Census Bureau (White,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "Black, Asian, Hawaiian and Pacific Islanders, Native Americans, and Latino). Latino is often treated",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 530
+ ],
+ "score": 1.0,
+ "content": "as an ethnicity, but we consider Latino a race, which can be judged from the facial appearance. We",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "then further divided subgroups such as Middle Eastern, East Asian, Southeast Asian, and Indian,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "as they look clearly distinct. During the data collection, we found very few examples for Hawaiian",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "and Pacific Islanders and Native Americans and discarded these categories. All the experiments",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 383,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 383,
+ 574
+ ],
+ "score": 1.0,
+ "content": "conducted in this paper were therefore based on 7 race classification.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 578,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "An important criterion to measure dataset bias is on which basis the bias should be measured: skin",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "color or race? A few recent studies (Buolamwini & Gebru, 2018; Merler et al., 2019) use skin",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "color as a proxy to racial or ethnicity grouping. While skin color can be easily computed without",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 612,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "subjective annotations, it has limitations. First, skin color is heavily affected by illumination and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 634
+ ],
+ "score": 1.0,
+ "content": "light conditions. The Pilot Parliaments Benchmark (PPB) dataset (Buolamwini & Gebru, 2018) only",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "used profile photographs of government officials taken in well controlled lighting, which makes it",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "non-in-the-wild. Second, within-group variations of skin color are huge. Even same individuals can",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "show different skin colors over time. Third, most importantly, race is a multidimensional concept",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "whereas skin color (i.e. brightness) is one dimensional. Figure 5 in Appendix shows the distributions",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "of the skin color of multiple race groups, measured by Individual Typology Angle (ITA) (Wilkes",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 104,
+ 686,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 686,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "et al., 2015). As shown here, the skin color provides no information to differentiate many groups",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "such as East Asian and White. Therefore, we explicitly use race and annotate the physical race by",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "human annotators’ judgments. To complement the limits of race categorization, however, we also",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 477,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 477,
+ 733
+ ],
+ "score": 1.0,
+ "content": "use skin color, measured by ITA, following the same procedure used by Merler et al. (2019).",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "page_idx": 3,
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Zhang et al., 2018). There exist many different definitions of fairness used in the literature (Verma",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "& Rubin, 2018). In this paper, we focus on balanced accuracy–whether the attribute classification",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "accuracy is independent of race and gender. More generally, research in fairness is concerned with",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "a model’s ability to produce fair outcomes (e.g. loan approval) independent of protected or sensitive",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 238,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 238,
+ 139
+ ],
+ "score": 1.0,
+ "content": "attributes such as race or gender.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 139
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 143,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Studies in algorithmic fairness have focused on either 1) discovering (auditing) existing bias in",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 154,
+ 505,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 154,
+ 505,
+ 166
+ ],
+ "score": 1.0,
+ "content": "datasets or systems (Shankar et al., 2017; Buolamwini & Gebru, 2018; Kiritchenko & Mohammad,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "2018; McDuff et al., 2019), 2) making a better dataset (Merler et al., 2019; Alvi et al., 2018), or 3)",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "designing a better algorithm or model (Das et al., 2018; Alvi et al., 2018; Ryu et al., 2017; Zemel",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 186,
+ 408,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 186,
+ 408,
+ 200
+ ],
+ "score": 1.0,
+ "content": "et al., 2013; Zafar et al., 2017). Our paper falls into the first two categories.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 105,
+ 143,
+ 505,
+ 200
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 505,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "The main task of interest in our paper is (balanced) gender classification from facial images. Buo-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 227
+ ],
+ "score": 1.0,
+ "content": "lamwini & Gebru (2018) demonstrated many commercial gender classification systems are biased",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "score": 1.0,
+ "content": "and least accurate on dark-skinned females. The biased results may be caused by biased datasets,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 232,
+ 250
+ ],
+ "score": 1.0,
+ "content": "such as skewed image origins (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 237,
+ 252,
+ 248
+ ],
+ "score": 0.85,
+ "content": "45 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 237,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "of images are from the U.S. in Imagenet) (Suresh et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 260
+ ],
+ "score": 1.0,
+ "content": "or biased underlying associations between scene and race in images (Stock & Cisse, 2018). It is,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 271
+ ],
+ "score": 1.0,
+ "content": "however, “infeasible to balance across all possible co-occurrences” of attributes (Hendricks et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 269,
+ 270,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 270,
+ 283
+ ],
+ "score": 1.0,
+ "content": "2018), except in a lab-controlled setting.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 204,
+ 506,
+ 283
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 364
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Therefore, the contribution of our paper is to mitigate, not entirely solve, the current limitations and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 310
+ ],
+ "score": 1.0,
+ "content": "biases of existing databases by collecting more diverse face images from non-White race groups. We",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "empirically show this significantly improves the generalization performance to novel image datasets",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "whose racial compositions are not dominated by the White race. Furthermore, as shown in Table 1,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 505,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 505,
+ 342
+ ],
+ "score": 1.0,
+ "content": "our dataset is the first large scale in-the-wild face image dataset which includes Southeast Asian",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 341,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "and Middle Eastern races. While their faces share similarity with East Asian and White groups, we",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 473,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 473,
+ 365
+ ],
+ "score": 1.0,
+ "content": "argue that not having these major race groups in datasets is a strong form of discrimination.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 105,
+ 286,
+ 505,
+ 365
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 384,
+ 258,
+ 397
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 260,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 260,
+ 399
+ ],
+ "score": 1.0,
+ "content": "3 DATASET CONSTRUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 412,
+ 209,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 210,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 210,
+ 425
+ ],
+ "score": 1.0,
+ "content": "3.1 RACE TAXONOMY",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 434,
+ 505,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Our dataset defines 7 race groups: White, Black, Indian, East Asian, Southeast Asian, Middle East-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "ern, and Latino. Race and ethnicity are different categorizations of humans. Race is defined based",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 469
+ ],
+ "score": 1.0,
+ "content": "on physical traits and ethnicity is based on cultural similarities (Schaefer, 2008). For example, Asian",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "immigrants in Latin America can be of Latino ethnicity. In practice, these two terms are often used",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 477,
+ 174,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 174,
+ 493
+ ],
+ "score": 1.0,
+ "content": "interchangeably.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 435,
+ 505,
+ 493
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 495,
+ 505,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "score": 1.0,
+ "content": "We first adopted a commonly accepted race classification from the U.S. Census Bureau (White,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "Black, Asian, Hawaiian and Pacific Islanders, Native Americans, and Latino). Latino is often treated",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 530
+ ],
+ "score": 1.0,
+ "content": "as an ethnicity, but we consider Latino a race, which can be judged from the facial appearance. We",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "then further divided subgroups such as Middle Eastern, East Asian, Southeast Asian, and Indian,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "as they look clearly distinct. During the data collection, we found very few examples for Hawaiian",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "and Pacific Islanders and Native Americans and discarded these categories. All the experiments",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 383,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 383,
+ 574
+ ],
+ "score": 1.0,
+ "content": "conducted in this paper were therefore based on 7 race classification.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 495,
+ 505,
+ 574
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 578,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "An important criterion to measure dataset bias is on which basis the bias should be measured: skin",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "color or race? A few recent studies (Buolamwini & Gebru, 2018; Merler et al., 2019) use skin",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "color as a proxy to racial or ethnicity grouping. While skin color can be easily computed without",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 612,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "subjective annotations, it has limitations. First, skin color is heavily affected by illumination and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 634
+ ],
+ "score": 1.0,
+ "content": "light conditions. The Pilot Parliaments Benchmark (PPB) dataset (Buolamwini & Gebru, 2018) only",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "used profile photographs of government officials taken in well controlled lighting, which makes it",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "non-in-the-wild. Second, within-group variations of skin color are huge. Even same individuals can",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "show different skin colors over time. Third, most importantly, race is a multidimensional concept",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "whereas skin color (i.e. brightness) is one dimensional. Figure 5 in Appendix shows the distributions",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "of the skin color of multiple race groups, measured by Individual Typology Angle (ITA) (Wilkes",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 104,
+ 686,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 686,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "et al., 2015). As shown here, the skin color provides no information to differentiate many groups",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "such as East Asian and White. Therefore, we explicitly use race and annotate the physical race by",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "human annotators’ judgments. To complement the limits of race categorization, however, we also",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 477,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 477,
+ 733
+ ],
+ "score": 1.0,
+ "content": "use skin color, measured by ITA, following the same procedure used by Merler et al. (2019).",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 104,
+ 578,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 300,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 302,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 302,
+ 95
+ ],
+ "score": 1.0,
+ "content": "3.2 IMAGE COLLECTION AND ANNOTATION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 103,
+ 505,
+ 191
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": "Many existing face datasets have been sourced from photographs of public figures such as politicians",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "or celebrities (Kumar et al., 2011; Huang et al., 2007; Joo et al., 2015; Rothe et al., 2016; Liu",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "et al., 2015). Despite the easiness of collecting images and ground truth attributes, the selection",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "of these populations may be biased. For example, politicians may be older and actors may be more",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 147,
+ 505,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 505,
+ 160
+ ],
+ "score": 1.0,
+ "content": "attractive than typical faces. Their images are usually taken by professional photographers in limited",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "situations, leading to the quality bias. Some datasets were collected via web search using keywords",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "such as “Asian boy” (Zhang et al., 2017). These queries may return only stereotypical faces or",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 180,
+ 480,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 180,
+ 480,
+ 193
+ ],
+ "score": 1.0,
+ "content": "prioritize celebrities in those categories rather than diverse individuals among general public.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 197,
+ 505,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Our goal is to minimize the selection bias introduced by such filtering and maximize the diversity",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "and coverage of the dataset. We started from a huge public image dataset, Yahoo YFCC100M",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "dataset (Thomee et al.), and detected faces from the images without any preselection. A recent",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 229,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 229,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "work also used the same dataset to construct a huge unfiltered face dataset (Diversity in Faces, DiF)",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 451,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 451,
+ 253
+ ],
+ "score": 1.0,
+ "content": "(Merler et al., 2019). Our dataset is smaller but more balanced on race (See Figure 1).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 257,
+ 505,
+ 379
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 504,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 504,
+ 270
+ ],
+ "score": 1.0,
+ "content": "For an efficient collection, we incrementally increased the dataset size. We first detected and anno-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "tated 7,125 faces randomly sampled from the entire YFCC100M dataset ignoring the locations of",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "images. After obtaining annotations on this initial set, we estimated demographic compositions of",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "each country. Based on this statistic, we adaptively adjusted the number of images for each country",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "sampled from the dataset such that the dataset is not dominated by the White race. Consequently,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "we excluded the U.S. and European countries in the later stage of data collection after we sampled",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 336
+ ],
+ "score": 1.0,
+ "content": "enough White faces from those countries. The minimum size of a detected face was set to 50 by",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "50 pixels. This is a relatively smaller size compared to other datasets, but we find the attributes",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "are still recognizable and these examples can actually make the classifiers more robust against noisy",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "data. We only used images with “Attribution” and “Share Alike” Creative Commons licenses, which",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 367,
+ 293,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 293,
+ 380
+ ],
+ "score": 1.0,
+ "content": "allow derivative work and commercial usages.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 384,
+ 505,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "We used Amazon Mechanical Turk to annotate the race, gender and age group for each face. We",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "assigned three workers for each image. If two or three workers agreed on their judgements, we",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "took the values as ground-truth. If all three workers produced different responses, we republished",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "the image to another 3 workers and subsequently discarded the image if the new annotators did not",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 427,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 427,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "agree. These annotations at this stage were still noisy. We further refined the annotations by training",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "a model from the initial ground truth annotations and applying back to the dataset. We then manually",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 467,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 467,
+ 462
+ ],
+ "score": 1.0,
+ "content": "re-verified the annotations for images whose annotations differed from model predictions.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 478,
+ 200,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 201,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 201,
+ 493
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 504,
+ 267,
+ 515
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 269,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 269,
+ 517
+ ],
+ "score": 1.0,
+ "content": "4.1 MEASURING BIAS IN DATASETS",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 525,
+ 505,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "We first measure how skewed each dataset is in terms of its race composition. For the datasets with",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "race annotations, we use the reported statistics. For the other datasets, we annotated the race labels",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "for 3,000 random samples drawn from each dataset. See Figure 1 for the result. As expected, most",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 571
+ ],
+ "score": 1.0,
+ "content": "existing face attribute datasets, especially the ones focusing on celebrities or politicians, are biased",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 568,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "toward the White race. Unlike race, we find that most datasets are relatively more balanced on",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 580,
+ 278,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 580,
+ 190,
+ 592
+ ],
+ "score": 1.0,
+ "content": "gender ranging from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 191,
+ 580,
+ 210,
+ 591
+ ],
+ "score": 0.78,
+ "content": "40 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 580,
+ 212,
+ 592
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 580,
+ 232,
+ 591
+ ],
+ "score": 0.62,
+ "content": "60 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 580,
+ 278,
+ 592
+ ],
+ "score": 1.0,
+ "content": "male ratio.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 605,
+ 328,
+ 617
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 329,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 329,
+ 618
+ ],
+ "score": 1.0,
+ "content": "4.2 MODEL AND CROSS-DATASET PERFORMANCE",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 626,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "To compare model performance of different datasets, we used an identical model architecture,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "ResNet-34 (He et al., 2016), to be trained from each dataset. We used ADAM optimization (Kingma",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "& Ba, 2014) with a learning rate of 0.0001. Given an image, we detected faces using the dlib’s",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "(dlib.net) CNN-based face detector (King, 2015) and ran the attribute classifier on each face. The",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 242,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 242,
+ 683
+ ],
+ "score": 1.0,
+ "content": "experiment was done in PyTorch.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 688,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Throughout the evaluations, we compare our dataset with three other datasets: UTKFace (Zhang",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 425,
+ 711
+ ],
+ "score": 1.0,
+ "content": "et al., 2017), LFWA+, and CelebA (Liu et al., 2015). Both UTKFace and LFWA",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 700,
+ 432,
+ 709
+ ],
+ "score": 0.33,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "have race annota-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "tions, and thus, are suitable for comparison with our dataset. CelebA does not have race annotations,",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 489,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 489,
+ 732
+ ],
+ "score": 1.0,
+ "content": "so we only use it for gender classification. See Table 1 for more detailed dataset characteristics.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 47.5
+ }
+ ],
+ "page_idx": 4,
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 300,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 302,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 302,
+ 95
+ ],
+ "score": 1.0,
+ "content": "3.2 IMAGE COLLECTION AND ANNOTATION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 103,
+ 505,
+ 191
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": "Many existing face datasets have been sourced from photographs of public figures such as politicians",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "or celebrities (Kumar et al., 2011; Huang et al., 2007; Joo et al., 2015; Rothe et al., 2016; Liu",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "et al., 2015). Despite the easiness of collecting images and ground truth attributes, the selection",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "of these populations may be biased. For example, politicians may be older and actors may be more",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 147,
+ 505,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 505,
+ 160
+ ],
+ "score": 1.0,
+ "content": "attractive than typical faces. Their images are usually taken by professional photographers in limited",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "situations, leading to the quality bias. Some datasets were collected via web search using keywords",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "such as “Asian boy” (Zhang et al., 2017). These queries may return only stereotypical faces or",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 180,
+ 480,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 180,
+ 480,
+ 193
+ ],
+ "score": 1.0,
+ "content": "prioritize celebrities in those categories rather than diverse individuals among general public.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 104,
+ 506,
+ 193
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 197,
+ 505,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Our goal is to minimize the selection bias introduced by such filtering and maximize the diversity",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "and coverage of the dataset. We started from a huge public image dataset, Yahoo YFCC100M",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "dataset (Thomee et al.), and detected faces from the images without any preselection. A recent",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 229,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 229,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "work also used the same dataset to construct a huge unfiltered face dataset (Diversity in Faces, DiF)",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 451,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 451,
+ 253
+ ],
+ "score": 1.0,
+ "content": "(Merler et al., 2019). Our dataset is smaller but more balanced on race (See Figure 1).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 196,
+ 506,
+ 253
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 257,
+ 505,
+ 379
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 504,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 504,
+ 270
+ ],
+ "score": 1.0,
+ "content": "For an efficient collection, we incrementally increased the dataset size. We first detected and anno-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "tated 7,125 faces randomly sampled from the entire YFCC100M dataset ignoring the locations of",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "images. After obtaining annotations on this initial set, we estimated demographic compositions of",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "each country. Based on this statistic, we adaptively adjusted the number of images for each country",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "sampled from the dataset such that the dataset is not dominated by the White race. Consequently,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "we excluded the U.S. and European countries in the later stage of data collection after we sampled",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 336
+ ],
+ "score": 1.0,
+ "content": "enough White faces from those countries. The minimum size of a detected face was set to 50 by",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "50 pixels. This is a relatively smaller size compared to other datasets, but we find the attributes",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "are still recognizable and these examples can actually make the classifiers more robust against noisy",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "data. We only used images with “Attribution” and “Share Alike” Creative Commons licenses, which",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 367,
+ 293,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 293,
+ 380
+ ],
+ "score": 1.0,
+ "content": "allow derivative work and commercial usages.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 258,
+ 506,
+ 380
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 384,
+ 505,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "We used Amazon Mechanical Turk to annotate the race, gender and age group for each face. We",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "assigned three workers for each image. If two or three workers agreed on their judgements, we",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "took the values as ground-truth. If all three workers produced different responses, we republished",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "the image to another 3 workers and subsequently discarded the image if the new annotators did not",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 427,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 427,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "agree. These annotations at this stage were still noisy. We further refined the annotations by training",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "a model from the initial ground truth annotations and applying back to the dataset. We then manually",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 467,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 467,
+ 462
+ ],
+ "score": 1.0,
+ "content": "re-verified the annotations for images whose annotations differed from model predictions.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 104,
+ 384,
+ 506,
+ 462
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 478,
+ 200,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 201,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 201,
+ 493
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 504,
+ 267,
+ 515
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 269,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 269,
+ 517
+ ],
+ "score": 1.0,
+ "content": "4.1 MEASURING BIAS IN DATASETS",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 525,
+ 505,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "We first measure how skewed each dataset is in terms of its race composition. For the datasets with",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "race annotations, we use the reported statistics. For the other datasets, we annotated the race labels",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "for 3,000 random samples drawn from each dataset. See Figure 1 for the result. As expected, most",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 571
+ ],
+ "score": 1.0,
+ "content": "existing face attribute datasets, especially the ones focusing on celebrities or politicians, are biased",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 568,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "toward the White race. Unlike race, we find that most datasets are relatively more balanced on",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 580,
+ 278,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 580,
+ 190,
+ 592
+ ],
+ "score": 1.0,
+ "content": "gender ranging from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 191,
+ 580,
+ 210,
+ 591
+ ],
+ "score": 0.78,
+ "content": "40 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 580,
+ 212,
+ 592
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 580,
+ 232,
+ 591
+ ],
+ "score": 0.62,
+ "content": "60 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 580,
+ 278,
+ 592
+ ],
+ "score": 1.0,
+ "content": "male ratio.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 525,
+ 506,
+ 592
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 605,
+ 328,
+ 617
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 329,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 329,
+ 618
+ ],
+ "score": 1.0,
+ "content": "4.2 MODEL AND CROSS-DATASET PERFORMANCE",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 626,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "To compare model performance of different datasets, we used an identical model architecture,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "ResNet-34 (He et al., 2016), to be trained from each dataset. We used ADAM optimization (Kingma",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "& Ba, 2014) with a learning rate of 0.0001. Given an image, we detected faces using the dlib’s",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "(dlib.net) CNN-based face detector (King, 2015) and ran the attribute classifier on each face. The",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 242,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 242,
+ 683
+ ],
+ "score": 1.0,
+ "content": "experiment was done in PyTorch.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 627,
+ 505,
+ 683
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 688,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Throughout the evaluations, we compare our dataset with three other datasets: UTKFace (Zhang",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 425,
+ 711
+ ],
+ "score": 1.0,
+ "content": "et al., 2017), LFWA+, and CelebA (Liu et al., 2015). Both UTKFace and LFWA",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 700,
+ 432,
+ 709
+ ],
+ "score": 0.33,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "have race annota-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "tions, and thus, are suitable for comparison with our dataset. CelebA does not have race annotations,",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 489,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 489,
+ 732
+ ],
+ "score": 1.0,
+ "content": "so we only use it for gender classification. See Table 1 for more detailed dataset characteristics.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 47.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 92,
+ 506,
+ 149
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 177,
+ 81,
+ 432,
+ 91
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 177,
+ 78,
+ 434,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 177,
+ 78,
+ 434,
+ 94
+ ],
+ "score": 1.0,
+ "content": "Table 2: Cross-Dataset Classification Accuracy on White Race.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 92,
+ 506,
+ 149
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 92,
+ 506,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 92,
+ 506,
+ 149
+ ],
+ "score": 0.978,
+ "html": "| ICD-9 Code | Description |
| 398.91 | Rheumatic heart failure (congestive) |
| 402.01 | Malignant hypertensive heart disease with heart failure |
| 402.11 | Benign hypertensive heart disease with heart failure |
| 402.91 | Unspecified hypertensive heart disease with heart failure |
| 404.01 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.03 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and with chronic kidney disease stage V or end stage renal disease |
| 404.11 | Hypertensive heart and chronic kidney disease, benign, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.13 | Hypertensive heart and chronic kidney disease, benign, with heart failure and chronickidney disease stage V or end stage renal disease |
| 404.91 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure and with chronic kidney disease stage I through stage IV, or unspecified |
| 404.93 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure andchronic kidney disease stage V or end stage renal disease |
| 428.0 | Congestive heart failure, unspecified |
| 428.1 | Leftheart failure |
| 428.20 | Systolic heart failure, unspecified |
| 428.21 | Acute systolic heart failure |
| 428.22 | Chronic systolic heart failure |
| 428.23 | Acute on chronic systolic heart failure |
| 428.30 | Diastolic heart failure, unspecified |
| 428.31 | Acute diastolic heart failure |
| 428.32 | Chronic diastolic heart failure |
| 428.33 | Acute on chronic diastolic heart failure |
| 428.40 | Combined systolic and diastolic heart failure, unspecified |
| 428.41 | Acute combined systolic and diastolic heart failure |
| 428.42 | Chronic combined systolic and diastolic heart failure |
| 428.43 | Acute on chronic combined systolic and diastolic heart failure |
| 428.9 | Heart failure, unspecified |
+
+# C HYPER-PARAMETER TUNING
+
+We define five hyper-parameters for GRAM:
+
+• dimensionality $m$ of the basic embedding $\mathbf { e } _ { i }$ : [100, 200, 300, 400, 500] • dimensionality $r$ of the RNN hidden layer $\mathbf { h } _ { t }$ from Eq. (4): [100, 200, 300, 400, 500] • dimensionality $l$ of $\mathbf { W } _ { a }$ and $ { \mathbf { b } } _ { a }$ from Eq. (3): [100, 200, 300, 400, 500] • $L _ { 2 }$ regularization coefficient for all weights except RNN weights: [0.1, 0.01, 0.001, 0.0001] • dropout rate for the dropout on the RNN hidden layer: [0.0, 0.2, 0.4, 0.6, 0.8]
+
+We performed 100 iterations of the random search by using the above ranges for each of the three prediction experiments. For sequential diagnoses prediction on Sutter data, we used $10 \%$ of the training data to tune the hyper-parameters to balance the time and search space. To match the baselines’ number of parameters to GRAM’s, we add 550 to the list of $m$ ’s possible values. This will make the baseline’s largest possible number of parameters comparable to the GRAM’s largest possible number of parameters.
+
+For SimpleRollUp and RollUpRare, the number of input codes is smaller than other models due to the grouping. Therefore, to match their largest possible number of parameters to GRAM’s, we need to add much larger values to $m$ . However, after preliminary experiments, as expected, setting $m$ to too large a value degraded the performance due to overfitting. Since the number of input codes decreased due to the grouping, increasing the dimensionality of $\mathbf { e } _ { i }$ is not a logical thing to do. Therefore, for SimpleRollUp and RollUpRare, we use the same list of values for $m$ as other baselines.
+
+Table 5: Hyper-parameters used by the models in each predictive modeling experiments
+
+| ICD-9 Code | Description |
| 398.91 | Rheumatic heart failure (congestive) |
| 402.01 | Malignant hypertensive heart disease with heart failure |
| 402.11 | Benign hypertensive heart disease with heart failure |
| 402.91 | Unspecified hypertensive heart disease with heart failure |
| 404.01 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.03 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and with chronic kidney disease stage V or end stage renal disease |
| 404.11 | Hypertensive heart and chronic kidney disease, benign, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.13 | Hypertensive heart and chronic kidney disease, benign, with heart failure and chronickidney disease stage V or end stage renal disease |
| 404.91 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure and with chronic kidney disease stage I through stage IV, or unspecified |
| 404.93 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure andchronic kidney disease stage V or end stage renal disease |
| 428.0 | Congestive heart failure, unspecified |
| 428.1 | Leftheart failure |
| 428.20 | Systolic heart failure, unspecified |
| 428.21 | Acute systolic heart failure |
| 428.22 | Chronic systolic heart failure |
| 428.23 | Acute on chronic systolic heart failure |
| 428.30 | Diastolic heart failure, unspecified |
| 428.31 | Acute diastolic heart failure |
| 428.32 | Chronic diastolic heart failure |
| 428.33 | Acute on chronic diastolic heart failure |
| 428.40 | Combined systolic and diastolic heart failure, unspecified |
| 428.41 | Acute combined systolic and diastolic heart failure |
| 428.42 | Chronic combined systolic and diastolic heart failure |
| 428.43 | Acute on chronic combined systolic and diastolic heart failure |
| 428.9 | Heart failure, unspecified |
",
+ "bbox": [
+ 196,
+ 79,
+ 797,
+ 488
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 174,
+ 539,
+ 825,
+ 594
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "C HYPER-PARAMETER TUNING ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 614,
+ 449,
+ 631
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "We define five hyper-parameters for GRAM: ",
+ "bbox": [
+ 176,
+ 645,
+ 455,
+ 660
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "• dimensionality $m$ of the basic embedding $\\mathbf { e } _ { i }$ : [100, 200, 300, 400, 500] • dimensionality $r$ of the RNN hidden layer $\\mathbf { h } _ { t }$ from Eq. (4): [100, 200, 300, 400, 500] • dimensionality $l$ of $\\mathbf { W } _ { a }$ and $ { \\mathbf { b } } _ { a }$ from Eq. (3): [100, 200, 300, 400, 500] • $L _ { 2 }$ regularization coefficient for all weights except RNN weights: [0.1, 0.01, 0.001, 0.0001] • dropout rate for the dropout on the RNN hidden layer: [0.0, 0.2, 0.4, 0.6, 0.8] ",
+ "bbox": [
+ 181,
+ 670,
+ 802,
+ 760
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "We performed 100 iterations of the random search by using the above ranges for each of the three prediction experiments. For sequential diagnoses prediction on Sutter data, we used $10 \\%$ of the training data to tune the hyper-parameters to balance the time and search space. To match the baselines’ number of parameters to GRAM’s, we add 550 to the list of $m$ ’s possible values. This will make the baseline’s largest possible number of parameters comparable to the GRAM’s largest possible number of parameters. ",
+ "bbox": [
+ 174,
+ 770,
+ 825,
+ 853
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "For SimpleRollUp and RollUpRare, the number of input codes is smaller than other models due to the grouping. Therefore, to match their largest possible number of parameters to GRAM’s, we need to add much larger values to $m$ . However, after preliminary experiments, as expected, setting $m$ to too large a value degraded the performance due to overfitting. Since the number of input codes decreased due to the grouping, increasing the dimensionality of $\\mathbf { e } _ { i }$ is not a logical thing to do. Therefore, for SimpleRollUp and RollUpRare, we use the same list of values for $m$ as other baselines. ",
+ "bbox": [
+ 176,
+ 859,
+ 823,
+ 901
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "table",
+ "img_path": "images/85e8dbbb6270b8ac2c0437d095afab49764c1b8875e71a1cba24aac6fee9d5d6.jpg",
+ "table_caption": [
+ "Table 5: Hyper-parameters used by the models in each predictive modeling experiments "
+ ],
+ "table_footnote": [],
+ "table_body": "| ICD-9 Code | Description |
| 398.91 | Rheumatic heart failure (congestive) |
| 402.01 | Malignant hypertensive heart disease with heart failure |
| 402.11 | Benign hypertensive heart disease with heart failure |
| 402.91 | Unspecified hypertensive heart disease with heart failure |
| 404.01 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.03 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and with chronic kidney disease stage V or end stage renal disease |
| 404.11 | Hypertensive heart and chronic kidney disease, benign, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.13 | Hypertensive heart and chronic kidney disease, benign, with heart failure and chronickidney disease stage V or end stage renal disease |
| 404.91 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure and with chronic kidney disease stage I through stage IV, or unspecified |
| 404.93 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure andchronic kidney disease stage V or end stage renal disease |
| 428.0 | Congestive heart failure, unspecified |
| 428.1 | Leftheart failure |
| 428.20 | Systolic heart failure, unspecified |
| 428.21 | Acute systolic heart failure |
| 428.22 | Chronic systolic heart failure |
| 428.23 | Acute on chronic systolic heart failure |
| 428.30 | Diastolic heart failure, unspecified |
| 428.31 | Acute diastolic heart failure |
| 428.32 | Chronic diastolic heart failure |
| 428.33 | Acute on chronic diastolic heart failure |
| 428.40 | Combined systolic and diastolic heart failure, unspecified |
| 428.41 | Acute combined systolic and diastolic heart failure |
| 428.42 | Chronic combined systolic and diastolic heart failure |
| 428.43 | Acute on chronic combined systolic and diastolic heart failure |
| 428.9 | Heart failure, unspecified |
",
+ "type": "table",
+ "image_path": "e5968d93609e6441fd2ec790f233442c1d47389cbc1c60c9ba33c95298a5628e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 120,
+ 63,
+ 488,
+ 171.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 120,
+ 171.0,
+ 488,
+ 279.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 120,
+ 279.0,
+ 488,
+ 387.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 206,
+ 394,
+ 405,
+ 407
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 206,
+ 394,
+ 405,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 206,
+ 394,
+ 405,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Table 4: Qualifying ICD-9 codes for heart failure",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 427,
+ 505,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "their corresponding case. Control subjects were required to have their first office encounter within",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "one year of the matching HF case patient’s first office visit, and have at least one office encounter 30",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "days before or any time after the case’s HFDx date to ensure similar duration of observations among",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 183,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 183,
+ 470
+ ],
+ "score": 1.0,
+ "content": "cases and controls.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 487,
+ 275,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 277,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 277,
+ 501
+ ],
+ "score": 1.0,
+ "content": "C HYPER-PARAMETER TUNING",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 511,
+ 279,
+ 523
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 280,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 280,
+ 524
+ ],
+ "score": 1.0,
+ "content": "We define five hyper-parameters for GRAM:",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 111,
+ 531,
+ 491,
+ 602
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 532,
+ 408,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 532,
+ 183,
+ 545
+ ],
+ "score": 1.0,
+ "content": "• dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 534,
+ 194,
+ 542
+ ],
+ "score": 0.74,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 532,
+ 290,
+ 545
+ ],
+ "score": 1.0,
+ "content": "of the basic embedding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 533,
+ 300,
+ 543
+ ],
+ "score": 0.75,
+ "content": "\\mathbf { e } _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 532,
+ 408,
+ 545
+ ],
+ "score": 1.0,
+ "content": ": [100, 200, 300, 400, 500]",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 546,
+ 464,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 546,
+ 183,
+ 559
+ ],
+ "score": 1.0,
+ "content": "• dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 549,
+ 190,
+ 556
+ ],
+ "score": 0.77,
+ "content": "r",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 546,
+ 291,
+ 559
+ ],
+ "score": 1.0,
+ "content": "of the RNN hidden layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 547,
+ 303,
+ 558
+ ],
+ "score": 0.88,
+ "content": "\\mathbf { h } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 546,
+ 464,
+ 559
+ ],
+ "score": 1.0,
+ "content": "from Eq. (4): [100, 200, 300, 400, 500]",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 111,
+ 561,
+ 410,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 561,
+ 183,
+ 574
+ ],
+ "score": 1.0,
+ "content": "• dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 562,
+ 189,
+ 571
+ ],
+ "score": 0.66,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 561,
+ 200,
+ 574
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 561,
+ 218,
+ 573
+ ],
+ "score": 0.91,
+ "content": "\\mathbf { W } _ { a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 561,
+ 236,
+ 574
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 561,
+ 249,
+ 573
+ ],
+ "score": 0.89,
+ "content": " { \\mathbf { b } } _ { a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 561,
+ 410,
+ 574
+ ],
+ "score": 1.0,
+ "content": "from Eq. (3): [100, 200, 300, 400, 500]",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 112,
+ 576,
+ 491,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 576,
+ 121,
+ 588
+ ],
+ "score": 1.0,
+ "content": "•",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 576,
+ 135,
+ 587
+ ],
+ "score": 0.85,
+ "content": "L _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 576,
+ 491,
+ 588
+ ],
+ "score": 1.0,
+ "content": "regularization coefficient for all weights except RNN weights: [0.1, 0.01, 0.001, 0.0001]",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 590,
+ 433,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 590,
+ 433,
+ 603
+ ],
+ "score": 1.0,
+ "content": "• dropout rate for the dropout on the RNN hidden layer: [0.0, 0.2, 0.4, 0.6, 0.8]",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 610,
+ 505,
+ 676
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "We performed 100 iterations of the random search by using the above ranges for each of the three",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 457,
+ 633
+ ],
+ "score": 1.0,
+ "content": "prediction experiments. For sequential diagnoses prediction on Sutter data, we used",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 622,
+ 477,
+ 632
+ ],
+ "score": 0.86,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 621,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 644
+ ],
+ "score": 1.0,
+ "content": "training data to tune the hyper-parameters to balance the time and search space. To match the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 383,
+ 655
+ ],
+ "score": 1.0,
+ "content": "baselines’ number of parameters to GRAM’s, we add 550 to the list of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 644,
+ 393,
+ 653
+ ],
+ "score": 0.68,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 642,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "’s possible values. This will",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "make the baseline’s largest possible number of parameters comparable to the GRAM’s largest possible",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 664,
+ 199,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 199,
+ 676
+ ],
+ "score": 1.0,
+ "content": "number of parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 681,
+ 504,
+ 714
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "score": 1.0,
+ "content": "For SimpleRollUp and RollUpRare, the number of input codes is smaller than other models due to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 693,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 693,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "the grouping. Therefore, to match their largest possible number of parameters to GRAM’s, we need to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 702,
+ 505,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 702,
+ 211,
+ 715
+ ],
+ "score": 1.0,
+ "content": "add much larger values to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 705,
+ 221,
+ 712
+ ],
+ "score": 0.67,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 221,
+ 702,
+ 468,
+ 715
+ ],
+ "score": 1.0,
+ "content": ". However, after preliminary experiments, as expected, setting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 704,
+ 478,
+ 712
+ ],
+ "score": 0.73,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 702,
+ 505,
+ 715
+ ],
+ "score": 1.0,
+ "content": "to too",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "page_idx": 11,
+ "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 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 733,
+ 311,
+ 742
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 312,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 312,
+ 746
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 120,
+ 63,
+ 488,
+ 387
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 120,
+ 63,
+ 488,
+ 387
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 63,
+ 488,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 63,
+ 488,
+ 387
+ ],
+ "score": 0.985,
+ "html": "| ICD-9 Code | Description |
| 398.91 | Rheumatic heart failure (congestive) |
| 402.01 | Malignant hypertensive heart disease with heart failure |
| 402.11 | Benign hypertensive heart disease with heart failure |
| 402.91 | Unspecified hypertensive heart disease with heart failure |
| 404.01 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.03 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and with chronic kidney disease stage V or end stage renal disease |
| 404.11 | Hypertensive heart and chronic kidney disease, benign, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.13 | Hypertensive heart and chronic kidney disease, benign, with heart failure and chronickidney disease stage V or end stage renal disease |
| 404.91 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure and with chronic kidney disease stage I through stage IV, or unspecified |
| 404.93 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure andchronic kidney disease stage V or end stage renal disease |
| 428.0 | Congestive heart failure, unspecified |
| 428.1 | Leftheart failure |
| 428.20 | Systolic heart failure, unspecified |
| 428.21 | Acute systolic heart failure |
| 428.22 | Chronic systolic heart failure |
| 428.23 | Acute on chronic systolic heart failure |
| 428.30 | Diastolic heart failure, unspecified |
| 428.31 | Acute diastolic heart failure |
| 428.32 | Chronic diastolic heart failure |
| 428.33 | Acute on chronic diastolic heart failure |
| 428.40 | Combined systolic and diastolic heart failure, unspecified |
| 428.41 | Acute combined systolic and diastolic heart failure |
| 428.42 | Chronic combined systolic and diastolic heart failure |
| 428.43 | Acute on chronic combined systolic and diastolic heart failure |
| 428.9 | Heart failure, unspecified |
",
+ "type": "table",
+ "image_path": "e5968d93609e6441fd2ec790f233442c1d47389cbc1c60c9ba33c95298a5628e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 120,
+ 63,
+ 488,
+ 171.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 120,
+ 171.0,
+ 488,
+ 279.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 120,
+ 279.0,
+ 488,
+ 387.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 206,
+ 394,
+ 405,
+ 407
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 206,
+ 394,
+ 405,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 206,
+ 394,
+ 405,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Table 4: Qualifying ICD-9 codes for heart failure",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 427,
+ 505,
+ 471
+ ],
+ "lines": [],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 428,
+ 506,
+ 470
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 487,
+ 275,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 277,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 277,
+ 501
+ ],
+ "score": 1.0,
+ "content": "C HYPER-PARAMETER TUNING",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 511,
+ 279,
+ 523
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 280,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 280,
+ 524
+ ],
+ "score": 1.0,
+ "content": "We define five hyper-parameters for GRAM:",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 106,
+ 511,
+ 280,
+ 524
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 111,
+ 531,
+ 491,
+ 602
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 532,
+ 408,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 532,
+ 183,
+ 545
+ ],
+ "score": 1.0,
+ "content": "• dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 534,
+ 194,
+ 542
+ ],
+ "score": 0.74,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 532,
+ 290,
+ 545
+ ],
+ "score": 1.0,
+ "content": "of the basic embedding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 533,
+ 300,
+ 543
+ ],
+ "score": 0.75,
+ "content": "\\mathbf { e } _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 532,
+ 408,
+ 545
+ ],
+ "score": 1.0,
+ "content": ": [100, 200, 300, 400, 500]",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 546,
+ 464,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 546,
+ 183,
+ 559
+ ],
+ "score": 1.0,
+ "content": "• dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 549,
+ 190,
+ 556
+ ],
+ "score": 0.77,
+ "content": "r",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 546,
+ 291,
+ 559
+ ],
+ "score": 1.0,
+ "content": "of the RNN hidden layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 547,
+ 303,
+ 558
+ ],
+ "score": 0.88,
+ "content": "\\mathbf { h } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 546,
+ 464,
+ 559
+ ],
+ "score": 1.0,
+ "content": "from Eq. (4): [100, 200, 300, 400, 500]",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 111,
+ 561,
+ 410,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 561,
+ 183,
+ 574
+ ],
+ "score": 1.0,
+ "content": "• dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 562,
+ 189,
+ 571
+ ],
+ "score": 0.66,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 561,
+ 200,
+ 574
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 561,
+ 218,
+ 573
+ ],
+ "score": 0.91,
+ "content": "\\mathbf { W } _ { a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 561,
+ 236,
+ 574
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 561,
+ 249,
+ 573
+ ],
+ "score": 0.89,
+ "content": " { \\mathbf { b } } _ { a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 561,
+ 410,
+ 574
+ ],
+ "score": 1.0,
+ "content": "from Eq. (3): [100, 200, 300, 400, 500]",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 112,
+ 576,
+ 491,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 576,
+ 121,
+ 588
+ ],
+ "score": 1.0,
+ "content": "•",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 576,
+ 135,
+ 587
+ ],
+ "score": 0.85,
+ "content": "L _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 576,
+ 491,
+ 588
+ ],
+ "score": 1.0,
+ "content": "regularization coefficient for all weights except RNN weights: [0.1, 0.01, 0.001, 0.0001]",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 590,
+ 433,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 590,
+ 433,
+ 603
+ ],
+ "score": 1.0,
+ "content": "• dropout rate for the dropout on the RNN hidden layer: [0.0, 0.2, 0.4, 0.6, 0.8]",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 111,
+ 532,
+ 491,
+ 603
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 610,
+ 505,
+ 676
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "We performed 100 iterations of the random search by using the above ranges for each of the three",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 457,
+ 633
+ ],
+ "score": 1.0,
+ "content": "prediction experiments. For sequential diagnoses prediction on Sutter data, we used",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 622,
+ 477,
+ 632
+ ],
+ "score": 0.86,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 621,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 644
+ ],
+ "score": 1.0,
+ "content": "training data to tune the hyper-parameters to balance the time and search space. To match the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 383,
+ 655
+ ],
+ "score": 1.0,
+ "content": "baselines’ number of parameters to GRAM’s, we add 550 to the list of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 644,
+ 393,
+ 653
+ ],
+ "score": 0.68,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 642,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "’s possible values. This will",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "make the baseline’s largest possible number of parameters comparable to the GRAM’s largest possible",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 664,
+ 199,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 199,
+ 676
+ ],
+ "score": 1.0,
+ "content": "number of parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 610,
+ 505,
+ 676
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 681,
+ 504,
+ 714
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "score": 1.0,
+ "content": "For SimpleRollUp and RollUpRare, the number of input codes is smaller than other models due to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 693,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 693,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "the grouping. Therefore, to match their largest possible number of parameters to GRAM’s, we need to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 702,
+ 505,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 702,
+ 211,
+ 715
+ ],
+ "score": 1.0,
+ "content": "add much larger values to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 705,
+ 221,
+ 712
+ ],
+ "score": 0.67,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 221,
+ 702,
+ 468,
+ 715
+ ],
+ "score": 1.0,
+ "content": ". However, after preliminary experiments, as expected, setting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 704,
+ 478,
+ 712
+ ],
+ "score": 0.73,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 702,
+ 505,
+ 715
+ ],
+ "score": 1.0,
+ "content": "to too",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 331
+ ],
+ "score": 1.0,
+ "content": "large a value degraded the performance due to overfitting. Since the number of input codes decreased",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 323,
+ 341
+ ],
+ "score": 1.0,
+ "content": "due to the grouping, increasing the dimensionality of",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 323,
+ 330,
+ 333,
+ 340
+ ],
+ "score": 0.87,
+ "content": "\\mathbf { e } _ { i }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 334,
+ 329,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "is not a logical thing to do. Therefore, for",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 340,
+ 458,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 371,
+ 352
+ ],
+ "score": 1.0,
+ "content": "SimpleRollUp and RollUpRare, we use the same list of values for",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 371,
+ 342,
+ 382,
+ 349
+ ],
+ "score": 0.63,
+ "content": "m",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 382,
+ 340,
+ 458,
+ 352
+ ],
+ "score": 1.0,
+ "content": "as other baselines.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 681,
+ 506,
+ 715
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 132,
+ 76,
+ 476,
+ 309
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 128,
+ 62,
+ 481,
+ 74
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 60,
+ 483,
+ 76
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 60,
+ 483,
+ 76
+ ],
+ "score": 1.0,
+ "content": "Table 5: Hyper-parameters used by the models in each predictive modeling experiments",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 132,
+ 76,
+ 476,
+ 309
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 76,
+ 476,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 76,
+ 476,
+ 309
+ ],
+ "score": 0.984,
+ "html": "| ICD-9 Code | Description |
| 398.91 | Rheumatic heart failure (congestive) |
| 402.01 | Malignant hypertensive heart disease with heart failure |
| 402.11 | Benign hypertensive heart disease with heart failure |
| 402.91 | Unspecified hypertensive heart disease with heart failure |
| 404.01 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.03 | Hypertensive heart and chronic kidney disease, malignant, with heart failure and with chronic kidney disease stage V or end stage renal disease |
| 404.11 | Hypertensive heart and chronic kidney disease, benign, with heart failure and withchronic kidney disease stage I through stage IV, or unspecified |
| 404.13 | Hypertensive heart and chronic kidney disease, benign, with heart failure and chronickidney disease stage V or end stage renal disease |
| 404.91 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure and with chronic kidney disease stage I through stage IV, or unspecified |
| 404.93 | Hypertensive heart and chronic kidney disease, unspecified, with heart failure andchronic kidney disease stage V or end stage renal disease |
| 428.0 | Congestive heart failure, unspecified |
| 428.1 | Leftheart failure |
| 428.20 | Systolic heart failure, unspecified |
| 428.21 | Acute systolic heart failure |
| 428.22 | Chronic systolic heart failure |
| 428.23 | Acute on chronic systolic heart failure |
| 428.30 | Diastolic heart failure, unspecified |
| 428.31 | Acute diastolic heart failure |
| 428.32 | Chronic diastolic heart failure |
| 428.33 | Acute on chronic diastolic heart failure |
| 428.40 | Combined systolic and diastolic heart failure, unspecified |
| 428.41 | Acute combined systolic and diastolic heart failure |
| 428.42 | Chronic combined systolic and diastolic heart failure |
| 428.43 | Acute on chronic combined systolic and diastolic heart failure |
| 428.9 | Heart failure, unspecified |
"
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1188,
+ 1403,
+ 1188,
+ 1403,
+ 1309,
+ 299,
+ 1309
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1697,
+ 1404,
+ 1697,
+ 1404,
+ 1878,
+ 298,
+ 1878
+ ],
+ "score": 0.975
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 311,
+ 1477,
+ 1365,
+ 1477,
+ 1365,
+ 1674,
+ 311,
+ 1674
+ ],
+ "score": 0.97
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1893,
+ 1402,
+ 1893,
+ 1402,
+ 1985,
+ 300,
+ 1985
+ ],
+ "score": 0.969
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 301,
+ 1422,
+ 777,
+ 1422,
+ 777,
+ 1455,
+ 301,
+ 1455
+ ],
+ "score": 0.926
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 574,
+ 1097,
+ 1126,
+ 1097,
+ 1126,
+ 1132,
+ 574,
+ 1132
+ ],
+ "score": 0.905
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 297,
+ 75,
+ 857,
+ 75,
+ 857,
+ 105,
+ 297,
+ 105
+ ],
+ "score": 0.905
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 1353,
+ 766,
+ 1353,
+ 766,
+ 1389,
+ 302,
+ 1389
+ ],
+ "score": 0.9
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2038,
+ 864,
+ 2038,
+ 864,
+ 2063,
+ 836,
+ 2063
+ ],
+ "score": 0.854
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 558,
+ 1560,
+ 607,
+ 1560,
+ 607,
+ 1592,
+ 558,
+ 1592
+ ],
+ "score": 0.91,
+ "latex": "\\mathbf { W } _ { a }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 657,
+ 1561,
+ 692,
+ 1561,
+ 692,
+ 1592,
+ 657,
+ 1592
+ ],
+ "score": 0.89,
+ "latex": " { \\mathbf { b } } _ { a }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 812,
+ 1520,
+ 843,
+ 1520,
+ 843,
+ 1551,
+ 812,
+ 1551
+ ],
+ "score": 0.88,
+ "latex": "\\mathbf { h } _ { t }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1271,
+ 1728,
+ 1326,
+ 1728,
+ 1326,
+ 1757,
+ 1271,
+ 1757
+ ],
+ "score": 0.86,
+ "latex": "10 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 339,
+ 1600,
+ 375,
+ 1600,
+ 375,
+ 1631,
+ 339,
+ 1631
+ ],
+ "score": 0.85,
+ "latex": "L _ { 2 }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 512,
+ 1527,
+ 529,
+ 1527,
+ 529,
+ 1547,
+ 512,
+ 1547
+ ],
+ "score": 0.77,
+ "latex": "r"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 807,
+ 1483,
+ 834,
+ 1483,
+ 834,
+ 1511,
+ 807,
+ 1511
+ ],
+ "score": 0.75,
+ "latex": "\\mathbf { e } _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 512,
+ 1485,
+ 540,
+ 1485,
+ 540,
+ 1508,
+ 512,
+ 1508
+ ],
+ "score": 0.74,
+ "latex": "m"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1303,
+ 1958,
+ 1330,
+ 1958,
+ 1330,
+ 1980,
+ 1303,
+ 1980
+ ],
+ "score": 0.73,
+ "latex": "m"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1066,
+ 1791,
+ 1094,
+ 1791,
+ 1094,
+ 1815,
+ 1066,
+ 1815
+ ],
+ "score": 0.68,
+ "latex": "m"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 588,
+ 1959,
+ 615,
+ 1959,
+ 615,
+ 1980,
+ 588,
+ 1980
+ ],
+ "score": 0.67,
+ "latex": "m"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 512,
+ 1562,
+ 525,
+ 1562,
+ 525,
+ 1588,
+ 512,
+ 1588
+ ],
+ "score": 0.66,
+ "latex": "l"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 573.0,
+ 1097.0,
+ 1126.0,
+ 1097.0,
+ 1126.0,
+ 1133.0,
+ 573.0,
+ 1133.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1348.0,
+ 772.0,
+ 1348.0,
+ 772.0,
+ 1394.0,
+ 293.0,
+ 1394.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 831.0,
+ 2035.0,
+ 868.0,
+ 2035.0,
+ 868.0,
+ 2074.0,
+ 831.0,
+ 2074.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1190.0,
+ 1403.0,
+ 1190.0,
+ 1403.0,
+ 1222.0,
+ 295.0,
+ 1222.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1220.0,
+ 1404.0,
+ 1220.0,
+ 1404.0,
+ 1250.0,
+ 293.0,
+ 1250.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1246.0,
+ 1407.0,
+ 1246.0,
+ 1407.0,
+ 1285.0,
+ 293.0,
+ 1285.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1281.0,
+ 511.0,
+ 1281.0,
+ 511.0,
+ 1308.0,
+ 294.0,
+ 1308.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1696.0,
+ 1405.0,
+ 1696.0,
+ 1405.0,
+ 1731.0,
+ 296.0,
+ 1731.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1727.0,
+ 1270.0,
+ 1727.0,
+ 1270.0,
+ 1761.0,
+ 293.0,
+ 1761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1327.0,
+ 1727.0,
+ 1404.0,
+ 1727.0,
+ 1404.0,
+ 1761.0,
+ 1327.0,
+ 1761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1759.0,
+ 1402.0,
+ 1759.0,
+ 1402.0,
+ 1790.0,
+ 296.0,
+ 1790.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1786.0,
+ 1065.0,
+ 1786.0,
+ 1065.0,
+ 1821.0,
+ 294.0,
+ 1821.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1095.0,
+ 1786.0,
+ 1404.0,
+ 1786.0,
+ 1404.0,
+ 1821.0,
+ 1095.0,
+ 1821.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1816.0,
+ 1404.0,
+ 1816.0,
+ 1404.0,
+ 1851.0,
+ 294.0,
+ 1851.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1845.0,
+ 553.0,
+ 1845.0,
+ 553.0,
+ 1880.0,
+ 295.0,
+ 1880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 1478.0,
+ 511.0,
+ 1478.0,
+ 511.0,
+ 1514.0,
+ 311.0,
+ 1514.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 541.0,
+ 1478.0,
+ 806.0,
+ 1478.0,
+ 806.0,
+ 1514.0,
+ 541.0,
+ 1514.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 835.0,
+ 1478.0,
+ 1135.0,
+ 1478.0,
+ 1135.0,
+ 1514.0,
+ 835.0,
+ 1514.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 1517.0,
+ 511.0,
+ 1517.0,
+ 511.0,
+ 1553.0,
+ 311.0,
+ 1553.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 530.0,
+ 1517.0,
+ 811.0,
+ 1517.0,
+ 811.0,
+ 1553.0,
+ 530.0,
+ 1553.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1517.0,
+ 1290.0,
+ 1517.0,
+ 1290.0,
+ 1553.0,
+ 844.0,
+ 1553.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 1559.0,
+ 511.0,
+ 1559.0,
+ 511.0,
+ 1595.0,
+ 311.0,
+ 1595.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 526.0,
+ 1559.0,
+ 557.0,
+ 1559.0,
+ 557.0,
+ 1595.0,
+ 526.0,
+ 1595.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 608.0,
+ 1559.0,
+ 656.0,
+ 1559.0,
+ 656.0,
+ 1595.0,
+ 608.0,
+ 1595.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 693.0,
+ 1559.0,
+ 1140.0,
+ 1559.0,
+ 1140.0,
+ 1595.0,
+ 693.0,
+ 1595.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 313.0,
+ 1601.0,
+ 338.0,
+ 1601.0,
+ 338.0,
+ 1634.0,
+ 313.0,
+ 1634.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 376.0,
+ 1601.0,
+ 1365.0,
+ 1601.0,
+ 1365.0,
+ 1634.0,
+ 376.0,
+ 1634.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 1639.0,
+ 1205.0,
+ 1639.0,
+ 1205.0,
+ 1675.0,
+ 311.0,
+ 1675.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1892.0,
+ 1407.0,
+ 1892.0,
+ 1407.0,
+ 1929.0,
+ 294.0,
+ 1929.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1925.0,
+ 1406.0,
+ 1925.0,
+ 1406.0,
+ 1958.0,
+ 295.0,
+ 1958.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1952.0,
+ 587.0,
+ 1952.0,
+ 587.0,
+ 1988.0,
+ 294.0,
+ 1988.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 616.0,
+ 1952.0,
+ 1302.0,
+ 1952.0,
+ 1302.0,
+ 1988.0,
+ 616.0,
+ 1988.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1331.0,
+ 1952.0,
+ 1404.0,
+ 1952.0,
+ 1404.0,
+ 1988.0,
+ 1331.0,
+ 1988.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1422.0,
+ 780.0,
+ 1422.0,
+ 780.0,
+ 1458.0,
+ 297.0,
+ 1458.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 11,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 5,
+ "poly": [
+ 367,
+ 212,
+ 1323,
+ 212,
+ 1323,
+ 860,
+ 367,
+ 860
+ ],
+ "score": 0.984,
+ "html": "| Algorithm 1: Self-supervised outlier detection framework (SSD) | |
| Output :Is outlier or not? ∀x E Xtest | Input :Xin, Xtest, feature extractor (f), projection head (h),Required True-positive rate (T), |
| Function getFeatures(X): return {f(xi)/llf(xi)ll2,∀ xi ∈ X}; | |
| Function SSDkScore(Z, μin,Σin, μood,Σood): | Function SSDScore(Z,μ,∑): return {(z - μ)T∑-1(z - μ),∀ z ∈ Z}; |
| end | return{(z-μin)T∑-1(z-μin)-(z-μood)T∑-ld(z-μood),∀z∈Z}; |
| Parition Xin in training set (Xtrain) and calibration set (Xcal); | |
| if Vin isnot available then | |
| Lbatch = N∑²N1 uj/T -log eu | h(f(xi)) |
| N 1（k≠i）eu uk/T | ；ui= |
| else | |
| 1 | |
| 2Nyi -log | 11(k≠i)1(yk=yi）eTuk/T |
| 1（k≠i)e²Tuk/T |
| end | |
| |
| Train feature extractor(f) by minimizing Lbatch over Xtrain; |
| Ztrain = getFeatures(Xtrain),Zcal = getFeatures(Xcal) |
| Ztest = getFeatures(Xtest),if Xood is available: Zood = getFeatures(Xood); |
| if Xood is not available then | |
| |
| Scal = SSDScore(Zcal, μtrain,∑train); # Sample mean and convariance of Ztrain |
| Stest = SSDScore(Ztest, μtrain,∑train) # outlier score; |
| else | |
| # Using convariance estimation techniques from Section 3.2 for Zood |
| Scal = SSDkScore(Zcal, μtrain,∑train, μood,∑ood); | |
| |
| Stest = SSDkScore(Ztest, μtrain,∑train,μood,∑ood); |
| end | |
| |
| |
| |
| |
| Xi E Xtest is an outlier if stest > (Scal threshold at TPR = T); | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
+
+# 4 EXPERIMENTAL RESULTS
+
+# 4.1 COMMON SETUP ACROSS ALL EXPERIMENTS
+
+We use recently proposed NT-Xent loss function from SimCLR (Chen et al., 2020) method for self-supervised training. We use the ResNet-50 network in all key experiments but also provide ablation with ResNet-18, ResNet-34, and ResNet-101 network architecture. We train each network, for both supervised and self-supervised training, with stochastic gradient descent for 500 epochs, 0.5 starting learning rate with cosine decay, and weight decay and batch size set to 1e-4 and 512, respectively. We set the temperature parameter to 0.5 in the NT-Xent loss. We evaluate our detector with three performance metrics, namely FPR (at $\mathrm { T P R } { = } 9 5 \%$ ), AUROC, and AUPR. For the supervised training baseline, we use identical training budget as SSD while also using the Mahalanobis distance based detection in the feature space. Due to space constraints, we present results with AUROC, which refers to area under the receiver operating characteristic curve, in the main paper and provide detailed results with other performance metrics in Appendix B.5. Our setup incorporates six image datasets along with additional synthetic datasets based on random noise. We report the average results over three independent runs in most experiments.
+
+Number of clusters. We find the choice of the number of clusters dependent on which layer we extract the features from in the Residual neural networks. While for the first three blocks, we find an increase in AUROC with the number of clusters, the trend is reversed for the last block (Appendix B.2). Since the last block features achieve the highest detection performance, we model the in-distribution features as a single cluster in subsequent experiments.
+
+# 4.2 PERFORMANCE OF SSD
+
+Comparison with unsupervised learning based detectors. We present this comparison in Table 1. We find that SSD improves average AUROC by up to 55, compared to standard outlier detectors based on Density modeling $\mathrm { ( P i x e l C N N + + }$ (Salimans et al., 2017)), input reconstruction (Autoencoder (Hawkins et al., 2002), Variational Auto-encoder (Kingma & Welling, 2014)), and One-class classification (Deep-SVDD Ruff et al. (2018)). A common limitation of each of these three detectors is to find images from the SVHN dataset as more in-distribution when trained on CIFAR-10 or CIFAR-100 dataset. In contrast, SSD is able to successfully detect a large fraction of outliers from SVHN dataset. We also experiment with Rotation-loss Gidaris et al. (2018), a non-contrastive self-supervised training objective. We find that SSD with contrastive NT-Xent loss achieves $9 . 6 \%$ higher average AUROC compared to using Rotation-loss.
+
+Table 1: Comparison of SSD with different outlier detectors using only unlabeled training data.
+
+ | Airplane | Automobile | Bird | Cat | Deer | Dog | Frog | Horse | Ship | Truck | Average |
| Randomly Initialized network | 77.4 | 44.1 | 62.4 | 44.1 | 62.1 | 49.6 | 59.8 | 48.0 | 73.8 | 53.7 | 57.5 |
| VAE(Kingma & Welling,2014) | 70.0 | 38.6 | 67.9 | 53.5 | 74.8 | 52.3 | 68.7 | 49.3 | 69.6 | 38.6 | 58.3 |
| OCSVM (Scholkopf et al., 2001) | 63.0 | 44.0 | 64.9 | 48.7 | 73.5 | 50.0 | 72.5 | 53.3 | 64.9 | 50.8 | 58.5 |
| AnoGAN (Schlegl et al., 2017) | 67.1 | 54.7 | 52.9 | 54.5 | 65.1 | 60.3 | 58.5 | 62.5 | 75.8 | 66.5 | 61.8 |
| PixelCNN(Van den Oord et al.,2016) | 53.1 | 99.5 | 47.6 | 51.7 | 73.9 | 54.2 | 59.2 | 78.9 | 34.0 | 66.2 | 61.8 |
| DSVDD (Ruff et al., 2018) | 61.7 | 65.9 | 50.8 | 59.1 | 60.9 | 65.7 | 67.7 | 67.3 | 75.9 | 73.1 | 64.8 |
| OCGAN (Perera et al.,2019) | 75.7 | 53.1 | 64.0 | 62.0 | 72.3 | 62.0 | 72.3 | 57.5 | 82.0 | 55.4 | 65.6 |
| RCAE(Chalapathy et al., 2018b) | 72.0 | 63.1 | 71.7 | 60.6 | 72.8 | 64.0 | 64.9 | 63.6 | 74.7 | 74.5 | 68.2 |
| DROCC (Goyal et al.,2020) | 81.7 | 76.7 | 66.7 | 67.1 | 73.6 | 74.4 | 74.4 | 71.4 | 80.0 | 76.2 | 74.2 |
| Deep-SAD (Ruff et al., 2019) | 二 | 二 | 二 | | | 一 | 一 | | | 一 | 77.9 |
| E3Outlier (Wang et al.,2019) | 79.4 | 95.3 | 75.4 | 73.9 | 84.1 | 87.9 | 85.0 | 93.4 | 92.3 | 89.7 | 85.6 |
| GT(Golan & El-Yaniv,2018) | 74.7 | 95.7 | 78.1 | 72.4 | 87.8 | 87.8 | 83.4 | 95.5 | 93.3 | 91.3 | 86.0 |
| InvAE (Huang et al., 2019) | 78.5 | 89.8 | 86.1 | 77.4 | 90.5 | 84.5 | 89.2 | 92.9 | 92.0 | 85.5 | 86.6 |
| GOAD (Bergman & Hoshen,2020) | 77.2 | 96.7 | 83.3 | 77.7 | 87.8 | 87.8 | 90.0 | 96.1 | 93.8 | 92.0 | 88.2 |
| CSI (Tack et al., 2020) | 89.9 | 99.9 | 93.1 | 86.4 | 93.9 | 93.2 | 95.1 | 98.7 | 97.9 | 95.5 | 94.3 |
| SSD | 82.7 | 98.5 | 84.2 | 84.5 | 84.8 | 90.9 | 91.7 | 95.2 | 92.9 | 94.4 | 90.0 |
+
+Results. Compared to the baseline SSD detector, one-shot and five-shot settings improve the average AUROC, across all OOD datasets, by 1.6 and 2.1, respectively (Table 1, 2). In particular, we observe large gains with CIFAR-100 as in-distribution and CIFAR-10 as OOD where five-shot detection improves the AUROC from 69.6 to 78.3. We find the use of shrunk covariance estimator most critical in the success of our approach. Use of shrunk covariance estimation itself improves the AUROC from 69.6 to 77.1. Then data augmentation further improves it 78.3 for the five-shot detection. With an increasing number of transformed copies of each sample, we also observe improvement in AUROC, though it later plateaus close to ten copies (Appendix B.3).
+
+What if additional OOD images are available Note that some earlier works, such as Liang et al. (2018), assume that 1000 OOD inputs are available for tuning the detector. We find that with access to this large amount of OOD samples, $S S D _ { k }$ can improve the state-of-the-art by an even larger margin. For example, with CIFAR-100 as in-distribution and CIFAR-10 as out-of-distribution, it achieves 89.4 AUROC, which is $1 4 . 2 \%$ higher than the current state-of-the-art (Winkens et al., 2020).
+
+# 4.4 SUCCESS WHEN USING DATA LABELS $( S S D + )$
+
+Now we integrate labels of training data in our framework and compare it with the existing state-ofthe-art detectors. We report our results in Table 4. Our approach improves the average AUROC by 0.8 over the previous state-of-the-art detector. Our approach also achieves equal or better performance than previous state-of-the-art across individual pairs of in and out-distribution dataset. For example, using labels in our framework improves the AUROC of Mahalanobis detector from 55.5 to 72.1 for CIFAR-100 as in-distribution and CIFAR-10 as the OOD dataset. Using the simple softmax probabilities, training a two-layer MLP on learned representations further improves the AUROC to 78.3. Combining $S S D +$ with a five-shot OOD detection method further brings a gain of 1.4 in the average AUROC.
+
+# 5 DISCUSSION AND CONCLUSION
+
+On tuning hyperparameters in SSD. In our framework, we either explicitly avoid the use of additional tuning-parameters (such as when combining self-supervised and supervised loss functions in $S S D +$ ) or refrain from tuning the existing set of parameters for each OOD dataset. For example, we use a standard set of parameters for self-supervised training and model the learned features with a single-cluster.
+
+Table 4: Comparison of $S S D +$ , i.e., incorporating labels in the SSD detector, with state-of-the-art detectors based on supervised training.
+
+| Algorithm 1: Self-supervised outlier detection framework (SSD) | |
| Output :Is outlier or not? ∀x E Xtest | Input :Xin, Xtest, feature extractor (f), projection head (h),Required True-positive rate (T), |
| Function getFeatures(X): return {f(xi)/llf(xi)ll2,∀ xi ∈ X}; | |
| Function SSDkScore(Z, μin,Σin, μood,Σood): | Function SSDScore(Z,μ,∑): return {(z - μ)T∑-1(z - μ),∀ z ∈ Z}; |
| end | return{(z-μin)T∑-1(z-μin)-(z-μood)T∑-ld(z-μood),∀z∈Z}; |
| Parition Xin in training set (Xtrain) and calibration set (Xcal); | |
| if Vin isnot available then | |
| Lbatch = N∑²N1 uj/T -log eu | h(f(xi)) |
| N 1（k≠i）eu uk/T | ；ui= |
| else | |
| 1 | |
| 2Nyi -log | 11(k≠i)1(yk=yi）eTuk/T |
| 1（k≠i)e²Tuk/T |
| end | |
| |
| Train feature extractor(f) by minimizing Lbatch over Xtrain; |
| Ztrain = getFeatures(Xtrain),Zcal = getFeatures(Xcal) |
| Ztest = getFeatures(Xtest),if Xood is available: Zood = getFeatures(Xood); |
| if Xood is not available then | |
| |
| Scal = SSDScore(Zcal, μtrain,∑train); # Sample mean and convariance of Ztrain |
| Stest = SSDScore(Ztest, μtrain,∑train) # outlier score; |
| else | |
| # Using convariance estimation techniques from Section 3.2 for Zood |
| Scal = SSDkScore(Zcal, μtrain,∑train, μood,∑ood); | |
| |
| Stest = SSDkScore(Ztest, μtrain,∑train,μood,∑ood); |
| end | |
| |
| |
| |
| |
| Xi E Xtest is an outlier if stest > (Scal threshold at TPR = T); | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
",
+ "bbox": [
+ 171,
+ 107,
+ 807,
+ 527
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4 EXPERIMENTAL RESULTS ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 542,
+ 416,
+ 558
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.1 COMMON SETUP ACROSS ALL EXPERIMENTS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 568,
+ 524,
+ 582
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We use recently proposed NT-Xent loss function from SimCLR (Chen et al., 2020) method for self-supervised training. We use the ResNet-50 network in all key experiments but also provide ablation with ResNet-18, ResNet-34, and ResNet-101 network architecture. We train each network, for both supervised and self-supervised training, with stochastic gradient descent for 500 epochs, 0.5 starting learning rate with cosine decay, and weight decay and batch size set to 1e-4 and 512, respectively. We set the temperature parameter to 0.5 in the NT-Xent loss. We evaluate our detector with three performance metrics, namely FPR (at $\\mathrm { T P R } { = } 9 5 \\%$ ), AUROC, and AUPR. For the supervised training baseline, we use identical training budget as SSD while also using the Mahalanobis distance based detection in the feature space. Due to space constraints, we present results with AUROC, which refers to area under the receiver operating characteristic curve, in the main paper and provide detailed results with other performance metrics in Appendix B.5. Our setup incorporates six image datasets along with additional synthetic datasets based on random noise. We report the average results over three independent runs in most experiments. ",
+ "bbox": [
+ 173,
+ 593,
+ 826,
+ 773
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Number of clusters. We find the choice of the number of clusters dependent on which layer we extract the features from in the Residual neural networks. While for the first three blocks, we find an increase in AUROC with the number of clusters, the trend is reversed for the last block (Appendix B.2). Since the last block features achieve the highest detection performance, we model the in-distribution features as a single cluster in subsequent experiments. ",
+ "bbox": [
+ 174,
+ 781,
+ 823,
+ 852
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.2 PERFORMANCE OF SSD ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 868,
+ 380,
+ 882
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Comparison with unsupervised learning based detectors. We present this comparison in Table 1. We find that SSD improves average AUROC by up to 55, compared to standard outlier detectors based on Density modeling $\\mathrm { ( P i x e l C N N + + }$ (Salimans et al., 2017)), input reconstruction (Autoencoder (Hawkins et al., 2002), Variational Auto-encoder (Kingma & Welling, 2014)), and One-class classification (Deep-SVDD Ruff et al. (2018)). A common limitation of each of these three detectors is to find images from the SVHN dataset as more in-distribution when trained on CIFAR-10 or CIFAR-100 dataset. In contrast, SSD is able to successfully detect a large fraction of outliers from SVHN dataset. We also experiment with Rotation-loss Gidaris et al. (2018), a non-contrastive self-supervised training objective. We find that SSD with contrastive NT-Xent loss achieves $9 . 6 \\%$ higher average AUROC compared to using Rotation-loss. ",
+ "bbox": [
+ 174,
+ 895,
+ 823,
+ 924
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/60ff791a73e029afb920442a77d07c7a80e68c334b9cd71275a95852c952af5d.jpg",
+ "table_caption": [
+ "Table 1: Comparison of SSD with different outlier detectors using only unlabeled training data. "
+ ],
+ "table_footnote": [],
+ "table_body": " | Airplane | Automobile | Bird | Cat | Deer | Dog | Frog | Horse | Ship | Truck | Average |
| Randomly Initialized network | 77.4 | 44.1 | 62.4 | 44.1 | 62.1 | 49.6 | 59.8 | 48.0 | 73.8 | 53.7 | 57.5 |
| VAE(Kingma & Welling,2014) | 70.0 | 38.6 | 67.9 | 53.5 | 74.8 | 52.3 | 68.7 | 49.3 | 69.6 | 38.6 | 58.3 |
| OCSVM (Scholkopf et al., 2001) | 63.0 | 44.0 | 64.9 | 48.7 | 73.5 | 50.0 | 72.5 | 53.3 | 64.9 | 50.8 | 58.5 |
| AnoGAN (Schlegl et al., 2017) | 67.1 | 54.7 | 52.9 | 54.5 | 65.1 | 60.3 | 58.5 | 62.5 | 75.8 | 66.5 | 61.8 |
| PixelCNN(Van den Oord et al.,2016) | 53.1 | 99.5 | 47.6 | 51.7 | 73.9 | 54.2 | 59.2 | 78.9 | 34.0 | 66.2 | 61.8 |
| DSVDD (Ruff et al., 2018) | 61.7 | 65.9 | 50.8 | 59.1 | 60.9 | 65.7 | 67.7 | 67.3 | 75.9 | 73.1 | 64.8 |
| OCGAN (Perera et al.,2019) | 75.7 | 53.1 | 64.0 | 62.0 | 72.3 | 62.0 | 72.3 | 57.5 | 82.0 | 55.4 | 65.6 |
| RCAE(Chalapathy et al., 2018b) | 72.0 | 63.1 | 71.7 | 60.6 | 72.8 | 64.0 | 64.9 | 63.6 | 74.7 | 74.5 | 68.2 |
| DROCC (Goyal et al.,2020) | 81.7 | 76.7 | 66.7 | 67.1 | 73.6 | 74.4 | 74.4 | 71.4 | 80.0 | 76.2 | 74.2 |
| Deep-SAD (Ruff et al., 2019) | 二 | 二 | 二 | | | 一 | 一 | | | 一 | 77.9 |
| E3Outlier (Wang et al.,2019) | 79.4 | 95.3 | 75.4 | 73.9 | 84.1 | 87.9 | 85.0 | 93.4 | 92.3 | 89.7 | 85.6 |
| GT(Golan & El-Yaniv,2018) | 74.7 | 95.7 | 78.1 | 72.4 | 87.8 | 87.8 | 83.4 | 95.5 | 93.3 | 91.3 | 86.0 |
| InvAE (Huang et al., 2019) | 78.5 | 89.8 | 86.1 | 77.4 | 90.5 | 84.5 | 89.2 | 92.9 | 92.0 | 85.5 | 86.6 |
| GOAD (Bergman & Hoshen,2020) | 77.2 | 96.7 | 83.3 | 77.7 | 87.8 | 87.8 | 90.0 | 96.1 | 93.8 | 92.0 | 88.2 |
| CSI (Tack et al., 2020) | 89.9 | 99.9 | 93.1 | 86.4 | 93.9 | 93.2 | 95.1 | 98.7 | 97.9 | 95.5 | 94.3 |
| SSD | 82.7 | 98.5 | 84.2 | 84.5 | 84.8 | 90.9 | 91.7 | 95.2 | 92.9 | 94.4 | 90.0 |
",
+ "bbox": [
+ 207,
+ 303,
+ 790,
+ 472
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Results. Compared to the baseline SSD detector, one-shot and five-shot settings improve the average AUROC, across all OOD datasets, by 1.6 and 2.1, respectively (Table 1, 2). In particular, we observe large gains with CIFAR-100 as in-distribution and CIFAR-10 as OOD where five-shot detection improves the AUROC from 69.6 to 78.3. We find the use of shrunk covariance estimator most critical in the success of our approach. Use of shrunk covariance estimation itself improves the AUROC from 69.6 to 77.1. Then data augmentation further improves it 78.3 for the five-shot detection. With an increasing number of transformed copies of each sample, we also observe improvement in AUROC, though it later plateaus close to ten copies (Appendix B.3). ",
+ "bbox": [
+ 173,
+ 486,
+ 825,
+ 598
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "What if additional OOD images are available Note that some earlier works, such as Liang et al. (2018), assume that 1000 OOD inputs are available for tuning the detector. We find that with access to this large amount of OOD samples, $S S D _ { k }$ can improve the state-of-the-art by an even larger margin. For example, with CIFAR-100 as in-distribution and CIFAR-10 as out-of-distribution, it achieves 89.4 AUROC, which is $1 4 . 2 \\%$ higher than the current state-of-the-art (Winkens et al., 2020). ",
+ "bbox": [
+ 174,
+ 604,
+ 825,
+ 674
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "4.4 SUCCESS WHEN USING DATA LABELS $( S S D + )$ ",
+ "text_level": 1,
+ "bbox": [
+ 178,
+ 691,
+ 527,
+ 705
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Now we integrate labels of training data in our framework and compare it with the existing state-ofthe-art detectors. We report our results in Table 4. Our approach improves the average AUROC by 0.8 over the previous state-of-the-art detector. Our approach also achieves equal or better performance than previous state-of-the-art across individual pairs of in and out-distribution dataset. For example, using labels in our framework improves the AUROC of Mahalanobis detector from 55.5 to 72.1 for CIFAR-100 as in-distribution and CIFAR-10 as the OOD dataset. Using the simple softmax probabilities, training a two-layer MLP on learned representations further improves the AUROC to 78.3. Combining $S S D +$ with a five-shot OOD detection method further brings a gain of 1.4 in the average AUROC. ",
+ "bbox": [
+ 173,
+ 717,
+ 826,
+ 843
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5 DISCUSSION AND CONCLUSION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 863,
+ 467,
+ 880
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "On tuning hyperparameters in SSD. In our framework, we either explicitly avoid the use of additional tuning-parameters (such as when combining self-supervised and supervised loss functions in $S S D +$ ) or refrain from tuning the existing set of parameters for each OOD dataset. For example, we use a standard set of parameters for self-supervised training and model the learned features with a single-cluster. ",
+ "bbox": [
+ 173,
+ 895,
+ 823,
+ 924
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "table",
+ "img_path": "images/0f2c3d1659e2ecd80afa79297ac05e0ba820d3617688f579475376a9386aeaf4.jpg",
+ "table_caption": [
+ "Table 4: Comparison of $S S D +$ , i.e., incorporating labels in the SSD detector, with state-of-the-art detectors based on supervised training. "
+ ],
+ "table_footnote": [
+ "∗ Uses $4 \\times$ wider ResNet-50 network, † Requires additional out-of-distribution data for tuning. "
+ ],
+ "table_body": "| Algorithm 1: Self-supervised outlier detection framework (SSD) | |
| Output :Is outlier or not? ∀x E Xtest | Input :Xin, Xtest, feature extractor (f), projection head (h),Required True-positive rate (T), |
| Function getFeatures(X): return {f(xi)/llf(xi)ll2,∀ xi ∈ X}; | |
| Function SSDkScore(Z, μin,Σin, μood,Σood): | Function SSDScore(Z,μ,∑): return {(z - μ)T∑-1(z - μ),∀ z ∈ Z}; |
| end | return{(z-μin)T∑-1(z-μin)-(z-μood)T∑-ld(z-μood),∀z∈Z}; |
| Parition Xin in training set (Xtrain) and calibration set (Xcal); | |
| if Vin isnot available then | |
| Lbatch = N∑²N1 uj/T -log eu | h(f(xi)) |
| N 1（k≠i）eu uk/T | ；ui= |
| else | |
| 1 | |
| 2Nyi -log | 11(k≠i)1(yk=yi）eTuk/T |
| 1（k≠i)e²Tuk/T |
| end | |
| |
| Train feature extractor(f) by minimizing Lbatch over Xtrain; |
| Ztrain = getFeatures(Xtrain),Zcal = getFeatures(Xcal) |
| Ztest = getFeatures(Xtest),if Xood is available: Zood = getFeatures(Xood); |
| if Xood is not available then | |
| |
| Scal = SSDScore(Zcal, μtrain,∑train); # Sample mean and convariance of Ztrain |
| Stest = SSDScore(Ztest, μtrain,∑train) # outlier score; |
| else | |
| # Using convariance estimation techniques from Section 3.2 for Zood |
| Scal = SSDkScore(Zcal, μtrain,∑train, μood,∑ood); | |
| |
| Stest = SSDkScore(Ztest, μtrain,∑train,μood,∑ood); |
| end | |
| |
| |
| |
| |
| Xi E Xtest is an outlier if stest > (Scal threshold at TPR = T); | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
",
+ "type": "table",
+ "image_path": "9b3bf7a857cd6e3d37a67efbba4ac26978860f0be5a6b5cb6f1f6c5dd8d7ee40.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 85,
+ 494,
+ 196.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 196.0,
+ 494,
+ 307.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 307.0,
+ 494,
+ 418.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 430,
+ 255,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 256,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 256,
+ 444
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTAL RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 450,
+ 321,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 321,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 321,
+ 462
+ ],
+ "score": 1.0,
+ "content": "4.1 COMMON SETUP ACROSS ALL EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "We use recently proposed NT-Xent loss function from SimCLR (Chen et al., 2020) method for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "self-supervised training. We use the ResNet-50 network in all key experiments but also provide",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "ablation with ResNet-18, ResNet-34, and ResNet-101 network architecture. We train each network,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "for both supervised and self-supervised training, with stochastic gradient descent for 500 epochs,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "0.5 starting learning rate with cosine decay, and weight decay and batch size set to 1e-4 and 512,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "score": 1.0,
+ "content": "respectively. We set the temperature parameter to 0.5 in the NT-Xent loss. We evaluate our detector",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 506,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 296,
+ 550
+ ],
+ "score": 1.0,
+ "content": "with three performance metrics, namely FPR (at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 537,
+ 339,
+ 547
+ ],
+ "score": 0.83,
+ "content": "\\mathrm { T P R } { = } 9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 537,
+ 506,
+ 550
+ ],
+ "score": 1.0,
+ "content": "), AUROC, and AUPR. For the supervised",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "training baseline, we use identical training budget as SSD while also using the Mahalanobis distance",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "based detection in the feature space. Due to space constraints, we present results with AUROC, which",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 582
+ ],
+ "score": 1.0,
+ "content": "refers to area under the receiver operating characteristic curve, in the main paper and provide detailed",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "results with other performance metrics in Appendix B.5. Our setup incorporates six image datasets",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "score": 1.0,
+ "content": "along with additional synthetic datasets based on random noise. We report the average results over",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 603,
+ 285,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 285,
+ 615
+ ],
+ "score": 1.0,
+ "content": "three independent runs in most experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 619,
+ 504,
+ 675
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "Number of clusters. We find the choice of the number of clusters dependent on which layer we",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "extract the features from in the Residual neural networks. While for the first three blocks, we",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "find an increase in AUROC with the number of clusters, the trend is reversed for the last block",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 664
+ ],
+ "score": 1.0,
+ "content": "(Appendix B.2). Since the last block features achieve the highest detection performance, we model",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 398,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 398,
+ 676
+ ],
+ "score": 1.0,
+ "content": "the in-distribution features as a single cluster in subsequent experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 688,
+ 233,
+ 699
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 234,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 234,
+ 701
+ ],
+ "score": 1.0,
+ "content": "4.2 PERFORMANCE OF SSD",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Comparison with unsupervised learning based detectors. We present this comparison in Table 1.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "We find that SSD improves average AUROC by up to 55, compared to standard outlier detectors",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 25,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 25,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 105,
+ 85,
+ 494,
+ 418
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 105,
+ 85,
+ 494,
+ 418
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 85,
+ 494,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 85,
+ 494,
+ 418
+ ],
+ "score": 0.873,
+ "html": "| Algorithm 1: Self-supervised outlier detection framework (SSD) | |
| Output :Is outlier or not? ∀x E Xtest | Input :Xin, Xtest, feature extractor (f), projection head (h),Required True-positive rate (T), |
| Function getFeatures(X): return {f(xi)/llf(xi)ll2,∀ xi ∈ X}; | |
| Function SSDkScore(Z, μin,Σin, μood,Σood): | Function SSDScore(Z,μ,∑): return {(z - μ)T∑-1(z - μ),∀ z ∈ Z}; |
| end | return{(z-μin)T∑-1(z-μin)-(z-μood)T∑-ld(z-μood),∀z∈Z}; |
| Parition Xin in training set (Xtrain) and calibration set (Xcal); | |
| if Vin isnot available then | |
| Lbatch = N∑²N1 uj/T -log eu | h(f(xi)) |
| N 1（k≠i）eu uk/T | ；ui= |
| else | |
| 1 | |
| 2Nyi -log | 11(k≠i)1(yk=yi）eTuk/T |
| 1（k≠i)e²Tuk/T |
| end | |
| |
| Train feature extractor(f) by minimizing Lbatch over Xtrain; |
| Ztrain = getFeatures(Xtrain),Zcal = getFeatures(Xcal) |
| Ztest = getFeatures(Xtest),if Xood is available: Zood = getFeatures(Xood); |
| if Xood is not available then | |
| |
| Scal = SSDScore(Zcal, μtrain,∑train); # Sample mean and convariance of Ztrain |
| Stest = SSDScore(Ztest, μtrain,∑train) # outlier score; |
| else | |
| # Using convariance estimation techniques from Section 3.2 for Zood |
| Scal = SSDkScore(Zcal, μtrain,∑train, μood,∑ood); | |
| |
| Stest = SSDkScore(Ztest, μtrain,∑train,μood,∑ood); |
| end | |
| |
| |
| |
| |
| Xi E Xtest is an outlier if stest > (Scal threshold at TPR = T); | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
",
+ "type": "table",
+ "image_path": "9b3bf7a857cd6e3d37a67efbba4ac26978860f0be5a6b5cb6f1f6c5dd8d7ee40.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 85,
+ 494,
+ 196.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 196.0,
+ 494,
+ 307.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 307.0,
+ 494,
+ 418.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 430,
+ 255,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 256,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 256,
+ 444
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTAL RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 450,
+ 321,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 321,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 321,
+ 462
+ ],
+ "score": 1.0,
+ "content": "4.1 COMMON SETUP ACROSS ALL EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "We use recently proposed NT-Xent loss function from SimCLR (Chen et al., 2020) method for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "self-supervised training. We use the ResNet-50 network in all key experiments but also provide",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "ablation with ResNet-18, ResNet-34, and ResNet-101 network architecture. We train each network,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "for both supervised and self-supervised training, with stochastic gradient descent for 500 epochs,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "0.5 starting learning rate with cosine decay, and weight decay and batch size set to 1e-4 and 512,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "score": 1.0,
+ "content": "respectively. We set the temperature parameter to 0.5 in the NT-Xent loss. We evaluate our detector",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 506,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 296,
+ 550
+ ],
+ "score": 1.0,
+ "content": "with three performance metrics, namely FPR (at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 537,
+ 339,
+ 547
+ ],
+ "score": 0.83,
+ "content": "\\mathrm { T P R } { = } 9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 537,
+ 506,
+ 550
+ ],
+ "score": 1.0,
+ "content": "), AUROC, and AUPR. For the supervised",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "training baseline, we use identical training budget as SSD while also using the Mahalanobis distance",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "based detection in the feature space. Due to space constraints, we present results with AUROC, which",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 582
+ ],
+ "score": 1.0,
+ "content": "refers to area under the receiver operating characteristic curve, in the main paper and provide detailed",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "results with other performance metrics in Appendix B.5. Our setup incorporates six image datasets",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "score": 1.0,
+ "content": "along with additional synthetic datasets based on random noise. We report the average results over",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 603,
+ 285,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 285,
+ 615
+ ],
+ "score": 1.0,
+ "content": "three independent runs in most experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 470,
+ 506,
+ 615
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 619,
+ 504,
+ 675
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "Number of clusters. We find the choice of the number of clusters dependent on which layer we",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "extract the features from in the Residual neural networks. While for the first three blocks, we",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "find an increase in AUROC with the number of clusters, the trend is reversed for the last block",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 664
+ ],
+ "score": 1.0,
+ "content": "(Appendix B.2). Since the last block features achieve the highest detection performance, we model",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 398,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 398,
+ 676
+ ],
+ "score": 1.0,
+ "content": "the in-distribution features as a single cluster in subsequent experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 105,
+ 619,
+ 505,
+ 676
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 688,
+ 233,
+ 699
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 234,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 234,
+ 701
+ ],
+ "score": 1.0,
+ "content": "4.2 PERFORMANCE OF SSD",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Comparison with unsupervised learning based detectors. We present this comparison in Table 1.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "We find that SSD improves average AUROC by up to 55, compared to standard outlier detectors",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 203,
+ 506,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 203,
+ 226,
+ 217
+ ],
+ "score": 1.0,
+ "content": "based on Density modeling",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 227,
+ 204,
+ 284,
+ 214
+ ],
+ "score": 0.64,
+ "content": "\\mathrm { ( P i x e l C N N + + }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 284,
+ 203,
+ 506,
+ 217
+ ],
+ "score": 1.0,
+ "content": "(Salimans et al., 2017)), input reconstruction (Auto-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 215,
+ 505,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 215,
+ 505,
+ 227
+ ],
+ "score": 1.0,
+ "content": "encoder (Hawkins et al., 2002), Variational Auto-encoder (Kingma & Welling, 2014)), and One-class",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 238
+ ],
+ "score": 1.0,
+ "content": "classification (Deep-SVDD Ruff et al. (2018)). A common limitation of each of these three detectors",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "is to find images from the SVHN dataset as more in-distribution when trained on CIFAR-10 or",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 248,
+ 504,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 504,
+ 259
+ ],
+ "score": 1.0,
+ "content": "CIFAR-100 dataset. In contrast, SSD is able to successfully detect a large fraction of outliers from",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "SVHN dataset. We also experiment with Rotation-loss Gidaris et al. (2018), a non-contrastive",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 482,
+ 281
+ ],
+ "score": 1.0,
+ "content": "self-supervised training objective. We find that SSD with contrastive NT-Xent loss achieves",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 482,
+ 270,
+ 505,
+ 280
+ ],
+ "score": 0.83,
+ "content": "9 . 6 \\%",
+ "type": "inline_equation",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 281,
+ 338,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 338,
+ 293
+ ],
+ "score": 1.0,
+ "content": "higher average AUROC compared to using Rotation-loss.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 106,
+ 709,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 118,
+ 91,
+ 495,
+ 192
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 112,
+ 75,
+ 494,
+ 87
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 73,
+ 496,
+ 89
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 73,
+ 496,
+ 89
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparison of SSD with different outlier detectors using only unlabeled training data.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 118,
+ 91,
+ 495,
+ 192
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 91,
+ 495,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 91,
+ 495,
+ 192
+ ],
+ "score": 0.982,
+ "html": " | Airplane | Automobile | Bird | Cat | Deer | Dog | Frog | Horse | Ship | Truck | Average |
| Randomly Initialized network | 77.4 | 44.1 | 62.4 | 44.1 | 62.1 | 49.6 | 59.8 | 48.0 | 73.8 | 53.7 | 57.5 |
| VAE(Kingma & Welling,2014) | 70.0 | 38.6 | 67.9 | 53.5 | 74.8 | 52.3 | 68.7 | 49.3 | 69.6 | 38.6 | 58.3 |
| OCSVM (Scholkopf et al., 2001) | 63.0 | 44.0 | 64.9 | 48.7 | 73.5 | 50.0 | 72.5 | 53.3 | 64.9 | 50.8 | 58.5 |
| AnoGAN (Schlegl et al., 2017) | 67.1 | 54.7 | 52.9 | 54.5 | 65.1 | 60.3 | 58.5 | 62.5 | 75.8 | 66.5 | 61.8 |
| PixelCNN(Van den Oord et al.,2016) | 53.1 | 99.5 | 47.6 | 51.7 | 73.9 | 54.2 | 59.2 | 78.9 | 34.0 | 66.2 | 61.8 |
| DSVDD (Ruff et al., 2018) | 61.7 | 65.9 | 50.8 | 59.1 | 60.9 | 65.7 | 67.7 | 67.3 | 75.9 | 73.1 | 64.8 |
| OCGAN (Perera et al.,2019) | 75.7 | 53.1 | 64.0 | 62.0 | 72.3 | 62.0 | 72.3 | 57.5 | 82.0 | 55.4 | 65.6 |
| RCAE(Chalapathy et al., 2018b) | 72.0 | 63.1 | 71.7 | 60.6 | 72.8 | 64.0 | 64.9 | 63.6 | 74.7 | 74.5 | 68.2 |
| DROCC (Goyal et al.,2020) | 81.7 | 76.7 | 66.7 | 67.1 | 73.6 | 74.4 | 74.4 | 71.4 | 80.0 | 76.2 | 74.2 |
| Deep-SAD (Ruff et al., 2019) | 二 | 二 | 二 | | | 一 | 一 | | | 一 | 77.9 |
| E3Outlier (Wang et al.,2019) | 79.4 | 95.3 | 75.4 | 73.9 | 84.1 | 87.9 | 85.0 | 93.4 | 92.3 | 89.7 | 85.6 |
| GT(Golan & El-Yaniv,2018) | 74.7 | 95.7 | 78.1 | 72.4 | 87.8 | 87.8 | 83.4 | 95.5 | 93.3 | 91.3 | 86.0 |
| InvAE (Huang et al., 2019) | 78.5 | 89.8 | 86.1 | 77.4 | 90.5 | 84.5 | 89.2 | 92.9 | 92.0 | 85.5 | 86.6 |
| GOAD (Bergman & Hoshen,2020) | 77.2 | 96.7 | 83.3 | 77.7 | 87.8 | 87.8 | 90.0 | 96.1 | 93.8 | 92.0 | 88.2 |
| CSI (Tack et al., 2020) | 89.9 | 99.9 | 93.1 | 86.4 | 93.9 | 93.2 | 95.1 | 98.7 | 97.9 | 95.5 | 94.3 |
| SSD | 82.7 | 98.5 | 84.2 | 84.5 | 84.8 | 90.9 | 91.7 | 95.2 | 92.9 | 94.4 | 90.0 |
",
+ "type": "table",
+ "image_path": "933bd53076a6f2cd02bb095c84b505888b0de7616e85209204997e7b07c365d2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 127,
+ 240,
+ 484,
+ 284.6666666666667
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 284.6666666666667,
+ 484,
+ 329.33333333333337
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 127,
+ 329.33333333333337,
+ 484,
+ 374.00000000000006
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ }
+ ],
+ "index": 6.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "Results. Compared to the baseline SSD detector, one-shot and five-shot settings improve the average",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "AUROC, across all OOD datasets, by 1.6 and 2.1, respectively (Table 1, 2). In particular, we observe",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "large gains with CIFAR-100 as in-distribution and CIFAR-10 as OOD where five-shot detection",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "improves the AUROC from 69.6 to 78.3. We find the use of shrunk covariance estimator most critical",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "in the success of our approach. Use of shrunk covariance estimation itself improves the AUROC from",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "69.6 to 77.1. Then data augmentation further improves it 78.3 for the five-shot detection. With an",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "increasing number of transformed copies of each sample, we also observe improvement in AUROC,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 343,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 343,
+ 474
+ ],
+ "score": 1.0,
+ "content": "though it later plateaus close to ten copies (Appendix B.3).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 479,
+ 505,
+ 534
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "What if additional OOD images are available Note that some earlier works, such as Liang et al.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "(2018), assume that 1000 OOD inputs are available for tuning the detector. We find that with access",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 257,
+ 514
+ ],
+ "score": 1.0,
+ "content": "to this large amount of OOD samples,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 501,
+ 279,
+ 512
+ ],
+ "score": 0.9,
+ "content": "S S D _ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 500,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "can improve the state-of-the-art by an even larger margin.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 524
+ ],
+ "score": 1.0,
+ "content": "For example, with CIFAR-100 as in-distribution and CIFAR-10 as out-of-distribution, it achieves",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 477,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 201,
+ 536
+ ],
+ "score": 1.0,
+ "content": "89.4 AUROC, which is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 523,
+ 228,
+ 533
+ ],
+ "score": 0.86,
+ "content": "1 4 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 522,
+ 477,
+ 536
+ ],
+ "score": 1.0,
+ "content": "higher than the current state-of-the-art (Winkens et al., 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 548,
+ 323,
+ 559
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 322,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 291,
+ 561
+ ],
+ "score": 1.0,
+ "content": "4.4 SUCCESS WHEN USING DATA LABELS",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 548,
+ 322,
+ 559
+ ],
+ "score": 0.64,
+ "content": "( S S D + )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 568,
+ 506,
+ 668
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 506,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 506,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Now we integrate labels of training data in our framework and compare it with the existing state-of-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 579,
+ 506,
+ 593
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 579,
+ 506,
+ 593
+ ],
+ "score": 1.0,
+ "content": "the-art detectors. We report our results in Table 4. Our approach improves the average AUROC by 0.8",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 591,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "over the previous state-of-the-art detector. Our approach also achieves equal or better performance",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 602,
+ 506,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 602,
+ 506,
+ 614
+ ],
+ "score": 1.0,
+ "content": "than previous state-of-the-art across individual pairs of in and out-distribution dataset. For example,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 506,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 506,
+ 625
+ ],
+ "score": 1.0,
+ "content": "using labels in our framework improves the AUROC of Mahalanobis detector from 55.5 to 72.1",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 623,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 623,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "for CIFAR-100 as in-distribution and CIFAR-10 as the OOD dataset. Using the simple softmax",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 647
+ ],
+ "score": 1.0,
+ "content": "probabilities, training a two-layer MLP on learned representations further improves the AUROC to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 177,
+ 658
+ ],
+ "score": 1.0,
+ "content": "78.3. Combining",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 646,
+ 205,
+ 656
+ ],
+ "score": 0.79,
+ "content": "S S D +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 644,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": "with a five-shot OOD detection method further brings a gain of 1.4 in the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 657,
+ 178,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 178,
+ 669
+ ],
+ "score": 1.0,
+ "content": "average AUROC.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 684,
+ 286,
+ 697
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 289,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 289,
+ 700
+ ],
+ "score": 1.0,
+ "content": "5 DISCUSSION AND CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "On tuning hyperparameters in SSD. In our framework, we either explicitly avoid the use of",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "additional tuning-parameters (such as when combining self-supervised and supervised loss functions",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 107,
+ 486,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 80,
+ 502,
+ 102
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 504,
+ 93
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 504,
+ 93
+ ],
+ "score": 1.0,
+ "content": "Table 2: Comparing performance of self-supervised (SSD) and supervised representations. We also",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 90,
+ 502,
+ 104
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 90,
+ 283,
+ 104
+ ],
+ "score": 1.0,
+ "content": "provide results for few-shot OOD detection",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 92,
+ 309,
+ 102
+ ],
+ "score": 0.84,
+ "content": "( S S D _ { k } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 90,
+ 502,
+ 104
+ ],
+ "score": 1.0,
+ "content": ") for comparison with our baseline SSD detector.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 107,
+ 486,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 107,
+ 486,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 107,
+ 486,
+ 217
+ ],
+ "score": 0.98,
+ "html": " | Airplane | Automobile | Bird | Cat | Deer | Dog | Frog | Horse | Ship | Truck | Average |
| Randomly Initialized network | 77.4 | 44.1 | 62.4 | 44.1 | 62.1 | 49.6 | 59.8 | 48.0 | 73.8 | 53.7 | 57.5 |
| VAE(Kingma & Welling,2014) | 70.0 | 38.6 | 67.9 | 53.5 | 74.8 | 52.3 | 68.7 | 49.3 | 69.6 | 38.6 | 58.3 |
| OCSVM (Scholkopf et al., 2001) | 63.0 | 44.0 | 64.9 | 48.7 | 73.5 | 50.0 | 72.5 | 53.3 | 64.9 | 50.8 | 58.5 |
| AnoGAN (Schlegl et al., 2017) | 67.1 | 54.7 | 52.9 | 54.5 | 65.1 | 60.3 | 58.5 | 62.5 | 75.8 | 66.5 | 61.8 |
| PixelCNN(Van den Oord et al.,2016) | 53.1 | 99.5 | 47.6 | 51.7 | 73.9 | 54.2 | 59.2 | 78.9 | 34.0 | 66.2 | 61.8 |
| DSVDD (Ruff et al., 2018) | 61.7 | 65.9 | 50.8 | 59.1 | 60.9 | 65.7 | 67.7 | 67.3 | 75.9 | 73.1 | 64.8 |
| OCGAN (Perera et al.,2019) | 75.7 | 53.1 | 64.0 | 62.0 | 72.3 | 62.0 | 72.3 | 57.5 | 82.0 | 55.4 | 65.6 |
| RCAE(Chalapathy et al., 2018b) | 72.0 | 63.1 | 71.7 | 60.6 | 72.8 | 64.0 | 64.9 | 63.6 | 74.7 | 74.5 | 68.2 |
| DROCC (Goyal et al.,2020) | 81.7 | 76.7 | 66.7 | 67.1 | 73.6 | 74.4 | 74.4 | 71.4 | 80.0 | 76.2 | 74.2 |
| Deep-SAD (Ruff et al., 2019) | 二 | 二 | 二 | | | 一 | 一 | | | 一 | 77.9 |
| E3Outlier (Wang et al.,2019) | 79.4 | 95.3 | 75.4 | 73.9 | 84.1 | 87.9 | 85.0 | 93.4 | 92.3 | 89.7 | 85.6 |
| GT(Golan & El-Yaniv,2018) | 74.7 | 95.7 | 78.1 | 72.4 | 87.8 | 87.8 | 83.4 | 95.5 | 93.3 | 91.3 | 86.0 |
| InvAE (Huang et al., 2019) | 78.5 | 89.8 | 86.1 | 77.4 | 90.5 | 84.5 | 89.2 | 92.9 | 92.0 | 85.5 | 86.6 |
| GOAD (Bergman & Hoshen,2020) | 77.2 | 96.7 | 83.3 | 77.7 | 87.8 | 87.8 | 90.0 | 96.1 | 93.8 | 92.0 | 88.2 |
| CSI (Tack et al., 2020) | 89.9 | 99.9 | 93.1 | 86.4 | 93.9 | 93.2 | 95.1 | 98.7 | 97.9 | 95.5 | 94.3 |
| SSD | 82.7 | 98.5 | 84.2 | 84.5 | 84.8 | 90.9 | 91.7 | 95.2 | 92.9 | 94.4 | 90.0 |
",
+ "type": "table",
+ "image_path": "933bd53076a6f2cd02bb095c84b505888b0de7616e85209204997e7b07c365d2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 127,
+ 240,
+ 484,
+ 284.6666666666667
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 284.6666666666667,
+ 484,
+ 329.33333333333337
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 127,
+ 329.33333333333337,
+ 484,
+ 374.00000000000006
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ }
+ ],
+ "index": 6.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "Results. Compared to the baseline SSD detector, one-shot and five-shot settings improve the average",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "AUROC, across all OOD datasets, by 1.6 and 2.1, respectively (Table 1, 2). In particular, we observe",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "large gains with CIFAR-100 as in-distribution and CIFAR-10 as OOD where five-shot detection",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "improves the AUROC from 69.6 to 78.3. We find the use of shrunk covariance estimator most critical",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "in the success of our approach. Use of shrunk covariance estimation itself improves the AUROC from",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "69.6 to 77.1. Then data augmentation further improves it 78.3 for the five-shot detection. With an",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "increasing number of transformed copies of each sample, we also observe improvement in AUROC,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 343,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 343,
+ 474
+ ],
+ "score": 1.0,
+ "content": "though it later plateaus close to ten copies (Appendix B.3).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 384,
+ 506,
+ 474
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 479,
+ 505,
+ 534
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "What if additional OOD images are available Note that some earlier works, such as Liang et al.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "(2018), assume that 1000 OOD inputs are available for tuning the detector. We find that with access",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 257,
+ 514
+ ],
+ "score": 1.0,
+ "content": "to this large amount of OOD samples,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 501,
+ 279,
+ 512
+ ],
+ "score": 0.9,
+ "content": "S S D _ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 500,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "can improve the state-of-the-art by an even larger margin.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 524
+ ],
+ "score": 1.0,
+ "content": "For example, with CIFAR-100 as in-distribution and CIFAR-10 as out-of-distribution, it achieves",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 477,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 201,
+ 536
+ ],
+ "score": 1.0,
+ "content": "89.4 AUROC, which is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 523,
+ 228,
+ 533
+ ],
+ "score": 0.86,
+ "content": "1 4 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 522,
+ 477,
+ 536
+ ],
+ "score": 1.0,
+ "content": "higher than the current state-of-the-art (Winkens et al., 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 478,
+ 506,
+ 536
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 548,
+ 323,
+ 559
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 322,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 291,
+ 561
+ ],
+ "score": 1.0,
+ "content": "4.4 SUCCESS WHEN USING DATA LABELS",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 548,
+ 322,
+ 559
+ ],
+ "score": 0.64,
+ "content": "( S S D + )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 568,
+ 506,
+ 668
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 506,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 506,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Now we integrate labels of training data in our framework and compare it with the existing state-of-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 579,
+ 506,
+ 593
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 579,
+ 506,
+ 593
+ ],
+ "score": 1.0,
+ "content": "the-art detectors. We report our results in Table 4. Our approach improves the average AUROC by 0.8",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 591,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "over the previous state-of-the-art detector. Our approach also achieves equal or better performance",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 602,
+ 506,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 602,
+ 506,
+ 614
+ ],
+ "score": 1.0,
+ "content": "than previous state-of-the-art across individual pairs of in and out-distribution dataset. For example,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 506,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 506,
+ 625
+ ],
+ "score": 1.0,
+ "content": "using labels in our framework improves the AUROC of Mahalanobis detector from 55.5 to 72.1",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 623,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 623,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "for CIFAR-100 as in-distribution and CIFAR-10 as the OOD dataset. Using the simple softmax",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 647
+ ],
+ "score": 1.0,
+ "content": "probabilities, training a two-layer MLP on learned representations further improves the AUROC to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 177,
+ 658
+ ],
+ "score": 1.0,
+ "content": "78.3. Combining",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 646,
+ 205,
+ 656
+ ],
+ "score": 0.79,
+ "content": "S S D +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 644,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": "with a five-shot OOD detection method further brings a gain of 1.4 in the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 657,
+ 178,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 178,
+ 669
+ ],
+ "score": 1.0,
+ "content": "average AUROC.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 569,
+ 506,
+ 669
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 684,
+ 286,
+ 697
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 289,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 289,
+ 700
+ ],
+ "score": 1.0,
+ "content": "5 DISCUSSION AND CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "On tuning hyperparameters in SSD. In our framework, we either explicitly avoid the use of",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "additional tuning-parameters (such as when combining self-supervised and supervised loss functions",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 117,
+ 293
+ ],
+ "score": 1.0,
+ "content": "in",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 117,
+ 280,
+ 144,
+ 291
+ ],
+ "score": 0.82,
+ "content": "S S D +",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 144,
+ 279,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": ") or refrain from tuning the existing set of parameters for each OOD dataset. For example,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 292,
+ 506,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 292,
+ 506,
+ 303
+ ],
+ "score": 1.0,
+ "content": "we use a standard set of parameters for self-supervised training and model the learned features with a",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 301,
+ 165,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 301,
+ 165,
+ 315
+ ],
+ "score": 1.0,
+ "content": "single-cluster.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 33.5,
+ "bbox_fs": [
+ 105,
+ 709,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 109,
+ 107,
+ 503,
+ 228
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 80,
+ 504,
+ 103
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 506,
+ 93
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 206,
+ 93
+ ],
+ "score": 1.0,
+ "content": "Table 4: Comparison of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 206,
+ 81,
+ 232,
+ 91
+ ],
+ "score": 0.84,
+ "content": "S S D +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 79,
+ 506,
+ 93
+ ],
+ "score": 1.0,
+ "content": ", i.e., incorporating labels in the SSD detector, with state-of-the-art",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 89,
+ 263,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 89,
+ 263,
+ 105
+ ],
+ "score": 1.0,
+ "content": "detectors based on supervised training.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 109,
+ 107,
+ 503,
+ 228
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 107,
+ 503,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 107,
+ 503,
+ 228
+ ],
+ "score": 0.984,
+ "html": "| Algorithm 1: Self-supervised outlier detection framework (SSD) | |
| Output :Is outlier or not? ∀x E Xtest | Input :Xin, Xtest, feature extractor (f), projection head (h),Required True-positive rate (T), |
| Function getFeatures(X): return {f(xi)/llf(xi)ll2,∀ xi ∈ X}; | |
| Function SSDkScore(Z, μin,Σin, μood,Σood): | Function SSDScore(Z,μ,∑): return {(z - μ)T∑-1(z - μ),∀ z ∈ Z}; |
| end | return{(z-μin)T∑-1(z-μin)-(z-μood)T∑-ld(z-μood),∀z∈Z}; |
| Parition Xin in training set (Xtrain) and calibration set (Xcal); | |
| if Vin isnot available then | |
| Lbatch = N∑²N1 uj/T -log eu | h(f(xi)) |
| N 1（k≠i）eu uk/T | ；ui= |
| else | |
| 1 | |
| 2Nyi -log | 11(k≠i)1(yk=yi）eTuk/T |
| 1（k≠i)e²Tuk/T |
| end | |
| |
| Train feature extractor(f) by minimizing Lbatch over Xtrain; |
| Ztrain = getFeatures(Xtrain),Zcal = getFeatures(Xcal) |
| Ztest = getFeatures(Xtest),if Xood is available: Zood = getFeatures(Xood); |
| if Xood is not available then | |
| |
| Scal = SSDScore(Zcal, μtrain,∑train); # Sample mean and convariance of Ztrain |
| Stest = SSDScore(Ztest, μtrain,∑train) # outlier score; |
| else | |
| # Using convariance estimation techniques from Section 3.2 for Zood |
| Scal = SSDkScore(Zcal, μtrain,∑train, μood,∑ood); | |
| |
| Stest = SSDkScore(Ztest, μtrain,∑train,μood,∑ood); |
| end | |
| |
| |
| |
| |
| Xi E Xtest is an outlier if stest > (Scal threshold at TPR = T); | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
"
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 840,
+ 2089,
+ 858,
+ 2089,
+ 858,
+ 2112,
+ 840,
+ 2112
+ ],
+ "score": 0.77
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 676,
+ 368,
+ 1005,
+ 368,
+ 1005,
+ 401,
+ 676,
+ 401
+ ],
+ "score": 0.91,
+ "latex": "\\{ f ( x _ { i } ) / \\| f ( x _ { i } ) \\| _ { 2 } , \\forall x _ { i } \\in \\mathcal { X } \\}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 898,
+ 785,
+ 973,
+ 785,
+ 973,
+ 814,
+ 898,
+ 814
+ ],
+ "score": 0.91,
+ "latex": "\\chi _ { t r a i n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 392,
+ 529,
+ 437,
+ 529,
+ 437,
+ 558,
+ 392,
+ 558
+ ],
+ "score": 0.9,
+ "latex": "\\mathcal { X } _ { i n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 762,
+ 784,
+ 838,
+ 784,
+ 838,
+ 815,
+ 762,
+ 815
+ ],
+ "score": 0.9,
+ "latex": "\\mathcal { L } _ { b a t c h }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 716,
+ 403,
+ 1093,
+ 403,
+ 1093,
+ 437,
+ 716,
+ 437
+ ],
+ "score": 0.9,
+ "latex": "\\{ ( z - \\mu ) ^ { T } \\Sigma ^ { - 1 } ( z - \\mu ) , \\forall z \\in \\mathcal { Z } \\}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 1122,
+ 418,
+ 1122,
+ 418,
+ 1153,
+ 297,
+ 1153
+ ],
+ "score": 0.89,
+ "latex": "x _ { i } \\in \\mathcal { X } _ { t e s t }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 322,
+ 877,
+ 380,
+ 877,
+ 380,
+ 906,
+ 322,
+ 906
+ ],
+ "score": 0.89,
+ "latex": "\\mathcal { X } _ { o o d }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 321,
+ 559,
+ 365,
+ 559,
+ 365,
+ 589,
+ 321,
+ 589
+ ],
+ "score": 0.89,
+ "latex": "\\mathcal { \\ V } _ { i n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 694,
+ 846,
+ 752,
+ 846,
+ 752,
+ 875,
+ 694,
+ 875
+ ],
+ "score": 0.88,
+ "latex": "\\mathcal { X } _ { o o d }"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 336,
+ 678,
+ 949,
+ 678,
+ 949,
+ 757,
+ 336,
+ 757
+ ],
+ "score": 0.88,
+ "latex": "\\begin{array} { r } { \\mathcal { L } _ { b a t c h } = \\frac { 1 } { 2 N } \\sum _ { i = 1 } ^ { 2 N } - l o g \\frac { \\frac { 1 } { 2 N _ { y _ { i } } - 1 } \\sum _ { k = 1 } ^ { 2 N } 1 ( k \\neq i ) 1 ( y _ { k } = y _ { i } ) e ^ { u _ { i } ^ { T } u _ { k } / \\tau } } { \\sum _ { k = 1 } ^ { 2 N } 1 ( k \\neq i ) e ^ { u _ { i } ^ { T } u _ { k } / \\tau } } ; } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 456,
+ 277,
+ 517,
+ 277,
+ 517,
+ 307,
+ 456,
+ 307
+ ],
+ "score": 0.86,
+ "latex": "\\chi _ { t e s t }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 605,
+ 529,
+ 694,
+ 529,
+ 694,
+ 559,
+ 605,
+ 559
+ ],
+ "score": 0.86,
+ "latex": "( \\mathcal { X } _ { t r a i n } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 913,
+ 529,
+ 980,
+ 529,
+ 980,
+ 559,
+ 913,
+ 559
+ ],
+ "score": 0.86,
+ "latex": "( \\mathcal { X } _ { c a l } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 400,
+ 277,
+ 444,
+ 277,
+ 444,
+ 308,
+ 400,
+ 308
+ ],
+ "score": 0.85,
+ "latex": "\\mathcal { X } _ { i n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 552,
+ 785,
+ 589,
+ 785,
+ 589,
+ 816,
+ 552,
+ 816
+ ],
+ "score": 0.84,
+ "latex": "( f )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 513,
+ 308,
+ 570,
+ 308,
+ 570,
+ 338,
+ 513,
+ 338
+ ],
+ "score": 0.84,
+ "latex": "\\mathcal { X } _ { o o d }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 591,
+ 339,
+ 718,
+ 339,
+ 718,
+ 368,
+ 591,
+ 368
+ ],
+ "score": 0.83,
+ "latex": "\\forall x \\in \\mathcal { X } _ { t e s t }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 825,
+ 1493,
+ 944,
+ 1493,
+ 944,
+ 1522,
+ 825,
+ 1522
+ ],
+ "score": 0.83,
+ "latex": "\\mathrm { T P R } { = } 9 5 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 583,
+ 1119,
+ 730,
+ 1119,
+ 730,
+ 1156,
+ 583,
+ 1156
+ ],
+ "score": 0.81,
+ "latex": "s _ { t e s t } ^ { i } > ( s _ { c a l }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 580,
+ 308,
+ 625,
+ 308,
+ 625,
+ 338,
+ 580,
+ 338
+ ],
+ "score": 0.8,
+ "latex": "\\mathcal { \\gamma } _ { i n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 718,
+ 278,
+ 754,
+ 278,
+ 754,
+ 309,
+ 718,
+ 309
+ ],
+ "score": 0.8,
+ "latex": "( f )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 339,
+ 938,
+ 824,
+ 938,
+ 824,
+ 969,
+ 339,
+ 969
+ ],
+ "score": 0.78,
+ "latex": "s _ { t e s t } = { S S } \\mathsf { D S c o r e } \\left( \\mathcal { Z } _ { t e s t } , \\mu _ { t r a i n } , \\boldsymbol { \\Sigma } _ { t r a i n } \\right) \\mathrm { : }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 338,
+ 906,
+ 800,
+ 906,
+ 800,
+ 938,
+ 338,
+ 938
+ ],
+ "score": 0.77,
+ "latex": "s _ { c a l } = { \\tt S S D S c o r e } ( \\mathcal { Z } _ { c a l } , \\mu _ { t r a i n } , \\Sigma _ { t r a i n } ) { } _ { }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 534,
+ 436,
+ 812,
+ 436,
+ 812,
+ 465,
+ 534,
+ 465
+ ],
+ "score": 0.77,
+ "latex": "( \\mathcal { Z } , \\mu _ { i n } , \\Sigma _ { i n } , \\mu _ { o o d } , \\Sigma _ { o o d } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 870,
+ 1122,
+ 976,
+ 1122,
+ 976,
+ 1152,
+ 870,
+ 1152
+ ],
+ "score": 0.76,
+ "latex": "\\mathrm { T P R } = T "
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1052,
+ 996,
+ 1109,
+ 996,
+ 1109,
+ 1026,
+ 1052,
+ 1026
+ ],
+ "score": 0.76,
+ "latex": "{ \\mathcal { Z } } _ { o o d }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1203,
+ 908,
+ 1279,
+ 908,
+ 1279,
+ 937,
+ 1203,
+ 937
+ ],
+ "score": 0.74,
+ "latex": "\\mathcal { Z } _ { t r a i n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 945,
+ 278,
+ 980,
+ 278,
+ 980,
+ 307,
+ 945,
+ 307
+ ],
+ "score": 0.72,
+ "latex": "( h )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 594,
+ 815,
+ 688,
+ 815,
+ 688,
+ 846,
+ 594,
+ 846
+ ],
+ "score": 0.72,
+ "latex": "( \\mathcal { X } _ { t r a i n } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1305,
+ 278,
+ 1343,
+ 278,
+ 1343,
+ 308,
+ 1305,
+ 308
+ ],
+ "score": 0.7,
+ "latex": "( T )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 543,
+ 369,
+ 584,
+ 369,
+ 584,
+ 399,
+ 543,
+ 399
+ ],
+ "score": 0.67,
+ "latex": "( \\mathcal { X } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 847,
+ 393,
+ 847,
+ 393,
+ 876,
+ 298,
+ 876
+ ],
+ "score": 0.66,
+ "latex": "\\mathcal { Z } _ { t e s t } ="
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 651,
+ 307,
+ 922,
+ 307,
+ 922,
+ 341,
+ 651,
+ 341
+ ],
+ "score": 0.63,
+ "latex": "\\mathcal { X } _ { i n } \\in \\mathbb { P } _ { X } ^ { i n } , \\mathcal { X } _ { o o d } \\in \\mathbb { P } _ { X } ^ { o o d }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 815,
+ 409,
+ 815,
+ 409,
+ 845,
+ 298,
+ 845
+ ],
+ "score": 0.6,
+ "latex": "\\mathcal { Z } _ { t r a i n } ="
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 700,
+ 816,
+ 787,
+ 816,
+ 787,
+ 845,
+ 700,
+ 845
+ ],
+ "score": 0.58,
+ "latex": "\\mathcal { Z } _ { c a l } ="
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 414,
+ 463,
+ 1196,
+ 463,
+ 1196,
+ 503,
+ 414,
+ 503
+ ],
+ "score": 0.55,
+ "latex": "\\{ ( z - \\mu _ { i n } ) ^ { T } \\Sigma _ { i n } ^ { - 1 } ( z - \\mu _ { i n } ) - ( z - \\mu _ { o o d } ) ^ { T } \\Sigma _ { o o d } ^ { - 1 } ( z - \\mu _ { o o d } ) , \\forall z \\in \\mathcal { Z } \\} ;"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1160,
+ 844,
+ 1254,
+ 844,
+ 1254,
+ 878,
+ 1160,
+ 878
+ ],
+ "score": 0.54,
+ "latex": "s ( \\mathcal { X } _ { o o d } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 954,
+ 240,
+ 1018,
+ 240,
+ 1018,
+ 269,
+ 954,
+ 269
+ ],
+ "score": 0.53,
+ "latex": "( S S D )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 579,
+ 846,
+ 657,
+ 846,
+ 657,
+ 878,
+ 579,
+ 878
+ ],
+ "score": 0.52,
+ "latex": "( \\mathcal { X } _ { t e s t } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 522,
+ 404,
+ 625,
+ 404,
+ 625,
+ 435,
+ 522,
+ 435
+ ],
+ "score": 0.5,
+ "latex": "( \\mathcal { Z } , \\mu , \\Sigma )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 956,
+ 815,
+ 1044,
+ 815,
+ 1044,
+ 847,
+ 956,
+ 847
+ ],
+ "score": 0.49,
+ "latex": "s ( \\mathcal { X } _ { c a l } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 335,
+ 587,
+ 1154,
+ 587,
+ 1154,
+ 653,
+ 335,
+ 653
+ ],
+ "score": 0.42,
+ "latex": "\\begin{array} { r } { \\mathcal { L } _ { b a t c h } = \\frac { 1 } { 2 N } \\sum _ { i = 1 } ^ { 2 N } - l o g \\frac { e ^ { u _ { i } ^ { T } u _ { j } / \\tau } } { \\sum _ { k = 1 } ^ { 2 N } 1 ( k \\neq i ) e ^ { u _ { i } ^ { T } u _ { k } / \\tau } } ; u _ { i } = \\frac { h ( f ( x _ { i } ) ) } { \\| h ( f ( x _ { i } ) ) \\| _ { 2 } } ; \\# \\mathrm { T r a i n f e a t u r e ~ e x t r a c t o r } } \\end{array}"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 334,
+ 587,
+ 1167,
+ 587,
+ 1167,
+ 653,
+ 334,
+ 653
+ ],
+ "score": 0.4,
+ "latex": "\\begin{array} { r } { \\mathcal { L } _ { b a t c h } = \\frac { 1 } { 2 N } \\sum _ { i = 1 } ^ { 2 N } - l o g \\frac { e ^ { u _ { i } ^ { T } u _ { j } / \\tau } } { \\sum _ { k = 1 } ^ { 2 N } 1 ( k \\neq i ) e ^ { u _ { i } ^ { T } u _ { k } / \\tau } } ; u _ { i } = \\frac { h ( f ( x _ { i } ) ) } { \\| h ( f ( x _ { i } ) ) \\| _ { 2 } } ; \\# \\mathrm { T r a i n f e a t u r e ~ e x t r a c t o r } } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 337,
+ 1054,
+ 959,
+ 1054,
+ 959,
+ 1089,
+ 337,
+ 1089
+ ],
+ "score": 0.38,
+ "latex": "s _ { t e s t } = { \\tt S S D k S c o r e } ( \\mathcal { Z } _ { t e s t } , \\mu _ { t r a i n } , \\Sigma _ { t r a i n } , \\mu _ { o o d } , \\Sigma _ { o o d } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 847,
+ 660,
+ 847,
+ 660,
+ 877,
+ 298,
+ 877
+ ],
+ "score": 0.37,
+ "latex": "\\mathcal { Z } _ { t e s t } = \\mathsf { g e t F e a t u r e s } ( \\mathcal { X } _ { t e s t } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 901,
+ 846,
+ 990,
+ 846,
+ 990,
+ 876,
+ 901,
+ 876
+ ],
+ "score": 0.34,
+ "latex": "{ \\mathcal { Z } } _ { o o d } ="
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 375,
+ 1025,
+ 934,
+ 1025,
+ 934,
+ 1057,
+ 375,
+ 1057
+ ],
+ "score": 0.31,
+ "latex": "\\mathsf { \\Pi } _ { l } = \\mathtt { S S D k S c o r e } ( \\mathcal { Z } _ { c a l } , \\mu _ { t r a i n } , \\sum _ { t r a i n } , \\mu _ { o o d } , \\sum _ { o o d }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 900,
+ 847,
+ 1257,
+ 847,
+ 1257,
+ 877,
+ 900,
+ 877
+ ],
+ "score": 0.25,
+ "latex": "\\mathcal { Z } _ { o o d } = \\mathsf { g e t F e a t u r e s } ( \\mathcal { X } _ { o o d } ) ;"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1910.0,
+ 651.0,
+ 1910.0,
+ 651.0,
+ 1949.0,
+ 294.0,
+ 1949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1250.0,
+ 894.0,
+ 1250.0,
+ 894.0,
+ 1285.0,
+ 294.0,
+ 1285.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 71.0,
+ 815.0,
+ 71.0,
+ 815.0,
+ 108.0,
+ 294.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1192.0,
+ 713.0,
+ 1192.0,
+ 713.0,
+ 1235.0,
+ 293.0,
+ 1235.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 840.0,
+ 2087.0,
+ 861.0,
+ 2087.0,
+ 861.0,
+ 2118.0,
+ 840.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1308.0,
+ 1407.0,
+ 1308.0,
+ 1407.0,
+ 1345.0,
+ 295.0,
+ 1345.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1341.0,
+ 1404.0,
+ 1341.0,
+ 1404.0,
+ 1373.0,
+ 295.0,
+ 1373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1368.0,
+ 1407.0,
+ 1368.0,
+ 1407.0,
+ 1406.0,
+ 294.0,
+ 1406.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1398.0,
+ 1408.0,
+ 1398.0,
+ 1408.0,
+ 1437.0,
+ 292.0,
+ 1437.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1431.0,
+ 1408.0,
+ 1431.0,
+ 1408.0,
+ 1468.0,
+ 294.0,
+ 1468.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1461.0,
+ 1406.0,
+ 1461.0,
+ 1406.0,
+ 1496.0,
+ 292.0,
+ 1496.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1492.0,
+ 824.0,
+ 1492.0,
+ 824.0,
+ 1529.0,
+ 294.0,
+ 1529.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 945.0,
+ 1492.0,
+ 1406.0,
+ 1492.0,
+ 1406.0,
+ 1529.0,
+ 945.0,
+ 1529.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1522.0,
+ 1406.0,
+ 1522.0,
+ 1406.0,
+ 1557.0,
+ 292.0,
+ 1557.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1551.0,
+ 1406.0,
+ 1551.0,
+ 1406.0,
+ 1591.0,
+ 292.0,
+ 1591.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1583.0,
+ 1406.0,
+ 1583.0,
+ 1406.0,
+ 1618.0,
+ 292.0,
+ 1618.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1614.0,
+ 1404.0,
+ 1614.0,
+ 1404.0,
+ 1650.0,
+ 294.0,
+ 1650.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1642.0,
+ 1406.0,
+ 1642.0,
+ 1406.0,
+ 1681.0,
+ 294.0,
+ 1681.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1675.0,
+ 794.0,
+ 1675.0,
+ 794.0,
+ 1710.0,
+ 295.0,
+ 1710.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1721.0,
+ 1404.0,
+ 1721.0,
+ 1404.0,
+ 1754.0,
+ 297.0,
+ 1754.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1750.0,
+ 1404.0,
+ 1750.0,
+ 1404.0,
+ 1786.0,
+ 293.0,
+ 1786.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1782.0,
+ 1404.0,
+ 1782.0,
+ 1404.0,
+ 1815.0,
+ 294.0,
+ 1815.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1814.0,
+ 1403.0,
+ 1814.0,
+ 1403.0,
+ 1847.0,
+ 295.0,
+ 1847.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1840.0,
+ 1106.0,
+ 1840.0,
+ 1106.0,
+ 1880.0,
+ 293.0,
+ 1880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1971.0,
+ 1406.0,
+ 1971.0,
+ 1406.0,
+ 2007.0,
+ 297.0,
+ 2007.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 2002.0,
+ 1405.0,
+ 2002.0,
+ 1405.0,
+ 2037.0,
+ 296.0,
+ 2037.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 5,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1537,
+ 1404,
+ 1537,
+ 1404,
+ 1843,
+ 297,
+ 1843
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 5,
+ "poly": [
+ 328,
+ 253,
+ 1377,
+ 253,
+ 1377,
+ 534,
+ 328,
+ 534
+ ],
+ "score": 0.982,
+ "html": " | Airplane | Automobile | Bird | Cat | Deer | Dog | Frog | Horse | Ship | Truck | Average |
| Randomly Initialized network | 77.4 | 44.1 | 62.4 | 44.1 | 62.1 | 49.6 | 59.8 | 48.0 | 73.8 | 53.7 | 57.5 |
| VAE(Kingma & Welling,2014) | 70.0 | 38.6 | 67.9 | 53.5 | 74.8 | 52.3 | 68.7 | 49.3 | 69.6 | 38.6 | 58.3 |
| OCSVM (Scholkopf et al., 2001) | 63.0 | 44.0 | 64.9 | 48.7 | 73.5 | 50.0 | 72.5 | 53.3 | 64.9 | 50.8 | 58.5 |
| AnoGAN (Schlegl et al., 2017) | 67.1 | 54.7 | 52.9 | 54.5 | 65.1 | 60.3 | 58.5 | 62.5 | 75.8 | 66.5 | 61.8 |
| PixelCNN(Van den Oord et al.,2016) | 53.1 | 99.5 | 47.6 | 51.7 | 73.9 | 54.2 | 59.2 | 78.9 | 34.0 | 66.2 | 61.8 |
| DSVDD (Ruff et al., 2018) | 61.7 | 65.9 | 50.8 | 59.1 | 60.9 | 65.7 | 67.7 | 67.3 | 75.9 | 73.1 | 64.8 |
| OCGAN (Perera et al.,2019) | 75.7 | 53.1 | 64.0 | 62.0 | 72.3 | 62.0 | 72.3 | 57.5 | 82.0 | 55.4 | 65.6 |
| RCAE(Chalapathy et al., 2018b) | 72.0 | 63.1 | 71.7 | 60.6 | 72.8 | 64.0 | 64.9 | 63.6 | 74.7 | 74.5 | 68.2 |
| DROCC (Goyal et al.,2020) | 81.7 | 76.7 | 66.7 | 67.1 | 73.6 | 74.4 | 74.4 | 71.4 | 80.0 | 76.2 | 74.2 |
| Deep-SAD (Ruff et al., 2019) | 二 | 二 | 二 | | | 一 | 一 | | | 一 | 77.9 |
| E3Outlier (Wang et al.,2019) | 79.4 | 95.3 | 75.4 | 73.9 | 84.1 | 87.9 | 85.0 | 93.4 | 92.3 | 89.7 | 85.6 |
| GT(Golan & El-Yaniv,2018) | 74.7 | 95.7 | 78.1 | 72.4 | 87.8 | 87.8 | 83.4 | 95.5 | 93.3 | 91.3 | 86.0 |
| InvAE (Huang et al., 2019) | 78.5 | 89.8 | 86.1 | 77.4 | 90.5 | 84.5 | 89.2 | 92.9 | 92.0 | 85.5 | 86.6 |
| GOAD (Bergman & Hoshen,2020) | 77.2 | 96.7 | 83.3 | 77.7 | 87.8 | 87.8 | 90.0 | 96.1 | 93.8 | 92.0 | 88.2 |
| CSI (Tack et al., 2020) | 89.9 | 99.9 | 93.1 | 86.4 | 93.9 | 93.2 | 95.1 | 98.7 | 97.9 | 95.5 | 94.3 |
| SSD | 82.7 | 98.5 | 84.2 | 84.5 | 84.8 | 90.9 | 91.7 | 95.2 | 92.9 | 94.4 | 90.0 |
"
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1071,
+ 1404,
+ 1071,
+ 1404,
+ 1317,
+ 297,
+ 1317
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1331,
+ 1405,
+ 1331,
+ 1405,
+ 1485,
+ 298,
+ 1485
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 296,
+ 1972,
+ 1402,
+ 1972,
+ 1402,
+ 2034,
+ 296,
+ 2034
+ ],
+ "score": 0.944
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 299,
+ 224,
+ 1396,
+ 224,
+ 1396,
+ 286,
+ 299,
+ 286
+ ],
+ "score": 0.937
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 1902,
+ 797,
+ 1902,
+ 797,
+ 1938,
+ 302,
+ 1938
+ ],
+ "score": 0.906
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 76,
+ 812,
+ 76,
+ 812,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.893
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 304,
+ 1523,
+ 898,
+ 1523,
+ 898,
+ 1554,
+ 304,
+ 1554
+ ],
+ "score": 0.834
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2089,
+ 858,
+ 2089,
+ 858,
+ 2111,
+ 841,
+ 2111
+ ],
+ "score": 0.787
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 300,
+ 622,
+ 1394,
+ 622,
+ 1394,
+ 653,
+ 300,
+ 653
+ ],
+ "score": 0.625
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 299,
+ 622,
+ 1394,
+ 622,
+ 1394,
+ 653,
+ 299,
+ 653
+ ],
+ "score": 0.25
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 715,
+ 1394,
+ 776,
+ 1394,
+ 776,
+ 1423,
+ 715,
+ 1423
+ ],
+ "score": 0.9,
+ "latex": "S S D _ { k }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 562,
+ 1454,
+ 636,
+ 1454,
+ 636,
+ 1483,
+ 562,
+ 1483
+ ],
+ "score": 0.86,
+ "latex": "1 4 . 2 \\%"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 788,
+ 256,
+ 860,
+ 256,
+ 860,
+ 286,
+ 788,
+ 286
+ ],
+ "score": 0.84,
+ "latex": "( S S D _ { k } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 495,
+ 1795,
+ 570,
+ 1795,
+ 570,
+ 1824,
+ 495,
+ 1824
+ ],
+ "score": 0.79,
+ "latex": "S S D +"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 811,
+ 1524,
+ 895,
+ 1524,
+ 895,
+ 1554,
+ 811,
+ 1554
+ ],
+ "score": 0.64,
+ "latex": "( S S D + )"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 223.0,
+ 1401.0,
+ 223.0,
+ 1401.0,
+ 259.0,
+ 295.0,
+ 259.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 252.0,
+ 787.0,
+ 252.0,
+ 787.0,
+ 290.0,
+ 293.0,
+ 290.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 861.0,
+ 252.0,
+ 1397.0,
+ 252.0,
+ 1397.0,
+ 290.0,
+ 861.0,
+ 290.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1895.0,
+ 805.0,
+ 1895.0,
+ 805.0,
+ 1946.0,
+ 291.0,
+ 1946.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 815.0,
+ 73.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1520.0,
+ 810.0,
+ 1520.0,
+ 810.0,
+ 1560.0,
+ 295.0,
+ 1560.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 896.0,
+ 1520.0,
+ 904.0,
+ 1520.0,
+ 904.0,
+ 1560.0,
+ 896.0,
+ 1560.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2086.0,
+ 860.0,
+ 2086.0,
+ 860.0,
+ 2116.0,
+ 839.0,
+ 2116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 299.0,
+ 617.0,
+ 1398.0,
+ 617.0,
+ 1398.0,
+ 655.0,
+ 299.0,
+ 655.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 299.0,
+ 618.0,
+ 1399.0,
+ 618.0,
+ 1399.0,
+ 657.0,
+ 299.0,
+ 657.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1581.0,
+ 1408.0,
+ 1581.0,
+ 1408.0,
+ 1617.0,
+ 295.0,
+ 1617.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1611.0,
+ 1406.0,
+ 1611.0,
+ 1406.0,
+ 1648.0,
+ 295.0,
+ 1648.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1642.0,
+ 1406.0,
+ 1642.0,
+ 1406.0,
+ 1679.0,
+ 295.0,
+ 1679.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1674.0,
+ 1406.0,
+ 1674.0,
+ 1406.0,
+ 1706.0,
+ 296.0,
+ 1706.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1701.0,
+ 1406.0,
+ 1701.0,
+ 1406.0,
+ 1738.0,
+ 294.0,
+ 1738.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1732.0,
+ 1404.0,
+ 1732.0,
+ 1404.0,
+ 1768.0,
+ 295.0,
+ 1768.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1764.0,
+ 1406.0,
+ 1764.0,
+ 1406.0,
+ 1799.0,
+ 294.0,
+ 1799.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1791.0,
+ 494.0,
+ 1791.0,
+ 494.0,
+ 1830.0,
+ 292.0,
+ 1830.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 571.0,
+ 1791.0,
+ 1404.0,
+ 1791.0,
+ 1404.0,
+ 1830.0,
+ 571.0,
+ 1830.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1825.0,
+ 496.0,
+ 1825.0,
+ 496.0,
+ 1859.0,
+ 292.0,
+ 1859.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1068.0,
+ 1405.0,
+ 1068.0,
+ 1405.0,
+ 1109.0,
+ 292.0,
+ 1109.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1101.0,
+ 1405.0,
+ 1101.0,
+ 1405.0,
+ 1135.0,
+ 294.0,
+ 1135.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1133.0,
+ 1404.0,
+ 1133.0,
+ 1404.0,
+ 1165.0,
+ 294.0,
+ 1165.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1163.0,
+ 1405.0,
+ 1163.0,
+ 1405.0,
+ 1196.0,
+ 295.0,
+ 1196.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1193.0,
+ 1404.0,
+ 1193.0,
+ 1404.0,
+ 1226.0,
+ 294.0,
+ 1226.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1220.0,
+ 1405.0,
+ 1220.0,
+ 1405.0,
+ 1259.0,
+ 292.0,
+ 1259.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1255.0,
+ 1407.0,
+ 1255.0,
+ 1407.0,
+ 1289.0,
+ 295.0,
+ 1289.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1285.0,
+ 954.0,
+ 1285.0,
+ 954.0,
+ 1319.0,
+ 296.0,
+ 1319.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1330.0,
+ 1408.0,
+ 1330.0,
+ 1408.0,
+ 1367.0,
+ 295.0,
+ 1367.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1362.0,
+ 1405.0,
+ 1362.0,
+ 1405.0,
+ 1395.0,
+ 296.0,
+ 1395.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1391.0,
+ 714.0,
+ 1391.0,
+ 714.0,
+ 1429.0,
+ 292.0,
+ 1429.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 777.0,
+ 1391.0,
+ 1408.0,
+ 1391.0,
+ 1408.0,
+ 1429.0,
+ 777.0,
+ 1429.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1420.0,
+ 1406.0,
+ 1420.0,
+ 1406.0,
+ 1458.0,
+ 292.0,
+ 1458.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1452.0,
+ 561.0,
+ 1452.0,
+ 561.0,
+ 1489.0,
+ 295.0,
+ 1489.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 637.0,
+ 1452.0,
+ 1325.0,
+ 1452.0,
+ 1325.0,
+ 1489.0,
+ 637.0,
+ 1489.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1971.0,
+ 1406.0,
+ 1971.0,
+ 1406.0,
+ 2007.0,
+ 295.0,
+ 2007.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 2002.0,
+ 1404.0,
+ 2002.0,
+ 1404.0,
+ 2038.0,
+ 294.0,
+ 2038.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 7,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 5,
+ "poly": [
+ 303,
+ 298,
+ 1399,
+ 298,
+ 1399,
+ 634,
+ 303,
+ 634
+ ],
+ "score": 0.984,
+ "html": "