Slide 1

Slide 1 text

深層学習に潜むバイアス 2024.6.13 中島 悠太(⼤阪⼤学)

Slide 2

Slide 2 text

ࣗݾ঺հ • தౡ༔ଠ େࡕେֶσʔλϏϦςΟϑϩϯςΟΞػߏ ڭत • ུྺ • 2012 • 2012 • 2015 • 2017 • 2024 • ݚڀ෼໺ • ίϯϐϡʔλϏδϣϯ • ύλʔϯೝࣝ • ʢࣗવݴޠॲཧʣ • ಛʹɺVision and Language -2012: େࡕେֶେֶӃ޻ֶݚڀՊ ത࢜ޙظ՝ఔ -2012: UNC Charlotte, Visiting Scholar -2016: NAIST, ॿڭ -2016: CMU, Visiting Scholar -2024: େࡕେֶσʔλϏϦςΟϑϩϯςΟΞػߏ ।ڭत -2017: ݱ৬

Slide 3

Slide 3 text

ࠓ೔ͷ಺༰ • ը૾ɾө૾͸໾ʹཱͨͳ͍ʁ • ͳͥը૾ɾө૾͸໾ʹཱͨͳ͍ͷ͔ • ࣾձతόΠΞε

Slide 4

Slide 4 text

ө૾ɾը૾͸໾ʹཱͨͳ͍ʁ ਂ૚ֶशʹજΉόΠΞε

Slide 5

Slide 5 text

Vision and Languageͷ͍Ζ͍ΖͳλεΫͷྫ Visual Question Answering [Agrawal et al., “VQA: Visual question answering,” ICCV 2015] Image Captioning [Chen et al., “Microsoft COCO Captions,” arXiv:1504.00325, 2015] Dense Video Captioning [Krishna et al., “Dense-captioning events in videos,” ICCV 2017] Image Generation E.g. [Rombach et al., “High-resolution image synthesis with latent diffusion models,” CVPR 2022]

Slide 6

Slide 6 text

஌ࣝΛཁٻ͢ΔVideo QA [Garcial et al., “KnowIT VQA: Answering Knowledge-Based Questions about Videos,” AAAI 2020]

Slide 7

Slide 7 text

ϕʔεϥΠϯϞσϧͷਫ਼౓ൺֱ Text-only 🥺 [Garcial et al., “KnowIT VQA: Answering Knowledge-Based Questions about Videos,” AAAI 2020]

Slide 8

Slide 8 text

গ͠Ͱ΋ੑೳΛྑ͍ͨ͘͠ɾɾɾ [Garcial et al., “Knowledge-based video question answering with unsupervised scene descriptions,” ECCV 2020]

Slide 9

Slide 9 text

෦෼ө૾ݕࡧͰ΋ʁ 9 R@1 (IoU > 0.5) ߴੑೳ 映像を全く使わない手法が最新の手法と比肩 ΫΤϦ: A man is seen speaking to the camera and holding up a paper. ࣌ؒ ΫΤϦʹରԠ͢Δ۠ؒ ݕࡧΫΤϦ 🥺 [Otani et al., “Uncovering hidden challenges in query-based video moment retrieval,” BMVC 2020]

Slide 10

Slide 10 text

VQAλεΫͰ͸? From: [Antol et al., “VQA: Visual question answering,” ICCV 2015]

Slide 11

Slide 11 text

VQAλεΫͰ͸ʁ Text-only Text and image From: [Agrawal et al., “Don’t Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering,” CVPR 2018] 🥺

Slide 12

Slide 12 text

”Vision”ͷ෦෼ ຊ౰ʹ໾ʹཱͬͯΔʁ 🤔 Vision and languageΛ΍͖ͬͯͯࢥͬͨ͜ͱ

Slide 13

Slide 13 text

ͳͥը૾ɾө૾͸໾ʹཱͨͳ͍ͷ͔ʁ AIに潜むバイアス

Slide 14

Slide 14 text

ද૚త૬ؔͷ໰୊ From: [Agrawal et al., “Don’t Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering,” CVPR 2018] [Otani et al., “Uncovering hidden challenges in query-based video moment retrieval,” BMVC 2020] VQAλεΫ ෦෼ө૾ݕࡧλεΫ

Slide 15

Slide 15 text

ද૚త૬ؔͷߏ଄ • ςΩετ ͕ಛఆͷग़ྗ (ྫ͑͹ ”No”) Λڧྗʹαϙʔτ • ͦ΋ͦ΋ը૾→ग़ྗͷֶश͕೉͍͠ʁ … … Is the ? … Vision and LanguageϞσϧ “No” ग़ྗ ը૾ ςΩετ Is the person wearing shorts? “No” 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 t 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 v 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 y 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 t σʔληοτ

Slide 16

Slide 16 text

ಉ͜͡ͱ͸ ”vision” ͷΈͷλεΫͰ΋ى͖Δ From: [Beery et al., “Recognition in Terra Incognita” ECCV 2018] ϥϕϧ લܠ എܠ

Slide 17

Slide 17 text

੔ཧ͢Δͱɾɾɾʁ • ཧ૝తʁͳσʔληοτ • ͱ ʹґଘؔ܎͕͋Ε͹ɺ ͱ ʹ΋ ग़ྗ ग़ྗʹ ؔ͢Δಛ௃ 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 y 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 x ग़ྗͱ͸ ௚઀ؔ܎ͳ͍ ಛ௃ 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 c ೖྗ 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 p(y, c) = p(y)p(c) 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 p(y|c) = p(y0|c) for c 2 C ⇢ C ͱ ʹ௚઀ͷґଘؔ܎͕ͳͯ͘΋ɺ Λհͨ͠ґଘؔ܎͕͋Δ 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 y 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 c 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 x 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 x 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 c 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 y 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 c

Slide 18

Slide 18 text

ґଘؔ܎ͷݪҼ From: [Simone et al., “A survey on bias in visual datasets,” CVIU 2022]

Slide 19

Slide 19 text

੔ཧ͢Δͱɾɾɾʁ • σʔληοτΛ࡞Δͱ͖ʹɺ ͱ ͕ಠཱʹ ͳΔΑ͏ʹ͕Μ͹Δ • ఢରతֶशͳͲʹΑͬͯɺ ͔Β ͕༧ଌͰ ͖ͳ͍Α͏ʹ͢Δ • Seth et al., “DeAR: Vision-language models with additive residuals,” CVPR 2023 • Berg et al., “A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning,” AACL 2022 • CounterfactualαϯϓϧͰ ͱ ͕ಠཱʹͳ ΔΑ͏ʹ͢Δ • Chen et al., “Counterfactual sample synthesizing for robust visual question answering,” CVPR 2020 ϥϕϧ ग़ྗΛද͢ ಛ௃ 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 y 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 x ग़ྗͱ͸ ௚઀ؔ܎ͳ͍ ಛ௃ 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 c ೖྗ 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 x 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 c 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 y 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 c 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 x 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 c

Slide 20

Slide 20 text

ද૚త૬ؔʹରͯ͠Կ͕Ͱ͖Δ͔ʁ • ϕϯνϚʔΫ༻ͷσʔληοτ • Agrawal et al., “Don’t assume; look and answer: Overcoming priors for visual question answering,” CVPR 2018 • Hendrycks and Dietterich, “Benchmarking neural network robustness to common corruption and perturbations,” ICLR 2019 • Hendrycks et al., “Natural adversarial examples,” CVPR 2021 • Out-of-distributionݕग़ • Hendrycks and Gimpel, “A baseline for detecting misclassification and out- of-distribution examples in neural networks,” ICLR 2017 • Hein et al., “Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem,” CVPR 2019 • ද૚త૬ؔΛ࣋ͭಛ௃ྔͷݕग़ • Wong et al., “Leveraging sparse linear layers for debuggable deep networks,” ICML 2021 • Anders et al., “Finding and removing Clever Hans: Using explanation methods to debug and improve deep models,” Information Fusion, Vol. 77, 2022 • Neuhaus et al., “Spurious features everywhere – Large-scale detection of harmful spurious features in ImageNet,” ICCV 2023

Slide 21

Slide 21 text

Vision and languageλεΫ͸Өڹ͞Ε΍͍͢ʁ • ͦ΋ͦ΋ϚϧνϞʔμϧͳωοτϫʔΫ͸ಛఆͷϞμϦςΟʹΦʔόʔ ϑΟοτ͠΍͍͢ • Wang et al., “What makes training multi-modal classification network hard?” CVPR 2020 • ͓ͦΒ͘Vision and Languageͷ৔߹͸ςΩετͷϞμϦςΟ • Shah et al., “The pitfalls of simplicity bias in neural networks,” NeurIPS 2020 • ಛʹɺը૾ͱςΩετΛೖྗͱ͢Δ৔߹ • ςΩετ͸཭ࢄతͳͷͰɺೖྗͱग़ྗͱͷ૬ؔΛݟ͚ͭ΍͍͢ • ը૾͸·ͣը૾தͷ֓೦͕ݟ͖͔͑ͯͯΒ • ςΩετ→ग़ྗͷ૬ؔΛֶशͨ͠Βͦ͜Ͱऩଋͨ͠Α͏ʹݟ͑Δ • ΋͔ͨ͠͠ΒάϩοΩϯάͷΑ͏ͳݱ৅͕ى͖Δ͔΋ • Power et al., “Grokking: Generalization beyond overfitting on small algorithmic datasets,” ICLR Workshop 2021

Slide 22

Slide 22 text

ࣾձతόΠΞε AIに潜むバイアス

Slide 23

Slide 23 text

Ϟσϧ͕࣋ͭࣾձతόΠΞε Image from: [Hendricks et al., “Women also snowboard: Overcoming bias in captioning models,” ECCV 2018].

Slide 24

Slide 24 text

ը૾Ωϟϓγϣχϯάͷߏ଄

Slide 25

Slide 25 text

ࣾձతόΠΞεͷධՁํ๏ • ֤ଐੑάϧʔϓ͝ͱͷੑೳ • λεΫ(Ωϟϓγϣχϯά)ͷੑೳ • [Burns et al., “Women also snowboard: Overcoming bias in captioning models,” ECCV 2018] • άϧʔϓݻ༗ͷ୯ޠ(e.g., ”man”)͕ਖ਼͘͠ग़ྗ͞Ε͔ͨ • [Burns et al., “Women also snowboard: Overcoming bias in captioning models,” ECCV 2018] • [Zhao et al., “Understanding and evaluating racial biases in image captioning,” ICCV 2021] • Ωϟϓγϣϯͱଐੑάϧʔϓͱͷؔ࿈ͷڧ͞ • ୯ޠͱଐੑάϧʔϓͷڞى: σʔληοτͷΩϟϓγϣϯͱੜ੒͞Ε ͨΩϟϓγϣϯͰڞى͕ͲΕ͚ͩҧ͏͔ • [Zhao et al., “Men also like shopping: Reducing gender bias amplification using corpus-level constraints,” EMNLP 2017]

Slide 26

Slide 26 text

·ͣ͸ධՁํ๏ • COCOσʔληοτ಺Ͱͷ୯ޠͱੑผͷڞى [*] Lin et al., “Microsoft COCO: Common Objects in Context”, Proc. ECCV 2014.

Slide 27

Slide 27 text

୯ޠͷڞىͰ͸ෆे෼ʁ A is next to a person in a fire suit A is in a fire suit

Slide 28

Slide 28 text

ࣝผث͕ग़ྗ͢ΔείΞ͸όΠΞεͱؔ࿈ͦ͠͏

Slide 29

Slide 29 text

δΣϯμʔόΠΞεͷධՁࢦඪ: LIC [Hirota et al., “Quantifying societal bias amplification in image captioning,” CVPR 2022].

Slide 30

Slide 30 text

ը૾ΩϟϓγϣχϯάͷόΠΞε௿ݮख๏ • Ξϊςʔγϣϯͳ͠Ͱ࢖͑Δํ๏ • Liu et al., “Show, deconfound and tell: Image captioning with causal inference,” CVPR 2022 • Ξϊςʔγϣϯ(ਓ෺ྖҬɾଐੑάϧʔϓ)͕ඞཁͳํ๏ • Hendricks et al., “Women also snowboard: Overcoming bias in captioning models,” ECCV 2018

Slide 31

Slide 31 text

طଘख๏ΛධՁͯ͠ΈΔ Equalizer [Hendricks et al., “Women also snowboard: Overcoming bias in captioning models,” ECCV 2018]

Slide 32

Slide 32 text

݁Ռ όΠΞεେ -*$ Ϟσϧͷग़ྗͷείΞͷฏۉ஋ rσʔληοτͷείΞͷฏۉ஋

Slide 33

Slide 33 text

࠶ܝը૾Ωϟϓγϣχϯάͷߏ଄

Slide 34

Slide 34 text

࠶ܝ: طଘख๏ΛධՁͯ͠ΈΔ Equalizer [Hendricks et al., “Women also snowboard: Overcoming bias in captioning models,” ECCV 2018] ඇଐੑྖҬˠଐੑهड़ͷόΠΞεͷΈʂ

Slide 35

Slide 35 text

Equalizer͸όΠΞεΛ૿෯͢Δ

Slide 36

Slide 36 text

όΠΞε௿ݮͷͨΊͷΞΠσΞ • ଐੑྖҬ→ඇଐੑهड़ɺඇଐੑྖҬ→ଐੑهड़ͷґଘؔ܎Λແ͍ͨ͘͠ • ʹରͯ͠ Λɺ ʹରͯ͠ ΛϥϯμϜԽ͢Ε͹͍͍ ଐੑ هड़ ඇଐੑ هड़ ଐੑ ྖҬ ඇଐੑ ྖҬ ը૾ ऩू AAACZHichVG7SgNBFD1ZXzG+oiIIgohBsQo34gsrwcYyD6OBKLK7jrq4L3YngRj8AW0VCysFEfEzbPwBC39AEMsINhbe3SyIinqHmTlz5p47Z2Y01zR8SfQYU1pa29o74p2Jru6e3r5k/8Ca71Q8XRR1x3S8kqb6wjRsUZSGNEXJ9YRqaaZY1/aXg/31qvB8w7FXZc0Vm5a6axs7hq5KpnJyK5miNIUx9hNkIpBCFFkneY0NbMOBjgosCNiQjE2o8LmVkQHBZW4TdeY8Rka4L3CIBGsrnCU4Q2V2n8ddXpUj1uZ1UNMP1TqfYnL3WDmGCXqgG2rQPd3SM73/Wqse1gi81HjWmlrhbvUdDRfe/lVZPEvsfar+9Cyxg4XQq8He3ZAJbqE39dWDs0ZhMT9Rn6RLemH/F/RId3wDu/qqX+VE/hwJ/oDM9+f+Cdam05m59GxuJrVE0VfEMYJxTPF7z2MJK8iiyOcKHOMEp7EnpVsZVIaaqUos0gziSyijH+wiiew= t AAACanichVG7SgNBFD1ZXzE+EmOjpBGjYhVuxBdWgo1lNMYIJsjuOuqSze6yOwnE4A/YWQmmUhARP8PGH7DIJ4h2EWwsvNksiIp6h5k5c+aeO2dmNMc0PEnUDCld3T29feH+yMDg0HA0NhLf9uyKq4ucbpu2u6OpnjANS+SkIU2x47hCLWumyGultfZ+vipcz7CtLVlzRLGsHlrGgaGrkql8QVPdujzZiyUpRX5M/ATpACQRRMaO3aCAfdjQUUEZAhYkYxMqPG67SIPgMFdEnTmXkeHvC5wgwtoKZwnOUJkt8XjIq92AtXjdrun5ap1PMbm7rJzAND3SLbXoge7oid5/rVX3a7S91HjWOlrh7EVPx7Jv/6rKPEscfar+9CxxgGXfq8HeHZ9p30Lv6KvH563syuZ0fYau6Jn9X1KT7vkGVvVVv94Qmw1E+APS35/7J9ieS6UXUwsb88lVCr4ijAQmMcvvvYRVrCODnO/uDBdohF6UuDKuJDqpSijQjOJLKFMf5zCMsQ== ¯ t AAACZHichVG7SgNBFD1ZXzFGjYogCCIGxSrciC+sBBtLY4wKGsLuOurivtidLMTgD2irWFgpiIifYeMPWPgDglhGsLHwZrMgGtQ7zMyZM/fcOTOjuabhS6KnmNLS2tbeEe9MdCW7e3pTff3rvlP2dFHQHdPxNjXVF6Zhi4I0pCk2XU+olmaKDe1gqb6/EQjPNxx7TVZcUbTUPdvYNXRVMpULSqk0ZSiM0WaQjUAaUaw4qRtsYwcOdJRhQcCGZGxChc9tC1kQXOaKqDLnMTLCfYEjJFhb5izBGSqzBzzu8WorYm1e12v6oVrnU0zuHitHMU6PdEs1eqA7eqGPX2tVwxp1LxWetYZWuKXe46H8+78qi2eJ/S/Vn54ldjEfejXYuxsy9VvoDX1weF7LL6yOVyfoil7Z/yU90T3fwA7e9OucWL1Agj8g+/O5m8H6VCY7m5nJTacXKfqKOIYxhkl+7zksYhkrKPC5Aic4xVnsWUkqA8pgI1WJRZoBfAtl5BPwIonu v AAACanichVG7SgNBFD1Z3/GRqI1iI0bFKtwVX1gFbCxNNCYQJeyuoy7ui93Ngi75ATsrQSsFEfEzbPwBCz9BtFOwsfBmsyAa1DvMzJkz99w5M6M6hu75RI8Jqa29o7OruyfZ29c/kEoPDm16ds3VRFGzDdstq4onDN0SRV/3DVF2XKGYqiFK6sFKY78UCNfTbWvDP3TEtqnsWfqurik+U6UtVXHDoF5NZyhLUYy3AjkGGcSxZqevsYUd2NBQgwkBCz5jAwo8bhXIIDjMbSNkzmWkR/sCdSRZW+MswRkKswc87vGqErMWrxs1vUit8SkGd5eV45iiB7qhV7qnW3qij19rhVGNhpdDntWmVjjV1PHI+vu/KpNnH/tfqj89+9jFUuRVZ+9OxDRuoTX1wdHp6/pyYSqcpkt6Zv8X9Eh3fAMreNOu8qJwjiR/gPzzuVvB5mxWXsjO5+cyOYq/ohtjmMAMv/cicljFGoqRuxOc4TzxIg1Jo9JYM1VKxJphfAtp8hPrMoyz ¯ v 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 p(t|v, ¯ v, ¯ t) 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 p(¯ t|v, ¯ v, t) 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 p(¯ t|v, ¯ v, t) = p(¯ t|¯ v, t) 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 p(t|v, ¯ v, ¯ t) = p(t|v, ¯ t) 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 t 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 ¯ t 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 v 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 ¯ v ࠓͷͱ͜Ζɺͪΐͬͱ೉͍͠ɾɾɾ

Slide 37

Slide 37 text

όΠΞε௿ݮͷͨΊͷΞΠσΞ ଓ͖ • Ұ౓ςΩετʹͯ͠͠·͑͹؆୯ • ଐੑهड़ˠඇଐੑهड़ ଐੑهड़ΛϥϯμϜʹม͑Δ • ඇଐੑهड़ˠଐੑهड़ ඇଐੑهड़ΛϥϯμϜʹม͑Δ • ඇଐੑهड़ͷϥϯμϜԽ͸ࣗ༝౓Λ͋Δఔ౓ߜΔ • Ϟσϧ͕ग़ྗ͢ΔΩϟϓγϣϯͷ෼෍͔Β͋·Γ཭Εͳ͍Α͏ʹɺݴޠϞσ ϧΛར༻ͯ͠ϥϯμϜԽͨ͠ΩϟϓγϣϯΛੜ੒ ଐੑ هड़ ඇଐੑ هड़ ଐੑ ྖҬ ඇଐੑ ྖҬ ը૾ ऩू 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 t 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 ¯ t AAACZHichVG7SgNBFD1ZXzFGjYogCCIGxSrciC+sBBtLY4wKGsLuOurivtidLMTgD2irWFgpiIifYeMPWPgDglhGsLHwZrMgGtQ7zMyZM/fcOTOjuabhS6KnmNLS2tbeEe9MdCW7e3pTff3rvlP2dFHQHdPxNjXVF6Zhi4I0pCk2XU+olmaKDe1gqb6/EQjPNxx7TVZcUbTUPdvYNXRVMpULSqk0ZSiM0WaQjUAaUaw4qRtsYwcOdJRhQcCGZGxChc9tC1kQXOaKqDLnMTLCfYEjJFhb5izBGSqzBzzu8WorYm1e12v6oVrnU0zuHitHMU6PdEs1eqA7eqGPX2tVwxp1LxWetYZWuKXe46H8+78qi2eJ/S/Vn54ldjEfejXYuxsy9VvoDX1weF7LL6yOVyfoil7Z/yU90T3fwA7e9OucWL1Agj8g+/O5m8H6VCY7m5nJTacXKfqKOIYxhkl+7zksYhkrKPC5Aic4xVnsWUkqA8pgI1WJRZoBfAtl5BPwIonu v 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 ¯ v ଐੑ هड़ ඇଐੑ هड़ AAACa3ichVHLSsNAFD2Nr1pf1W5EXYil6qrcii9cCW5cttZqQaUkcazBNAnJtFBLf8ClGxe6URARP8ONP+DCTxBxVcGNC2/TgGhRb5jMmTP33DlzR3NMw5NETyGlo7OruyfcG+nrHxgcig6PbHl22dVFTrdN281rqidMwxI5aUhT5B1XqCXNFNva0Vpzf7siXM+wrU1ZdcReSS1axoGhq5Kp/K6mujVZnylE45QkPybbQSoAcQSRtqM32MU+bOgoowQBC5KxCRUefztIgeAwt4cacy4jw98XqCPC2jJnCc5QmT3if5FXOwFr8bpZ0/PVOp9i8nBZOYkEPdItNeiB7uiZPn6tVfNrNL1UedZaWuEUhk5Gs+//qko8Sxx+qf70LHGAZd+rwd4dn2neQm/pK8dnjezKRqI2TVf0wv4v6Ynu+QZW5U2/zoiNc0T4AVI/290OtuaSqcXkQmY+vkrBU4QxjinMcr+XsIp1pJHz+3yKc1yEXpWYMqZMtFKVUKCJ4VsoiU9dqIzi ¯ t0 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 t0 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 p(t0|t, ¯ t, ¯ t0) AAACf3ichVG7SgNBFD1Z3/GRqI1oEwwxChJuxHcl2Fj6SiKohN11oks2u8vuJKBRsPYHLKwUVMRC/8HGH7DwE8Qygo2FN+uCqKh3mJkzZ+65c2ZGc0zDk0SPIaWhsam5pbUt3N7R2RWJdvdkPbvs6iKj26btrmmqJ0zDEhlpSFOsOa5QS5opclpxvr6fqwjXM2xrVe46YrOkbltGwdBVyVQ+2u8Mb2iqW5UHyX05GgvwaEwmR/LROKXIj9hPkA5AHEEs2tFLbGALNnSUUYKABcnYhAqP2zrSIDjMbaLKnMvI8PcFDhBmbZmzBGeozBZ53ObVesBavK7X9Hy1zqeY3F1WxpCgB7qiGt3TNT3R26+1qn6NupddnrUPrXDykaO+ldd/VSWeJXY+VX96lihg2vdqsHfHZ+q30D/0lb3j2srscqI6RGf0zP5P6ZHu+AZW5UU/XxLLJwjzB6S/P/dPkB1LpSdTE0vj8TkKvqIVAxjEML/3FOawgEVk+NxDXOAGt0pISSopJchVQoGmF19CmXkHkFuSqQ== p(¯ t0|t, ¯ t, t0) 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 p(¯ t0|t, ¯ t, t0) = p(¯ t0|¯ t, t0) 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 p(t0|t, ¯ t, ¯ t0) = p(t0|t, ¯ t0)

Slide 38

Slide 38 text

ΑΓόΠΞε͕ڧ͍ΩϟϓγϣϯΛ໭͢ [Hirota et al., “Model-agnostic gender debiased image captioning,” CVPR 2023].

Slide 39

Slide 39 text

݁Ռ • ඇଐੑྖҬˠଐੑهड़͕ड͚ΔӨڹˣ • ଐੑྖҬˠඇଐੑهड़͕ड͚ΔӨڹˣ • Ωϟϓγϣϯͷ඼࣭͸΄ͱΜͲมΘΒͳ ͍͔ɺ্͕Δ ඇଐੑྖҬˠଐੑهड़ ଐੑྖҬˠඇଐੑهड़

Slide 40

Slide 40 text

ը૾ੜ੒Ϟσϧ΋όΠΞε From [Bianchi et al., “Easily accessible text-to-image generation amplifies demographic stereotypes at large scale,” FaaCT 2023].

Slide 41

Slide 41 text

όΠΞεͷධՁʹ࢖͑Δσʔληοτ [Garcia et al., “Uncurated image-text datasets: Shedding light on demographic bias,” CVPR 2023].

Slide 42

Slide 42 text

ը૾ੜ੒ϞσϧͷόΠΞεͷධՁ • Stable Diffusionͷੑผ-ΦϒδΣΫτؒͷόΠΞεΛධՁ [Wu et al., “Stable Diffusion exposed: Gender bias from prompt to image,” arXiv:2312.03027, 2023].

Slide 43

Slide 43 text

ະདྷͷσʔλͰֶश͞ΕͨϞσϧ͸Ͳ͏ͳΔʁ 0% 20% 100% ... ... Train model Evaluate gender bias skin tone bias age bias ethnicity bias Train model Evaluate gender bias skin tone bias age bias ethnicity bias Train model Evaluate gender bias skin tone bias age bias ethnicity bias 40% Train model Evaluate gender bias skin tone bias age bias ethnicity bias generated images real images

Slide 44

Slide 44 text

݁Ռ • CC3MʢҰ෦Λੜ੒ը૾ʹஔ͖׵͑ʣͰ܇࿅ͨ͠OpenCLIPͰɺText-to- Imageݕࡧ • إը૾ͷOpenCLIPۭؒͰͷʮ·ͱ·Γʯ౓߹͍

Slide 45

Slide 45 text

Take-home • Vision and LanguageλεΫͰ͸ɺ࣮͸Vision͕͋Μ·Γ໾ʹཱͨͳ͍ʁ • σʔληοτʹόΠΞε(ަབྷҼࢠ)͕͋ΔՄೳੑΛҙࣝ • Out-of-distributionͰͷධՁͷॏཁੑ • ʢओʹʣVision and LanguageλεΫͷόΠΞεͷ໰୊Λ၆ᛌ • όΠΞεʢද૚త૬ؔʣͱ͸ͳʹ͔ • ͳͥόΠΞε͕ى͖Δͷ͔ • Ωϟϓγϣϯੜ੒ʹ͓͚ΔࣾձతόΠΞεͷ໰୊ • σʔληοτͷόΠΞεͷͻͱͭ • Ωϟϓγϣϯੜ੒Ͱ͸ɺґଘؔ܎͕গ͠ෳࡶ • ίϯςΩετ͔Βੑผɺੑผ͔ΒίϯςΩετͷόΠΞεΛଧͪফ͢Α͏ͳ ֶशͰɺ໰୊Λ௿ݮ

Slide 46

Slide 46 text

ߟ͑Δ΂͖ϙΠϯτ • ࣮ੈքͷ෼෍͕ͦ͏ͳͷͰ͋Ε͹ɺͦͷ··Ͱ͍͍ͷͰ͸ʁ • ʢIn-distributionͷʣσʔληοτͳΒਫ਼౓ߴ͍͠ɾɾɾ vs • ֶशͰϞσϧͷόΠΞε͸૿෯͞ΕΔ͔΋͠Εͳ͍ • ը૾ɾςΩετͷҙຯཧղͷݚڀɺͦΕͰ͍͍ͷʁ ຊ౰ʹόΠΞεΛऔΓআ͘౒ྗΛ͢Δ΂͖ʁ

Slide 47

Slide 47 text

ँࣙ • ຊߨԋ͸ҎԼͷڞஶऀͷօ༷ͷଟେͳΔ͝ॿྗʹΑΔ΋ͷͰ͢ɻօ༷ʹ ײँਃ্͛͠·͢ɻ • ຊߨԋ͸ҎԼͷڝ૪తࢿۚͷαϙʔτʹΑΓ·͢ɻ • ө૾هड़ͷͨΊͷݴޠΛ૑ग़͢Δਓ޻஌ೳͷ࣮ݱɺ(୅ද)தౡɺ૑ൃతݚڀ ࢧԉࣄۀɺ2021೥౓࠾୒ • ࢹ֮తύλʔϯʹΑΔը૾هड़Λ༻͍ͨσΟʔϓχϡʔϥϧωοτͷόΠΞ ε௿ݮɺՊݚඅج൫(A)ɺ2023-2027 • ΠϯϑΥσϛοΫΛࠀ෰͢Διʔγϟϧ৘ใج൫ٕज़ɺCRESTʮ৴པ͞ΕΔ AIγεςϜΛࢧ͑Δج൫ٕज़ʯɺ2020-2024 େ୩·Ώ Noa Garcia ኍా༟྄ Yankun Wu Tianwei Chen

Slide 48

Slide 48 text

No content