Upgrade to Pro — share decks privately, control downloads, hide ads and more …

AI最新論文読み会2021年11月

M.Inomata
November 10, 2021

 AI最新論文読み会2021年11月

AI最新論文読み会2021年11月の発表資料です。

https://deeplearning-b.connpass.com/event/227234/

M.Inomata

November 10, 2021
Tweet

More Decks by M.Inomata

Other Decks in Science

Transcript

  1. AI࠷৽࿦จಡΈձ2021೥11݄
    ᷂tech vein ழມ ॆԝ

    View full-size slide

  2. ࣗݾ঺հ
    ழມ ॆԝ (͍ͷ·ͨ ΈͭͻΖ)


    גࣜձࣾ tech vein / DeepRad גࣜձࣾ


    ֤୅දऔక໾ ݉ σϕϩούʔ


    twitter: @ino2222

    View full-size slide

  3. Facebook άϧʔϓͷ঺հ
    IUUQTXXXGBDFCPPLDPNHSPVQT

    View full-size slide

  4. ΞδΣϯμ
    Archive Sanity (arxiv-sanity.com) ͔ΒϐοΫΞο
    ϓͨ͠ɺarxiv.org ͷաڈ1ϲ݄ؒͷ࿦จ঺հɻ


    ɾҰ൪ؾʹͳͬͨ࿦จͷ঺հ


    ɾtop recentͷ࿦จτοϓ10 Ϧετ


    ɾtop hype ͷ࿦จτοϓ10 Ϧετ


    View full-size slide

  5. Archive Sanity?
    https://www.arxiv-sanity.com/top

    View full-size slide

  6. Top10 Recent
    1. ResNet strikes back: An improved training procedure in timm


    2. Exploring the Limits of Large Scale Pre-training


    3. Deep Neural Networks and Tabular Data: A Survey


    4. Learning in High Dimension Always Amounts to Extrapolation


    5. ADOP: Approximate Differentiable One-Pixel Point Rendering


    6. Well-classi
    fi
    ed Examples are Underestimated in Classi
    fi
    cation with
    Deep Neural Networks


    7. ByteTrack: Multi-Object Tracking by Associating Every Detection Box


    8. MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision
    Transformer ← PickUp!


    9. Fast Model Editing at Scale


    10. Self-supervised Learning is More Robust to Dataset Imbalance

    View full-size slide

  7. Top10 Hype
    1. ADOP: Approximate Differentiable One-Pixel Point Rendering


    2. Real numbers, data science and chaos: How to
    fi
    t any dataset with a
    single parameter


    3. Delphi: Towards Machine Ethics and Norms


    4. Multitask Prompted Training Enables Zero-Shot Task Generalization


    5. Nonnegative spatial factorization


    6. Learning in High Dimension Always Amounts to Extrapolation


    7. StyleAlign: Analysis and Applications of Aligned StyleGAN Models


    8. Deep Learning Tools for Audacity: Helping Researchers Expand the
    Artist's Toolkit


    9. ECQx: Explainability-Driven Quantization for Low-Bit and Sparse
    DNNs


    10. Exploring the Limits of Large Scale Pre-training

    View full-size slide

  8. Top10 Recent (྘: CNN, ੺: Transformer)
    1. ResNet strikes back: An improved training procedure in timm


    2. Exploring the Limits of Large Scale Pre-training


    3. Deep Neural Networks and Tabular Data: A Survey


    4. Learning in High Dimension Always Amounts to Extrapolation


    5. ADOP: Approximate Differentiable One-Pixel Point Rendering


    6. Well-classi
    fi
    ed Examples are Underestimated in Classi
    fi
    cation with
    Deep Neural Networks


    7. ByteTrack: Multi-Object Tracking by Associating Every Detection Box


    8. MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision
    Transformer ← PickUp!


    9. Fast Model Editing at Scale


    10. Self-supervised Learning is More Robust to Dataset Imbalance

    View full-size slide

  9. Top10 Hype (྘: CNN, ੺: Transformer)
    2. Real numbers, data science and chaos: How to
    fi
    t any dataset with a
    single parameter


    3. Delphi: Towards Machine Ethics and Norms


    4. Multitask Prompted Training Enables Zero-Shot Task Generalization


    5. Nonnegative spatial factorization


    7. StyleAlign: Analysis and Applications of Aligned StyleGAN Models


    8. Deep Learning Tools for Audacity: Helping Researchers Expand the
    Artist's Toolkit


    9. ECQx: Explainability-Driven Quantization for Low-Bit and Sparse
    DNNs


    View full-size slide

  10. Pickup࿦จ

    View full-size slide

  11. 8. MobileViTɿܰྔɾ൚༻ɾϞόΠϧରԠͷϏδϣϯτϥϯεϑΥʔϚʔ


    (ݪจ: MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision


    Transformer)
    ܰྔͷ৞ΈࠐΈχϡʔϥϧωοτϫʔΫʢCNNʣ͸ɺϞόΠϧɾϏδϣϯɾλεΫͷσϑΝΫτͱͳ͍ͬͯΔɻͦͷۭ
    ؒ༠ಋόΠΞεʹΑΓɺҟͳΔࢹ֮λεΫؒͰΑΓগͳ͍ύϥϝʔλͰදݱΛֶश͢Δ͜ͱ͕Ͱ͖Δɻ͔͠͠ɺ͜ΕΒͷ
    ωοτϫʔΫ͸ۭؒతʹϩʔΧϧͳ΋ͷͰ͋ΔɻάϩʔόϧͳදݱΛֶश͢ΔͨΊʹɺࣗݾ஫ҙϕʔεͷϏδϣϯτϥϯ
    εϑΥʔϚʔʢViTʣ͕࠾༻͞Ε͍ͯΔɻCNNͱ͸ҟͳΓɺViTs͸ϔϏʔ΢ΣΠτͰ͋ΔɻຊߘͰ͸ɺCNNͱViTͷ௕ॴ
    Λ૊Έ߹ΘͤͯɺϞόΠϧɾϏδϣϯɾλεΫͷͨΊͷܰྔͰ௿஗ԆͷωοτϫʔΫΛߏங͢Δ͜ͱ͸Մೳ͔ɺͱ͍͏ٙ
    ໰Λ౤͔͚͍͛ͯΔɻ͜ͷ໨తͷͨΊʹɺϞόΠϧσόΠε༻ͷܰྔͰ൚༻తͳϏδϣϯม׵ثͰ͋ΔMobileViTΛ঺հ
    ͠·͢ɻMobileViT͸ɺม׵ثΛ༻͍ͨ৘ใͷάϩʔόϧॲཧͷͨΊͷҟͳΔࢹ఺ɺ͢ͳΘͪɺ৞ΈࠐΈͱͯ͠ͷม׵ث
    Λఏࣔ͠·͢ɻͦͷ݁ՌɺMobileViT͸ɺ͞·͟·ͳλεΫ΍σʔληοτʹ͓͍ͯɺCNN΍ViTϕʔεͷωοτϫʔΫ
    Λେ෯ʹ্ճΔ͜ͱ͕Θ͔Γ·ͨ͠ɻImageNet-1kσʔληοτͰ͸ɼMobileViT͸໿600ສݸͷύϥϝʔλͰ78.4%ͷ
    top-1ਫ਼౓Λୡ੒ͨ͠ɽ͜Ε͸ɼಉఔ౓ͷύϥϝʔλ਺ͷMobileNetv3ʢCNNϕʔεʣ͓ΑͼDeITʢViTϕʔεʣΑΓ΋
    3.2%͓Αͼ6.2%ߴ͍ਫ਼౓Ͱ͋ΔɽMS-COCOͷ෺ମݕग़λεΫͰ͸ɺMobileViT͸ಉ਺ͷύϥϝʔλͰMobileNetv3Α
    Γ΋5.7%ਫ਼౓͕ߴ͍ɻ
    http://arxiv.org/abs/2110.02178v1
    w ໨తɾ੒Ռɿ7J5ΛܰྔԽͨ͠৽Ϟσϧ.PCJMF7J5ͷ։ൃ
    w ํ๏ɿม׵ثΛ৞ΈࠐΈͱͯ͠࢖͍ɺάϩʔόϧͳදݱΛֶशͨ͠
    w ݻ༗໊ɿ.PCJMF7J5 IUUQTHJUIVCDPNBQQMFNMDWOFUT

    w ஶऀॴଐɿ"QQMF

    View full-size slide

  12. ࢀߟ: MobileNetV2

    MobileVit͸MobileNetV2 + Transformer
    IUUQTXXXSFTFBSDIHBUFOFU
    fi
    HVSF5IFBSDIJUFDUVSFPGUIF.PCJMF/FUWOFUXPSL@
    fi
    H@

    View full-size slide

  13. CNN+ViTͰશମΛΤϯίʔυ͢Δɻ


    CNNͰ੺ɾ੨υοτͷपΓΛ৞ΈࠐΜͩ͋ͱɺ


    TransformerͰ੺υοτ͔Β੨υοτΛݟΔ͜ͱͰɺ݁ՌతʹશମΛݟ͍ͯΔ

    View full-size slide

  14. MobileNetv3, DeITΛ্ճΔੑೳͰܰྔ

    View full-size slide

  15. ϞόΠϧ୺຤Ͱ΋ಈ͔͠΍͍͢ߴਫ਼౓ͳϞσϧ

    ʢͨͩ͠ɺMobileNetv2ΑΓ͸஗͍ʣ
    • Mobile Friendly: ܰྔɺ൚༻తɺ(ൺֱత)௿ϨΠςϯγʔͰ࢖͍΍͢
    ͦ͏ɻ


    • ͨͩ͠ϞόΠϧػث(=iPhone12)Ͱ͸MobileNetv2ΑΓਪ࿦͕8ഒ஗
    ͔ͬͨɻ

    →ϞόΠϧGPUʹ͸CUDAΧʔωϧ͕ແ͍ͷͱɺCNN͸৞ΈࠐΈͦ
    ͏ͱͷҰׅਖ਼نԽ༥߹ͳͲͷ࠷దԽ͕͋Δͷ͕ཧ༝ͱͷ͜ͱ

    View full-size slide

  16. GitHub


    https://github.com/apple/ml-cvnets

    View full-size slide

  17. Top recent: Best10

    View full-size slide

  18. 1. ResNetͷٯऻɻtimmͰͷֶशखॱͷվળ


    (ݪจ: ResNet strikes back: An improved training procedure in timm)
    HeΒʹΑͬͯઃܭ͞ΕͨӨڹྗͷ͋ΔResidual Networks͸ɼଟ͘ͷՊֶ࿦จͰۚࣈౝతͳΞʔΩςΫ
    νϟͱͯ͠औΓ্͛ΒΕ͍ͯ·͢ɽ͜ΕΒͷΞʔΩςΫνϟ͸௨ৗɺݚڀʹ͓͚ΔσϑΥϧτͷΞʔΩ
    ςΫνϟͱͯ͠ɺ͋Δ͍͸৽͍͠ΞʔΩςΫνϟ͕ఏҊ͞ΕͨࡍͷϕʔεϥΠϯͱͯ͠ػೳ͍ͯ͠·
    ͢ɻ͔͠͠ɺ2015೥ʹResNetΞʔΩςΫνϟ͕ൃද͞ΕͯҎདྷɺχϡʔϥϧωοτϫʔΫͷτϨʔχ
    ϯάͷϕετϓϥΫςΟεʹ͍ͭͯେ͖ͳਐల͕͋Γ·ͨ͠ɻ৽ͨͳ࠷దԽˍσʔλΦʔάϝϯςʔ
    γϣϯʹΑΓɺτϨʔχϯάϨγϐͷ༗ޮੑ͕ߴ·͍ͬͯ·͢ɻຊ࿦จͰ͸ɺ͜ͷΑ͏ͳਐาΛ౷߹͠
    ͨखॱͰτϨʔχϯάͨ͠৔߹ͷόχϥResNet-50ͷੑೳΛ࠶ධՁ͠·͢ɻզʑ͸ɺڝ૪ྗͷ͋Δֶश
    ઃఆͱࣄલʹֶश͞ΕͨϞσϧΛtimmΦʔϓϯιʔεϥΠϒϥϦͰڞ༗͠ɺকདྷͷݚڀͷͨΊͷΑΓྑ
    ͍ϕʔεϥΠϯͱͯ͠໾ཱͭ͜ͱΛظ଴͍ͯ͠·͢ɻྫ͑͹ɺզʑͷΑΓݫֶ͍͠शઃఆͰ͸ɺόχϥ
    ͷResNet-50͸ɺ௥Ճσʔλ΍ৠཹͳ͠ͰImageNet-valͷղ૾౓224x224Ͱ80.4%ͷτοϓ1ਫ਼౓Λ
    ୡ੒͍ͯ͠·͢ɻ·ͨɺҰൠతͳϞσϧʹ͍ͭͯɺզʑͷֶशํ๏ͰಘΒΕͨੑೳΛใࠂ͠·͢ɻ
    w ໨తɿΞʔΩςΫνϟͷมߋͳ͠ʹɺ3FT/FUͷ࠷ྑͷֶशखॱΛఏڙ͢Δ
    w ੒Ռɿ1Z5PSDI༻ͷUJNNϥΠϒϥϦͰϞσϧઃఆͱࣄલֶशࡁΈϞσϧΛఏڙ
    w ํ๏ɿϋΠύʔύϥϝʔλௐ੔ ΫϩεΤϯτϩϐʔଛࣦ͔Βͷ୤٫
    w ݻ༗໊ɿͳ͠
    w ஶऀॴଐɿ'BDFCPPL"*3FTFBSDI
    http://arxiv.org/abs/2110.00476v1

    View full-size slide

  19. A1~A3: ֶशίετͷҧ͍

    View full-size slide

  20. 2. େن໛ͳࣄલτϨʔχϯάͷݶքΛ୳Δ


    (ݪจ: Exploring the Limits of Large Scale Pre-training)
    ۙ೥ͷେن໛ػցֶशͷൃల͸ɺσʔλɺϞσϧαΠζɺֶश࣌ؒΛద੾ʹεέʔϧΞοϓ͢Δ͜ͱͰɺ
    ࣄલֶशͷվળ͕΄ͱΜͲͷԼྲྀλεΫʹ༗རʹҠߦ͢Δ͜ͱΛ؍࡯͢Δ͜ͱ͕Ͱ͖Δ͜ͱΛࣔࠦͯ͠
    ͍ΔɻຊݚڀͰ͸ɺ͜ͷݱ৅Λܥ౷తʹݚڀ͠ɺ্ྲྀͷਫ਼౓Λ্͛ΔͱԼྲྀͷλεΫͷੑೳ͕๞࿨͢Δ
    ͜ͱΛূ໌͠·ͨ͠ɻ۩ମతʹ͸ɺVision TransformersɺMLP-MixerɺResNetsʹ͍ͭͯɺύϥϝʔλ
    ਺͕1000ສ͔Β100ԯͷൣғͰ4800ճҎ্ͷ࣮ݧΛߦ͍ɺ࠷େن໛ͷը૾σʔλʢJFTɺ
    ImageNet21KʣͰֶश͠ɺ20Ҏ্ͷԼྲྀͷը૾ೝࣝλεΫͰධՁͨ͠ɻͦͷ݁Ռɺ๞࿨ݱ৅Λ൓ө
    ͠ɺ্ྲྀͱԼྲྀͷੑೳͷඇઢܗؔ܎Λଊ͑ͨԼྲྀੑೳͷϞσϧΛఏҊͨ͠ɻ͞Βʹɺ͜ͷΑ͏ͳݱ৅͕
    ൃੜ͢Δཧ༝Λ۷ΓԼ͛ͯཧղ͢ΔͨΊʹɺࢲ͕ͨͪ؍࡯ͨ͠๞࿨ݱ৅͸ɺϞσϧͷ૚Λ௨ͯ͠දݱ͕
    ਐԽ͢Δํ๏ͱີ઀ʹؔ܎͍ͯ͠Δ͜ͱΛࣔ͠·ͨ͠ɻ·ͨɺ͞Βʹۃ୺ͳྫͱͯ͠ɺΞοϓετϦʔ
    Ϝͱμ΢ϯετϦʔϜͷύϑΥʔϚϯε͕૬൓͢Δ৔߹Λ঺հ͠·͢ɻͭ·Γɺμ΢ϯετϦʔϜͷੑ
    ೳΛ޲্ͤ͞ΔͨΊʹ͸ɺΞοϓετϦʔϜͷਫ਼౓Λམͱ͢ඞཁ͕͋Δͱ͍͏͜ͱͰ͢ɻ
    http://arxiv.org/abs/2110.02095v1
    w ໨తɿը૾ೝࣝϞσϧͷGFXTIPUֶशʹ͍ͭͯͷܥ౷తͳݚڀ
    w ੒ՌɿGFXTIPUֶश༻ͷɺ൚༻తͳࣄલֶशϞσϧͷઃܭํ๏ͷఏࣔ
    w ํ๏ɿ7J5 .-1.JYFS 3FT/FUʹ͍ͭͯɺ6Q4USFBN%PXO4USFBNֶश࣌ͷ๞࿨ݱ৅ͷ૬ؔΛ

    λεΫผɾύϥϝʔλผɾϞσϧαΠζผʹௐ΂Δ
    w ݻ༗໊ɿͳ͠
    w ஶऀॴଐɿ(PPHMF3FTFBSDI

    View full-size slide

  21. ࣄલֶश(্ྲྀ)ͷਫ਼౓ vs సҠֶश(Լྲྀ)ͷ 1shot/25shot ਫ਼౓


    λεΫʹΑͬͯ͸্ྲྀͷਫ਼౓͕ߴ͍΄ͲԼྲྀͷਫ਼౓͕๞࿨͠΍͍͢(acc:0.2~0.5Ͱ΋ఀ଺)


    →্ྲྀλεΫͱԼྲྀλεΫͷؔ܎ੑ͕େࣄ

    View full-size slide

  22. ཁ໿
    • ൚༻తͳfew shot ֶशλεΫ޲͚ͷࣄલֶशϞσϧΛ࡞Δ
    ͱ͖ͷઃܭࢦඪɾվળࡦΛఏ͍ࣔͯ͠Δɻ


    • ࣄલֶशͷਫ਼౓Λߴ͗͘͢͠ͳ͍ɾֶश཰Λམͱ͢


    • ୯७ʹεέʔϧ͢Δ͜ͱ͕͢΂ͯΛղܾ͢Δͱ͸ݶΒͳ
    ͍


    • ̍ͭͷԼྲྀλεΫͰಛԽ͢ΔͷͰ͸ͳ͘ɺ෯޿͍Լྲྀͷ
    λεΫͰύϑΥʔϚϯεΛ޲্ͤ͞ΔΑ͏ͳઃܭ্ͷબ
    ୒Λ͢΂͖

    View full-size slide

  23. 3. σΟʔϓɾχϡʔϥϧɾωοτϫʔΫͱදܗࣜσʔλɻαʔϕΠ


    (ݪจ: Deep Neural Networks and Tabular Data: A Survey)
    ෆۉ࣭ͳදܗࣜͷσʔλ͸ɺ࠷΋Ұൠతʹ࢖༻͞Ε͍ͯΔσʔλܗࣜͰ͋Γɺଟ͘ͷॏཁ͔ͭܭࢉෛ
    ՙͷߴ͍ΞϓϦέʔγϣϯʹ͸͔ܽͤ·ͤΜɻಉछͷσʔληοτͰ͸ɺσΟʔϓχϡʔϥϧωοτ
    ϫʔΫ͕܁Γฦ͠༏ΕͨੑೳΛ͓ࣔͯ͠ΓɺͦͷͨΊ޿͘࠾༻͞Ε͍ͯ·͢ɻ͔͠͠ɺදܗࣜσʔλ
    ͷϞσϦϯάʢਪ࿦·ͨ͸ੜ੒ʣ΁ͷద༻͸ɺґવͱͯ͠ඇৗʹࠔ೉Ͱ͢ɻຊݚڀͰ͸ɺදܗࣜσʔ
    λʹର͢Δ࠷ઌ୺ͷਂ૚ֶशख๏ͷ֓ཁΛઆ໌͢Δɻ·ͣɺͦΕΒΛʮσʔλม׵ʯʮಛघͳΞʔΩ
    ςΫνϟʯʮਖ਼ଇԽϞσϧʯͷ3ͭͷάϧʔϓʹ෼ྨ͠·͢ɻͦͷޙɺ֤άϧʔϓͷओཁͳΞϓϩʔν
    ͷแׅతͳ֓ཁΛఏڙ͠·͢ɻλϏϡϥʔσʔλΛੜ੒͢ΔͨΊͷਂ૚ֶशΞϓϩʔνͷٞ࿦͸ɺλ
    Ϗϡϥʔσʔλ্Ͱਂ૚ϞσϧΛઆ໌͢ΔͨΊͷઓུʹΑͬͯิ׬͞ΕΔɻࢲͨͪͷओͳߩݙ͸ɺ͜
    ͷ෼໺ͷओͳݚڀͷྲྀΕͱطଘͷํ๏࿦ΛऔΓ্͛ɺؔ࿈͢Δ՝୊΍ະղܾͷݚڀ՝୊Λ໌Β͔ʹ͢
    Δ͜ͱͰ͢ɻզʑͷ஌ΔݶΓͰ͸ɺ͜Ε͸λϏϡϥʔσʔλʹର͢Δਂ૚ֶशͷΞϓϩʔνΛৄࡉʹ
    ݕ౼ͨ͠ॳΊͯͷ΋ͷͰ͢ɻຊ࿦จ͸ɺදܗࣜσʔλΛ༻͍ͨਂ૚ֶशʹڵຯΛ࣋ͭݚڀऀ΍࣮຿Ո
    ʹͱͬͯɺوॏͳग़ൃ఺ͱͳΓɺࢦ਑ͱͳΔͰ͠ΐ͏ɻ
    http://arxiv.org/abs/2110.01889v1
    w ໨తɿදܗࣜσʔλʹର͢Δ࠷ઌ୺ͷਂ૚ֶशख๏ͷ֓ཁΛઆ໌͢ΔϨϏϡʔ࿦จ
    w ੒ՌɿओͳݚڀͷྲྀΕͱطଘͷํ๏࿦ɾϞσϧ౳Λ·ͱΊͯɺະղܾͷݚڀ՝୊Λ໌Β͔ʹͨ͠
    w ํ๏ɿදܗࣜͷσʔληοτΛάϧʔϓ෼ྨͯ͠ɺάϧʔϓ͝ͱͷΞϓϩʔνͷ֓ཁΛ·ͱΊͨ
    w ݻ༗໊ɿͳ͠
    w ஶऀॴଐɿςϡʔϏϯήϯେֶ υΠπ
    4$)6'")PMEJOH"( υΠπ

    View full-size slide

  24. 4. ߴ࣍ݩͰͷֶͼ͸ɺৗʹ֎Ԇతͳ΋ͷͰ͋Δɻ


    (ݪจ: Learning in High Dimension Always Amounts to Extrapolation)
    ิؒͱ֎ૠͷ֓೦͸ɼਂ૚ֶश͔Βؔ਺ۙࣅ·Ͱ༷ʑͳ෼໺ͰجຊͱͳΔɽิؒ͸ɼ͋
    Δαϯϓϧx͕ɼ༩͑ΒΕͨσʔληοτͷತแͷ಺ଆ·ͨ͸ڥք্ʹ͋Δͱ͖ʹߦΘ
    ΕΔɽ֎ૠ͸ɼx͕ͦͷತแͷ֎ଆʹ͋Δͱ͖ʹߦΘΕΔɽ1ͭͷجຊతͳʢޡͬͨʣೝ
    ࣝ͸ɺ࠷ઌ୺ͷΞϧΰϦζϜ͕͏·͘ػೳ͢Δͷ͸ɺֶशσʔλΛਖ਼͘͠ิؒ͢Δೳྗ
    ͕͋Δ͔Βͩͱ͍͏΋ͷͰ͋Δɻ2ͭ໨ͷޡղ͸ɺิؒ͸λεΫ΍σʔληοτશମͰ
    ߦΘΕΔͱ͍͏΋ͷͰɺ࣮ࡍɺଟ͘ͷ௚؍΍ཧ࿦͕͜ͷԾఆʹґଘ͍ͯ͠Δɻզʑ͸ܦ
    ݧతɺཧ࿦తʹ͜ΕΒͷ2ͭͷ఺ʹ൓࿦͠ɺͲΜͳߴ࣍ݩʢ>100ʣͷσʔληοτͰ
    ΋ɺ΄ͱΜͲ࣮֬ʹิؒ͸ى͜Βͳ͍͜ͱΛ࣮ূͨ͠ɻ͜ΕΒͷ݁Ռ͸ɺҰൠԽੑೳͷ
    ࢦඪͱͯ͠ͷݱࡏͷิؒ/֎ૠͷఆٛͷଥ౰ੑʹٙ໰Λ౤͔͚͛Δ΋ͷͰ͋Δɻ
    http://arxiv.org/abs/2110.09485v1
    w ໨తɿߴ࣍ݩۭؒ Ҏ্
    Ͱ͸σʔληοτͰิ͕ؒى͜Βͳ͍ࣄΛཧ࿦తɾܦݧతʹ࣮ূ͢Δɻ
    w ੒Ռɿ৽͍͠αϯϓϧʹର͢ΔิؒΛҡ࣋͢ΔͨΊʹ͸ɺσʔληοτͷαΠζ͕σʔλͷ࣍ݩʹର͠
    ͯࢦ਺ؔ਺తʹେ͖͘ͳΔ͜ͱΛ࣮ূ͠ɺطଘͷࢦඪΛ൱ఆͨ͠ɻ
    w ํ๏ɿ࣍ݩ͕૿͑ͨ࣌ʹิؒ͢ΔͨΊͷσʔληοτྔΛཧ࿦ɾ࣮σʔληοτͷ྆໘Ͱܭࢉ
    w ݻ༗໊ɿ
    w ஶऀॴଐɿ'BDFCPPL"*3FTFBSDI

    View full-size slide

  25. ࢀߟ: ֎ૠɾ಺ૠ
    IUUQTBUNBSLJUJUNFEJBDPKQBJUBSUJDMFTOFXTIUNM

    View full-size slide

  26. σʔλͷ࣍ݩ͕૿͑Δఔɺ༻ҙͨ͠σʔλ
    ηοτͰิؒͰ͖Δׂ߹͕ࢦ਺తʹݮΔ

    View full-size slide

  27. 5. ADOP: ۙࣅࠩҟԽ1ϐΫηϧɾϙΠϯτɾϨϯμϦϯά


    ݪจ: ADOP: Approximate Differentiable One-Pixel Point Rendering)
    ຊݚڀͰ͸ɺγʔϯͷਫ਼ີԽͱ৽͍͠Ϗϡʔͷ߹੒ͷͨΊͷɺ৽͍͠ϙΠϯτϕʔεͷඍ෼ՄೳͳχϡʔϥϧϨϯμ
    ϦϯάύΠϓϥΠϯΛ঺հ͠·͢ɻೖྗ͸ɺ఺܈ͷॳظਪఆ஋ͱΧϝϥͷύϥϝʔλͰ͢ɻग़ྗ͸ɺ೚ҙͷΧϝϥ
    ϙʔζ͔Β߹੒͞Εͨը૾Ͱ͢ɻ఺܈ͷϨϯμϦϯά͸ɺඍ෼ՄೳͳϨϯμϥʔʹΑͬͯɺଟղ૾౓ͷ1ϐΫηϧ఺
    ϥελϥΠζΛ༻͍ͯߦΘΕ·͢ɻ཭ࢄతͳϥελϥΠζͷۭؒޯ഑͸ɺΰʔετδΦϝτϦͱ͍͏৽͍֓͠೦ʹ
    Αͬͯۙࣅ͞Ε·͢ɻϨϯμϦϯάޙɺχϡʔϥϧΠϝʔδϐϥϛου͸σΟʔϓχϡʔϥϧωοτϫʔΫʹ౉͞
    ΕɺγΣʔσΟϯάܭࢉͱϗʔϧϑΟϦϯά͕ߦΘΕ·͢ɻͦͯ͠ɺඍ෼Մೳͳ෺ཧϕʔεͷτʔϯϚούʔ͕ɺத
    ؒग़ྗΛλʔήοτը૾ʹม׵͠·͢ɻύΠϓϥΠϯͷ͢΂ͯͷεςʔδ͕ඍ෼ՄೳͰ͋ΔͨΊɺγʔϯͷ͢΂ͯͷ
    ύϥϝʔλʢΧϝϥϞσϧɺΧϝϥϙʔζɺϙΠϯτϙδγϣϯɺϙΠϯτΧϥʔɺ؀ڥϚοϓɺϨϯμϦϯάωο
    τϫʔΫͷॏΈɺϰΟωοτɺΧϝϥԠ౴ؔ਺ɺը૾͝ͱͷ࿐ग़ɺը૾͝ͱͷϗϫΠτόϥϯεʣΛ࠷దԽ͠·͢ɻ
    ຊγεςϜͰ͸ɺॳظ࠶ߏ੒ֶ͕शதʹվྑ͞ΕΔͨΊɺطଘͷΞϓϩʔνΑΓ΋γϟʔϓͰҰ؏ੑͷ͋Δ৽͍͠
    ϏϡʔΛ߹੒Ͱ͖Δ͜ͱΛ͍ࣔͯ͠·͢ɻ·ͨɺ1ϐΫηϧͷϙΠϯτΛޮ཰తʹϥελϥΠζ͢Δ͜ͱͰɺ೚ҙͷ
    ΧϝϥϞσϧΛ࢖༻͠ɺ100MϙΠϯτҎ্ͷγʔϯΛϦΞϧλΠϜͰදࣔ͢Δ͜ͱ͕Ͱ͖·͢ɻ
    http://arxiv.org/abs/2110.06635v2
    w ໨తɿΧϝϥը૾ɾ఺܈Λೖྗͱͨ͠ඍ෼Մೳͳө૾ϨϯμϦϯάύΠϓϥΠϯͷ঺հ
    w ੒Ռɿ೚ҙͷඃࣸମͷ -J%"3౳ͷ
    Χϝϥը૾͔Βө૾ΛϦΞϧλΠϜͰඳըͰ͖ΔϞσϧɾπʔϧΛެ։
    w ํ๏ɿ̏࣍ݩҐஔ৘ใ͖ͭͷը૾ɾ఺܈Λ̏࣍ݩతʹॲཧɺิؒɺ&YJG৘ใΛ࢖ͬͯ)%3ʹ࠶ߏ੒
    w ݻ༗໊ɿ"%01 IUUQTHJUIVCDPNEBSHMFJO"%01

    w ஶऀॴଐɿΤΞϥϯήϯʹχϡϧϯϕϧΫେֶ υΠπ

    View full-size slide

  28. ը૾+̏DҐஔ৘ใ͔Βɺ࿐ޫɾϗϫΠτόϥϯ
    εௐ੔ػೳ͖ͭͷө૾(novel frames)Λੜ੒͢Δ
    ೖྗΧϝϥը૾਺఺̏%Ґஔ৘ใ
    ग़ྗOPWFMGSBNFTө૾

    View full-size slide

  29. GitHubͰσϞಈըɾΞϓϦ͕ެ։͞Ε͍ͯΔ

    https://github.com/darglein/ADOP

    View full-size slide

  30. 1.఺܈(LiDARը૾౳)ͷϚοϐϯάɾϥελϥΠζɾิؒ


    2.HDRը૾ϨϯμϦϯά


    3.৭ௐิਖ਼ͯࣗ͠વͳ৭߹͍ʹ߹੒(LDRԽ)

    View full-size slide

  31. 6. ਂ૚χϡʔϥϧωοτϫʔΫʹΑΔ෼ྨͰ͸ɺΑ͘෼ྨ͞Εͨྫ͕աখධՁ͞ΕΔ


    (ݪจ: Well-classi
    fi
    ed Examples are Underestimated in Classi
    fi
    cation with Deep


    Neural Networks)
    ैདྷͷਂ૚෼ྨϞσϧͷֶशͰ͸ɺ෼ྨͷѱ͍ྫʹ஫໨͠ɺܾఆڥք͔Β཭Εͨ෼ྨͷྑ͍ྫΛແࢹ͢Δ͜ͱ
    ͕ৗࣝͰͨ͠ɻྫ͑͹ɺΫϩεΤϯτϩϐʔଛࣦΛ༻ֶ͍ͯश͢Δ৔߹ɺΑΓߴ͍໬౓Λ࣋ͭྫʢ͢ͳΘͪɺ
    Α͘෼ྨ͞Εͨྫʣ͸ɺόοΫϓϩύήʔγϣϯʹ͓͍ͯখ͞ͳޯ഑ʹد༩͠·͢ɻ͔͠͠ɺ͜ͷҰൠతͳख๏
    ͸ɺදݱֶशɺΤωϧΪʔͷ࠷దԽɺϚʔδϯͷ૿ՃΛ๦͛Δ͜ͱΛཧ࿦తʹ͍ࣔͯ͠·͢ɻ͜ͷܽؕΛଧͪফ
    ͨ͢Ίʹɺզʑ͸ɺ෼ྨͷྑ͍ྫʹՃࢉϘʔφεΛ༩͑ͯɺֶश΁ͷߩݙΛ෮׆ͤ͞Δ͜ͱΛఏҊ͢Δɻ͜ͷ
    ൓ྫ͸ɺ͜ΕΒ3ͭͷ໰୊Λཧ࿦తʹղܾ͢Δ΋ͷͰ͋Δɻຊ࿦จͰ͸ɺը૾෼ྨɺάϥϑ෼ྨɺػց຋༁ͳͲ
    ͷଟ༷ͳλεΫʹ͓͍ͯɺཧ࿦తͳ݁ՌΛ௚઀ݕূͨ͠Γɺຊ൓ྫΛ༻͍ͯେ෯ͳੑೳ޲্Λ࣮ݱ͢Δ͜ͱ
    Ͱɺ͜ͷओுΛ࣮ূతʹࢧ࣋͢Δɻ͞Βʹɺຊ࿦จ͸ɺզʑͷΞΠσΞ͕͜ΕΒ3ͭͷ໰୊ΛղܾͰ͖ΔͨΊɺ
    ෆۉߧͳ෼ྨɺOODݕग़ɺఢରత߈ܸԼͷΞϓϦέʔγϣϯͳͲͷෳࡶͳγφϦΦʹରԠͰ͖Δ͜ͱΛࣔͯ͠
    ͍Δɻίʔυ͸ɺhttps://github.com/lancopku/well-classi
    fi
    ed-examples-are-underestimated ɻ


    http://arxiv.org/abs/2110.06537v2
    w ໨తɾ੒ՌɿΫϩεΤϯτϩϐʔϩεΛվྑͨ͠৽͍͠-PTTؔ਺ͷఏҊ
    w ํ๏ɿόοΫϓϩύήʔγϣϯʹΑΔ$SPTT&OUSPQZ $&
    ϩεͷ໰୊఺Λ໌Β͔ʹ͢Δ
    w ݻ༗໊ɿ&ODPVSBHJOH-PTT &-

    w ஶऀॴଐɿ๺ژେֶ

    View full-size slide

  32. ELϩε=CEϩεʴ௥ՃϘʔφε

    View full-size slide

  33. ௥ՃϘʔφεͷՃࢉ஋͸LE=0~1
    Ͱௐ੔͢Δ

    View full-size slide

  34. CrossEntropy(CE) vs
    Encouraging(EL)

    View full-size slide

  35. 7. ByteTrack:͢΂ͯͷݕग़ϘοΫεΛؔ࿈෇͚Δ͜ͱʹΑΔෳ਺෺ମͷ௥੻


    (ݪจ: ByteTrack: Multi-Object Tracking by Associating Every Detection Box)


    ϚϧνɾΦϒδΣΫτɾτϥοΩϯάʢMOTʣ͸ɼಈը಺ͷΦϒδΣΫτͷό΢ϯσΟϯάɾϘοΫεͱΞΠσϯςΟ
    ςΟΛਪఆ͢Δ͜ͱΛ໨తͱ͍ͯ͠·͢ɽଟ͘ͷख๏Ͱ͸ɺᮢ஋ΑΓ΋ߴ͍είΞΛ࣋ͭݕग़ϘοΫεΛؔ࿈෇͚Δ͜
    ͱͰΞΠσϯςΟςΟΛಘ͍ͯ·͢ɻ͔͠͠ɺݕग़είΞͷ௿͍ΦϒδΣΫτʢྫ͑͹ɺӅ͞ΕͨΦϒδΣΫτʣ͸୯
    ७ʹࣺͯΒΕͯ͠·͏ͨΊɺແࢹͰ͖ͳ͍ਅͷΦϒδΣΫτͷܽམ΍ɺஅยతͳي੻͕ੜͯ͡͠·͏ɻ͜ͷ໰୊Λղܾ
    ͢ΔͨΊʹɺզʑ͸BYTEͱݺ͹ΕΔγϯϓϧͰޮՌత͔ͭ൚༻తͳؔ࿈෇͚ํ๏Λఏࣔ͢ΔɻBYTEͱ͸ɺߴείΞͷ
    ݕग़ϘοΫε͚ͩͰͳ͘ɺ͢΂ͯͷݕग़ϘοΫεΛؔ࿈෇͚ͯ௥੻͢Δํ๏Ͱ͋Δɻ௿είΞͷݕग़ϘοΫεʹରͯ͠
    ͸ɼτϥοΫϨοτͱͷྨࣅੑΛར༻ͯ͠ਅͷΦϒδΣΫτΛ෮ݩ͠ɼഎܠݕग़ΛϑΟϧλϦϯά͢ΔɽBYTEΛ9ͭͷ
    ҟͳΔ࠷ઌ୺ͷτϥοΧʔʹద༻ͨ͠ͱ͜ΖɺIDF1είΞΛ1ʙ10ϙΠϯτͷൣғͰҰ؏ͯ͠վળ͢Δ͜ͱ͕Ͱ͖·͠
    ͨɻMOTͷ࠷ઌ୺ͷੑೳΛ׆͔ͨ͢ΊʹɺByteTrackͱ໊෇͚ΒΕͨγϯϓϧͰڧྗͳτϥοΧʔΛઃܭ͠·ͨ͠ɻͦ
    ͷ݁ՌɺV100 GPUΛ༻͍ͨςετηοτʮMOT17ʯʹ͓͍ͯɺMOTA80.3ɺIDF1 77.3ɺHOTA63.1Λୡ੒͠ɺ30
    FPSͷಈ࡞଎౓Λ࣮ݱ͠·ͨ͠ɻιʔείʔυɺࣄલֶशࡁΈϞσϧɺσϓϩΠόʔδϣϯɺଞͷτϥοΧʔ΁ͷద༻ʹ
    ؔ͢ΔνϡʔτϦΞϧ͸ɺhttps://github.com/ifzhang/ByteTrack Ͱެ։͍ͯ͠·͢ɻ
    http://arxiv.org/abs/2110.06864v2
    w ໨తɾ੒ՌɿϚϧνΦϒδΣΫττϥοΩϯάΞϧΰϦζϜ#:5&Ͱ4P5"Λୡ੒ͨ͠
    w ํ๏ɿΧϧϚϯϑΟϧλʹΑΔΦϒδΣΫτҐஔਪఆ
    w ݻ༗໊ɿ#ZUF5SBDL IUUQTHJUIVCDPNJG[IBOH#ZUF5SBDL

    w ஶऀॴଐɿ՚தՊٕେֶɺ߳ߓେֶɺ#ZUF%BODF 5JL5PLͷձࣾ

    View full-size slide

  36. GitHubͰެ։͞Ε͍ͯΔ


    https://github.com/ifzhang/ByteTrack

    View full-size slide

  37. a: YOLOX౳ͷطଘख๏ͰObject Detection


    (t1~t3͸ө૾಺ͷ࿈ଓ͢ΔϑϨʔϜը૾)

    View full-size slide

  38. b: ߴείΞͷശʹඥͮ͘෺ମ(tracklet)ͷ࣍ϑϨʔϜҐஔΛ
    Kalman FilterͰ༧ଌ(IoUͰείΞԽͯۙ͠ࣅ͢Δ΋ͷΛબͿ)

    View full-size slide

  39. c: ௿είΞͷശ͔Βɺ(b)Ͱݕग़Ͱ͖ͳ
    ͔ͬͨtrackletΛਪଌͯ͠Ϛονϯά͢Δ

    View full-size slide

  40. ΧϧϚϯϑΟϧλ


    ϊΠζͷ͋Δෳ਺ͷ৘ใΛ༻͍ͯਅͷঢ়ଶΛਪఆ͢ΔϑΟϧλ
    (ྫɿϩέοτͷঢ়ଶਪఆ,ࣗಈӡస੍ޚ)



    View full-size slide

  41. MOTA,IDF1,FPSͰSoTAΛୡ੒

    View full-size slide

  42. 8. MobileViTɿܰྔɾ൚༻ɾϞόΠϧରԠͷϏδϣϯτϥϯεϑΥʔϚʔ


    (ݪจ: MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision


    Transformer)
    Pickup
    http://arxiv.org/abs/2110.02178v1

    View full-size slide

  43. 9. εέʔϧͰͷߴ଎Ϟσϧฤू


    (ݪจ: Fast Model Editing at Scale)
    ࣄલʹֶश͞Εͨେن໛ͳϞσϧ͸ɺ༷ʑͳμ΢ϯετϦʔϜͷλεΫͰૉ੖Β͍݁͠ՌΛ࣮ݱ͍ͯ͠·͕͢ɺطଘͷେن໛
    ͳϞσϧʹ͸·ͩΤϥʔ͕͋Γɺਖ਼֬ͳ༧ଌͰ͋ͬͯ΋࣌ؒͷܦաͱͱ΋ʹݹ͘ͳͬͯ͠·͏͜ͱ͕͋Γ·͢ɻ͜ͷΑ͏ͳࣦ
    ഊΛ͢΂ֶͯश࣌ʹݕग़͢Δ͜ͱ͸ෆՄೳͰ͋ΔͨΊɺ͜ͷΑ͏ͳϞσϧͷ։ൃऀͱΤϯυϢʔβͷ྆ํ͕ɺϞσϧΛͦͷ·
    ·ʹͯ͠ෆਖ਼֬ͳग़ྗΛमਖ਼Ͱ͖ΔΑ͏ʹ͢Δ͜ͱ͕๬·Ε·͢ɻ͔͠͠ɺେن໛ͳχϡʔϥϧωοτϫʔΫֶ͕श͢Δදݱ
    ͸෼ࢄ͓ͯ͠ΓɺϒϥοΫϘοΫεԽ͍ͯ͠ΔͨΊɺ͜ͷΑ͏ͳର৅Λߜͬͨฤू͸ࠔ೉Ͱ͢ɻ໰୊ͷ͋Δೖྗͱ৽ͨͳر๬
    ͷग़ྗ͕1͚ͭͩఏࣔ͞Εͨ৔߹ɺඍௐ੔Ξϓϩʔν͸ΦʔόʔϑΟοτ͢Δ܏޲͕͋Δɻ·ͨɺଞͷฤूΞϧΰϦζϜ͸ɺඇ
    ৗʹେ͖ͳϞσϧʹద༻͢Δ৔߹ɺܭࢉ͕ෆՄೳͰ͋Δ͔ɺ୯ʹޮՌ͕ͳ͍ɻେن໛ͳϞσϧͷϙετϗοΫฤूΛ༰қʹ͢
    ΔͨΊʹɺզʑ͸MEND (Model Editor Networks with Gradient Decomposition)ΛఏҊ͠·͢ɻMEND͸ɺඪ४తͳඍௐ
    ੔ʹΑͬͯಘΒΕͨޯ഑Λɺޯ഑ͷ௿ϥϯΫ෼ղΛ༻͍ͯม׵͢Δ͜ͱΛֶश͠ɺ͜ͷม׵ͷύϥϝʔλԽΛѻ͍΍ͯ͘͢͠
    ͍·͢ɻMEND͸ɺ100ԯҎ্ͷύϥϝʔλΛ࣋ͭϞσϧͰ͋ͬͯ΋ɺ1ͭͷGPU্Ͱ1೔Ҏ಺ʹֶश͢Δ͜ͱ͕Ͱ͖·͢ɻ·
    ͨɺҰ౓ֶशͨ͠MEND͸ɺࣄલʹֶशͨ͠Ϟσϧʹ৽ͨͳฤूΛՃ͑Δ͜ͱ͕Ͱ͖·͢ɻT5ɺGPTɺBERTɺBARTϞσϧΛ
    ༻͍࣮ͨݧʹΑΓɺMEND͸ɺ਺ઍສ͔Β100ԯҎ্ͷύϥϝʔλΛ࣋ͭϞσϧʹରͯ͠ޮՌతͳฤूΛߦ͏͜ͱ͕Ͱ͖Δ།
    ҰͷϞσϧฤूख๏Ͱ͋Δ͜ͱ͕Θ͔Γ·ͨ͠ɻ࣮૷͸ɺhttps://sites.google.com/view/mend-editing ɻ
    http://arxiv.org/abs/2110.11309v1
    w ໨తɿେن໛ͳ5SBOTGPSNFSϞσϧͰͷ௥Ճमਖ਼ͷͨΊͷ࠶ֶश࣌ͷ໰୊ʹରॲ͢Δ
    w ੒Ռɿߴ଎ʹɺΦʔόʔϑΟοτͳ͘ඍௐ੔Ͱ͖ΔϞσϧ.&/%ͷఏҊ
    w ํ๏ɿϞσϧฤू໰୊ࣗମΛֶश໰୊ͱͯ͠ѻ͏
    w ݻ༗໊ɿ.&/% IUUQTTJUFTHPPHMFDPNWJFXNFOEFEJUJOH

    w ஶऀॴଐɿελϯϑΥʔυେֶ

    View full-size slide

  44. ࣌୅ͷྲྀΕʹΑͬͯ౴͕͑มΘΔ


    →࠶ֶश͕ඞཁ
    • ΠΪϦεͷट૬͸୭ʁ

    ✗ 76୅ट૬(ςϦʔβɾϝΠ)

    ○ 77୅ट૬(ϘϦεɾδϣϯιϯ)
    IUUQTTJUFTHPPHMFDPNWJFXNFOEFEJUJOH

    View full-size slide

  45. ࣌୅ͷྲྀΕʹΑͬͯ౴͕͑มΘΔ


    →෦෼తͳ࠶ֶश͕ඞཁ(=Ϟσϧฤू)
    IUUQTTJUFTHPPHMFDPNWJFXNFOEFEJUJOH

    View full-size slide

  46. ࠶ֶशͯ͠΋ɺؔ܎ͳ͍࣭໰ʹӨڹ͠ͳ͍ࣄ͕େࣄ


    (ྫɿϝογ͕ॴଐ͢ΔεϙʔπνʔϜ͸Ͳ͔͜ʣ

    View full-size slide

  47. MENDͰϞσϧΛֶश͢Δͱɺ෦෼ฤू
    (࠶ֶश)͕Ͱ͖Δˠաֶश͠ͳ͍ˍߴ଎

    View full-size slide

  48. 10. ࣗݾڭࢣ෇ֶ͖श͸σʔληοτͷෆۉߧʹରͯ͠ΑΓؤ݈Ͱ͋Δ


    (ݪจ: Self-supervised Learning is More Robust to Dataset Imbalance)
    ࣗݾڭࢣ෇ֶ͖शʢSelf-Supervised Learning: SSLʣ͸ɼϥϕϧͳ͠Ͱֶश͢ΔͨΊɼҰൠతͳࢹ֮දݱΛֶश͢ΔͨΊͷ
    εέʔϥϒϧͳํ๏Ͱ͋Δɽ͔͠͠ɼେن໛ͳϥϕϧͳ͠σʔληοτͰ͸ɼϥϕϧͷ෼෍͕ϩϯάςʔϧͰ͋Δ͜ͱ͕ଟ
    ͘ɼSSLͷಈ࡞ʹ͍ͭͯ͸΄ͱΜͲ෼͔͍ͬͯͳ͍ɽຊݚڀͰ͸ɼσʔληοτෆۉߧԼͰͷࣗݾڭࢣ෇ֶ͖शΛܥ౷తʹ
    ௐࠪ͢Δɽ·ͣɼେن໛ͳ࣮ݧʹΑΓɼط੡ͷڭࢣ෇͖දݱ͸ɼڭࢣ෇͖දݱΑΓ΋Ϋϥεͷෆۉߧʹରͯ͠ΑΓؤ݈Ͱ͋
    Δ͜ͱ͕Θ͔ͬͨɽSSLΛ༻͍ͨόϥϯεܕͱΞϯόϥϯεܕͷࣄલֶशͷੑೳࠩ͸ɺαϯϓϧαΠζʹؔΘΒͣɺυϝΠ
    ϯ಺ɺಛʹυϝΠϯ֎ͷධՁʹ͓͍ͯɺڭࢣ෇ֶ͖शͷੑೳࠩΑΓ΋༗ҙʹখ͘͞ͳ͍ͬͯ·͢ɻୈೋʹɼSSLͷؤ݈ੑΛ
    ཧղ͢ΔͨΊʹɼSSL͸සग़σʔλ͔ΒΑΓ๛͔ͳಛ௃Λֶश͢Δͱ͍͏ԾઆΛཱͯͨɽͭ·ΓɼكͳΫϥε΍Լྲྀͷλε
    Ϋͷ෼ྨʹ໾ཱͭɼϥϕϧͱ͸ແؔ܎͕ͩ఻ୡՄೳͳಛ௃Λֶश͢ΔͷͰ͸ͳ͍͔ͱߟ͑ΒΕΔɽରরతʹɼڭࢣ෇ֶ͖
    शͰ͸ɼසग़͢Δྫ͔Βϥϕϧͱ͸ແؔ܎ͳಛ௃Λֶश͢Δಈػ͕ͳ͍ɽ͜ͷԾઆΛɺ୯७Խ͞ΕͨઃఆͰͷ൒߹੒࣮ݧ
    ͱཧ࿦త෼ੳʹΑͬͯݕূ͢Δɻୈࡾʹɺཧ࿦తಎ࡯ʹ৮ൃ͞Εͯɺ࠶ॏΈ෇͚ਖ਼ଇԽٕज़ΛߟҊͨ͠ɻ͜ͷٕज़͸ɺ͍͘
    ͔ͭͷධՁج४ʹج͍ͮͯɺΞϯόϥϯεͳσʔληοτʹ͓͚ΔSSLදݱͷ඼࣭ΛҰ؏ͯ͠޲্ͤ͞ɺಉ͡਺ͷྫΛ࣋ͭ
    όϥϯεͷऔΕͨσʔληοτͱΞϯόϥϯεͳσʔληοτͷؒͷখ͞ͳΪϟοϓΛຒΊΔɻ
    http://arxiv.org/abs/2110.05025v1
    w ໨తɿΫϥεෆۉߧԼͰͷʮࣗݾڭࢣֶ͖ͭश 44-
    ʯͷදݱ඼࣭Λܥ౷తʹௐࠪ͢Δ
    w ੒Ռɿ44-͕σʔληοτͷෆۉߧʹରͯ͠ؤڧͳࣄΛԠ༻ͯ͠SX4".ΛߟҊ
    w ݻ༗໊ɿ3FXFJHIUFE4". SX4".

    w ஶऀॴଐɿελϯϑΥʔυେֶɺτϤλɾϦαʔνɾΠϯεςΟςϡʔτ

    View full-size slide

  49. Top hype: Best10

    View full-size slide

  50. 1. ADOP: Approximate Differentiable One-Pixel Point Rendering (ۙࣅࠩҟԽ1ϐΫη
    ϧɾϙΠϯτɾϨϯμϦϯά)


    ॏෳ
    http://arxiv.org/abs/2110.06635v2

    View full-size slide

  51. 2. ࣮਺ɺσʔλαΠΤϯεɺΧΦεɿ͋ΒΏΔσʔληοτΛ୯ҰͷύϥϝʔλͰϑΟοτͤ͞Δํ๏


    (ݪจ: Real numbers, data science and chaos: How to
    fi
    t any dataset with a


    single parameter)
    ࣌ܥྻɺը૾ɺԻ੠ͳͲɺͲͷΑ͏ͳϞμϦςΟͷσʔλͰ͋ͬͯ
    ΋ɺ୯Ұͷ࣮਺஋ͷύϥϝʔλΛ࣋ͭྑ޷ͳεΧϥʔؔ਺ʢ࿈ଓɺඍ
    ෼Մೳ...ʣͰۙࣅͰ͖Δ͜ͱΛࣔ͠·͢ɻຊݚڀͰ͸ɺΧΦεཧ࿦ͷ
    جຊతͳ֓೦ʹج͍ͮͯɺσʔλͷ͢΂ͯͷαϯϓϧʹ೚ҙͷਫ਼౓Ͱ
    ϑΟοτͤ͞ΔͨΊʹɺ͜ͷύϥϝʔλΛௐ੔͢Δํ๏Λࣔ͢ڭҭత
    ͳΞϓϩʔνΛ࠾༻͍ͯ͠·͢ɻ޷ح৺Ԣ੝ͳσʔλαΠΤϯςΟε
    τΛର৅ʹɺػցֶशϞσϧͷදݱྗͱҰൠԽʹؔ͢Δ͜Ε·Ͱͷಉ
    ༷ͷ؍࡯݁ՌΛൃలͤͨ͞΋ͷͰ͢ɻ
    http://arxiv.org/abs/1904.12320v1
    w ໨తɿ೚ҙͷσʔληοτ9ͷ͢΂ͯͷαϯϓϧ͕ɼ୯७ͳඍ෼ํఔࣜʹΑͬͯ࠶ݱͰ͖Δ͜ͱΛࣔ͢͜ͱ
    w ੒ՌɿҰͭͷ࣮਺஋ύϥϝʔλ͚ͩͰશͯͷԻ੠ɾࢹ֮σʔλΛੜ੒Ͱ͖Δɺ୯७Ͱඍ෼ՄೳͳఆࣜԽ
    w ํ๏ɿΧΦεཧ࿦Λݩʹͨ͠Ի੠ɾը૾σʔλͷҰൠԽ
    w ݻ༗໊ɿ4JOHMF1BSBNFUFS'JU IUUQTHJUIVCDPN3BOMPUTJOHMFQBSBNFUFS
    fi
    U

    w ஶऀॴଐɿ4"1-BCT υΠπͷιϑτ΢ΣΞاۀͷݚڀػؔ

    View full-size slide

  52. 3. σϧϑΝΠػցͷྙཧɾنൣΛ໨ࢦͯ͠


    (ݪจ: Delphi: Towards Machine Ethics and Norms)
    ػցʹྙཧతͳߦಈΛڭ͑Δʹ͸ɺԿ͕ඞཁͰ͠ΐ͏͔ʁେ·͔ͳྙཧతϧʔϧ͸ɺʮೊɺࡴ͢ͳ͔Εʯͱ͍͏Α͏ʹ؆୯ʹड़΂Δ͜
    ͱ͕Ͱ͖Δ͔΋͠Ε·ͤΜ͕ɺͦͷΑ͏ͳϧʔϧΛݱ࣮ͷঢ়گʹద༻͢Δ͜ͱ͸͸Δ͔ʹෳࡶͰ͢ɻྫ͑͹ɺʮ༑ਓΛॿ͚Δʯ͜ͱ͸
    Ұൠతʹྑ͍͜ͱͰ͕͢ɺʮ༑ਓ͕ϑΣΠΫχϡʔεΛྲྀ͢ͷΛॿ͚Δʯ͜ͱ͸ྑ͍͜ͱͰ͸͋Γ·ͤΜɻࢲͨͪ͸ɺػցྙཧ΍نൣ
    ʹର͢Δ4ͭͷࠜຊతͳ՝୊Λಛఆ͍ͯ͠·͢ɻ(1) ಓಙతڭ܇ͱࣾձతنൣͷཧղɺ(2) ݱ࣮ੈքͷঢ়گΛࢹ֮తʹɺ͋Δ͍͸ࣗવݴޠ
    ʹΑΔهड़ΛಡΈऔΔೳྗɺ(3) ҟͳΔจ຺ʹ͓͚Δ୅ସతͳߦಈͷ݁ՌΛ༧ଌ͢ΔͨΊͷৗࣝతͳਪ࿦ɺ(4) ࠷΋ॏཁͳͷ͸ɺڝ߹͢
    ΔՁ஋ͷ૬ޓ࡞༻ͱɺҟͳΔจ຺ʹ͓͚ΔͦΕΒͷࠜڌΛߟྀͯ͠ྙཧత൑அΛԼ͢ೳྗʢྫɿදݱͷࣗ༝ͷݖརͱϑΣΠΫχϡʔε
    ͷ֦ࢄ๷ࢭʣɻ ຊߘͰ͸ɺ͜ΕΒͷٙ໰Λਂ૚ֶशͷύϥμΠϜͷதͰղܾ͠Α͏ͱ͢Δ΋ͷͰ͋ΔɻզʑͷϓϩτλΠϓϞσϧͰ͋
    ΔDelphi͸ɺݴޠϕʔεͷৗࣝతͳಓಙతਪ࿦ʹڧ͍ظ଴Λد͓ͤͯΓɺਓؒʹΑͬͯݕূ͞Εͨਫ਼౓͸࠷େͰ92.1%Ͱ͋Δɻ͜Ε
    ͸ɺGPT-3ͷθϩγϣοτੑೳ͕52.3%Ͱ͋Δͷͱ͸ରরతͰ͋Γɺେن໛ͳεέʔϧ͚ͩͰ͸ɺࣄલʹ܇࿅͞ΕͨਆܦݴޠϞσϧʹਓ
    ؒͷՁ஋Λ෇༩͢Δ͜ͱ͸Ͱ͖ͳ͍͜ͱΛ͍ࣔࠦͯ͠Δɻͦ͜Ͱࢲͨͪ͸ɺػց༻ʹΧελϚΠζ͞ΕͨಓಙͷڭՊॻ
    ʮCommonsense Norm BankʯΛൃද͠·ͨ͠ɻ͜ͷڭՊॻʹ͸ɺ೔ৗͷ͞·͟·ͳঢ়گʹ͓͚Δਓʑͷྙཧత൑அͷྫ͕170ສ݅ऩ
    ࿥͞Ε͍ͯ·͢ɻຊݚڀ͸ɺࠓޙͷݚڀͷͨΊͷ৽ͨͳϦιʔεͱϕʔεϥΠϯͱͳΔੑೳʹՃ͑ͯɺਓؒͷීวతͳՁ஋ͱݸਓతͳՁ
    ஋ͷ۠ผɺҟͳΔಓಙత࿮૊ΈͷϞσϧԽɺػցྙཧ΁ͷઆ໌ՄೳͰҰ؏ੑͷ͋ΔΞϓϩʔνͳͲɺ͍͔ͭ͘ͷॏཁͳະղܾͷݚڀ՝
    ୊ʹͭͳ͕Δ৽ͨͳಎ࡯Λఏڙ͍ͯ͠·͢ɻ
    http://arxiv.org/abs/2110.07574v1
    w ໨తɿಓಙతɾྙཧతͳֶशɾਪ࿦
    w ੒Ռɿػց༻ʹΧελϚΠζ͞ΕͨಓಙͷڭՊॻʮ$PNNPOTFOTF/PSN#BOLʯΛൃද
    w ํ๏ɿ
    w ݻ༗໊ɿϓϩτλΠϓϞσϧ%FMQIJσʔληοτ$PNNPOTFOTF/PSN#BOL
    w ஶऀॴଐɿϫγϯτϯେֶ ΞϨϯਓ޻஌ೳݚڀॴ

    View full-size slide

  53. Delphi: ྙཧΛ౿·͑ͯɺߦಈͷ
    ྑ͠ѱ͠Λ൑அ͢ΔϞσϧ

    View full-size slide

  54. σϞαΠτ:


    https://delphi.allenai.org/

    View full-size slide

  55. 4. ϚϧνλεΫϓϩϯϓτʹΑΔτϨʔχϯάͰθϩγϣοτλεΫͷҰൠԽΛ࣮ݱ


    (ݪจ: Multitask Prompted Training Enables Zero-Shot Task Generalization)
    ۙ೥ɺେن໛ͳݴޠϞσϧ͕ɺଟ༷ͳλεΫʹ͓͍ͯଥ౰ͳθϩγϣοτ൚ԽΛୡ੒͢Δ͜ͱ͕ࣔ͞Ε͍ͯ
    Δɻ͜Ε͸ɺݴޠϞσϧͷֶशʹ͓͚Δ҉໧ͷϚϧνλεΫֶशͷ݁ՌͰ͋Δͱ͍͏Ծઆཱ͕ͯΒΕ͍ͯ·
    ͢ɻͰ͸ɼ໌ࣔతͳϚϧνλεΫֶशʹΑͬͯɼθϩγϣοτͷ൚Խ͕௚઀Ҿ͖ى͜͞ΕΔͷͰ͠ΐ͏͔ʁ͜
    ͷٙ໰Λେن໛ʹݕূ͢ΔͨΊʹɺҰൠతͳࣗવݴޠλεΫΛਓ͕ؒಡΊΔϓϩϯϓτܗࣜʹ؆୯ʹϚοϐϯ
    ά͢ΔγεςϜΛ։ൃ͠·ͨ͠ɻେن໛ͳڭࢣ෇͖σʔληοτΛม׵͠ɺͦΕͧΕͷσʔληοτʹ͸༷ʑ
    ͳࣗવݴޠΛ༻͍ͨෳ਺ͷϓϩϯϓτ͕༻ҙ͞Ε͍ͯΔɻ͜ΕΒͷϓϩϯϓτσʔληοτ͸ɺࣗવݴޠͰࢦ
    ఆ͞Εͨશ͘ݟͨ͜ͱͷͳ͍λεΫΛ࣮ߦ͢ΔϞσϧͷೳྗΛϕϯνϚʔΫ͢Δ͜ͱ͕Ͱ͖Δɻࣄલʹֶश͠
    ͨΤϯίʔμʔͱσίʔμʔͷϞσϧΛɺ༷ʑͳλεΫΛؚΉϚϧνλεΫࠞ߹෺Ͱඍௐ੔ͨ͠ɻ͜ͷϞσϧ
    ͸ɺ͍͔ͭ͘ͷඪ४తͳσʔληοτʹ͓͍ͯڧྗͳθϩγϣοτੑೳΛୡ੒͠ɺ͠͹͠͹16ഒͷαΠζͷϞ
    σϧΑΓ΋༏Ε͍ͯΔɻ͞ΒʹɺBIG-BenchϕϯνϚʔΫͷλεΫͷαϒηοτʹ͓͍ͯ΋ɺ6ഒͷϞσϧΑΓ
    ΋ߴ͍ੑೳΛൃش͠·ͨ͠ɻ͢΂ͯͷϓϩϯϓτͱֶशࡁΈϞσϧ͸ɼgithub.com/bigscience-workshop/
    promptsource/Ͱެ։͞Ε͍ͯ·͢ɽ
    http://arxiv.org/abs/2110.08207v1
    w ໨తɿେن໛ͳݴޠϞσϧ͕θϩγϣοτ൚Խ͕Ͱ͖Δཧ༝ͷݚڀ
    w ੒ՌɿݴޠϞσϧͷڧྗͳθϩγϣοτҰൠԽೳྗΛ࣮ݱͰ͖ΔπʔϧͱֶशࡁΈϞσϧΛެ։
    w ํ๏ɿϚϧνλεΫϓϩϯϓττϨʔχϯά
    w ݻ༗໊ɿ5ϕʔεͷ5Ϟσϧ 1SPNQU4PVSDF HJUIVCDPNCJHTDJFODFXPSLTIPQ
    QSPNQUTPVSDF

    w ஶऀॴଐɿ)VHHJOH'BDF 🤗
    ϒϥ΢ϯେֶ΄͔

    View full-size slide

  56. Multitask Prompt Training


    ଟ༷ͳܗࣜͰ࣭໰ɾճ౴จΛ࡞ֶͬͯश͢Δ͜
    ͱͰɺະ஌ͷܗࣜͰͷθϩγϣοτʹڧ͘͢Δ

    View full-size slide

  57. T0: Google ͷ T5Ϟσϧϕʔε
    IUUQTHJUIVCDPNHPPHMFSFTFBSDIUFYUUPUFYUUSBOTGFSUSBOTGPSNFS

    View full-size slide

  58. GPT3(ύϥϝʔλ਺175B)ΑΓT0(ύϥϝʔλ਺10B)͕ߴ͍ਫ਼
    ౓ʹͳͬͨ.

    View full-size slide

  59. PromptSource: ϓϩϯϓτܗࣜσʔληοτ࡞੒πʔϧ


    https://github.com/bigscience-workshop/promptsource/

    View full-size slide

  60. 5. ඇෛͷۭؒతҼ਺෼ղ


    (ݪจ: Nonnegative spatial factorization)
    Ψ΢εաఔ͸ɺϊϯύϥϝτϦοΫͳॊೈੑͱෆ࣮֬ੑͷఆྔԽ͕ՄೳͰ͋Δ͜ͱ͔Βɺۭؒσʔλͷ
    ղੳʹ޿͘༻͍ΒΕ͓ͯΓɺ࠷ۙ։ൃ͞Εͨεέʔϥϒϧͳۙࣅ๏ʹΑΓɺ๲େͳσʔληοτ΁ͷద
    ༻͕༰қʹͳ͍ͬͯΔɻଟมྔͷ݁Ռʹରͯ͠͸ɺۭؒ૬ؔΛར༻ͨ࣍͠ݩ࡟ݮΛ૊Έ߹ΘͤͨίΞϦ
    φΠζͷઢܗϞσϧ͕͋Δɻ͔͠͠ɺඇෛϞσϧͱ͸ҟͳΓɺύʔπϕʔεͷදݱΛճ෮͠ͳ͍ͨΊɺ
    ͦͷ࣮਺જࡏҼࢠͱෛՙྔ͸ղऍ͕೉͍͠ɻຊݚڀͰ͸ɺۭؒΛߟྀͨ֬͠཰త࣍ݩ࡟ݮϞσϧͰ͋Δ
    ඇෛͷۭؒҼࢠԽʢNSFʣΛఏҊ͢ΔɻNSF͸ɺγϛϡϨʔγϣϯͱߴ࣍ݩۭؒτϥϯεΫϦϓτϛΫ
    εσʔλΛ༻͍ͯɺMEFISTOͷΑ͏ͳ࣮਺ۭؒҼࢠԽ΍ඇۭؒ࣍ݩ࡟ݮ๏ͱൺֱͨ͠ɻNSF͸ɺҨ఻ࢠ
    ൃݱͷҰൠԽՄೳͳۭؒύλʔϯΛಛఆ͠·͢ɻ͢΂ͯͷҨ఻ࢠൃݱύλʔϯ͕ۭؒతͰ͋Δͱ͸ݶΒ
    ͳ͍ͨΊɺۭؒతͳཁૉͱඇۭؒతͳཁૉΛ૊Έ߹ΘͤͨNSFͷϋΠϒϦου֦ுΛఏҊ͠ɺ؍ଌ஋ͱ
    ಛ௃ͷ྆ํʹ͍ۭͭͯؒతͳॏཁੑΛఆྔԽ͢Δ͜ͱΛՄೳʹ͍ͯ͠·͢ɻNSFͷTensorFlow࣮૷͸ɺ
    https://github.com/willtownes/nsf-paper ͔ΒೖखՄೳͰ͋Δɻ
    http://arxiv.org/abs/2110.06122v1
    w ໨తɿੜମ૊৫ͷݚڀʹ͓͚ΔۭؒతͳҨ఻ࢠൃݱͷଌఆ
    w ੒ՌɿҨ఻ࢠൃݱͷҰൠԽՄೳͳۭؒύλʔϯΛಛఆ
    w ํ๏ɿΨ΢εաఔΛ༻͍ͨσʔλղੳͰͷۭؒΛߟྀͨ֬͠཰త࣍ݩ࡟ݮϞσϧΛߟҊ
    w ݻ༗໊ɿ/4' /POOFHBUJWF4QBUJBM'BDUPSJ[BUJPO
    /4') /4')ZCSJE

    w ஶऀॴଐɿϓϦϯετϯେֶ άϥουετʔϯݚڀॴ αϯϑϥϯγεί

    View full-size slide

  61. Ϛ΢εͷ೴ͷVisiumۭؒతҨ఻ࢠൃݱղੳ


    View full-size slide

  62. 6. ߴ࣍ݩͰͷֶͼ͸ɺৗʹ֎Ԇతͳ΋ͷͰ͋Δɻ


    (ݪจ: Learning in High Dimension Always Amounts to Extrapolation)
    ॏෳ
    http://arxiv.org/abs/2110.09485v1

    View full-size slide

  63. 7. StyleAlignɿ੔ྻͨ͠StyleGANϞσϧͷ෼ੳͱԠ༻


    (ݪจ: StyleAlign: Analysis and Applications of Aligned StyleGAN Models)
    ຊ࿦จͰ͸ɺΞϥΠϝϯτ͞Εͨੜ੒ϞσϧͷಛੑͱͦͷԠ༻ʹ͍ͭͯৄࡉʹݕ౼ͨ͠ɻ͜͜Ͱ͸ɺ2ͭͷϞσϧ
    ͕ಉ͡ΞʔΩςΫνϟΛڞ༗͠ɺҰํʢࢠʣ͕ଞํʢ਌ʣ͔ΒผͷυϝΠϯ΁ͷඍௐ੔ΛܦͯಘΒΕͨ৔߹ɺ੔
    ྻͨ͠ϞσϧͱݺͿ͜ͱʹ͢Δɻ͢Ͱʹ͍͔ͭ͘ͷ࡞඼Ͱ͸ɺΞϥΠϝϯτ͞ΕͨStyleGANϞσϧͷجຊతͳಛ
    ੑΛར༻ͯ͠ɺը૾ؒͷ຋༁Λߦ͍ͬͯΔɻ͜͜Ͱ͸ɺStyleGANʹয఺Λ౰ͯͯɺϞσϧͷΞϥΠϯϝϯτΛॳ
    Ίͯৄࡉʹௐࠪ͢Δɻ·ͣɺ੔ྻͨ͠ϞσϧΛܦݧతʹ෼ੳ͠ɺͦͷੑ࣭ʹؔ͢Δॏཁͳٙ໰ʹର͢Δ౴͑Λఏ
    ڙ͢ΔɻಛʹɺࢠϞσϧͷજࡏۭؒ͸਌Ϟσϧͷજࡏۭؒͱҙຯతʹ੔߹͓ͯ͠Γɺਓͷإ΍ڭձͳͲͷԕ͍
    σʔλྖҬͰ͋ͬͯ΋ɺ৴͡ΒΕͳ͍΄Ͳ๛͔ͳҙຯΛܧঝ͍ͯ͠Δ͜ͱ͕Θ͔Γ·ͨ͠ɻ࣍ʹɺ͜ͷΑ͏ʹ͠
    ͯಘΒΕͨཧղΛ΋ͱʹɺ੔ྻͨ͠ϞσϧΛ׆༻ͯ͠͞·͟·ͳ՝୊Λղܾ͠·͢ɻը૾຋༁ʹՃ͑ͯɺ׬શʹ
    ࣗಈԽ͞ΕͨΫϩευϝΠϯͷը૾ϞʔϑΟϯάΛ࣮ূ͠·ͨ͠ɻ͞Βʹɺ਌ྖҬͰͷ؂ࢹͷΈʹཔΓͳ͕Βɺ
    ࢠྖҬͰ͸θϩγϣοτͷࢹ֮λεΫΛ࣮ߦͰ͖Δ͜ͱΛࣔ͠·͢ɻ͞Βʹɺ਌ྖҬͷ؂ࢹͷΈʹґଘ͠ͳ͕
    ΒɺࢠྖҬͰθϩγϣοτɾϏδϣϯɾλεΫΛ࣮ߦ͢Δ͜ͱ͕Ͱ͖Δ͜ͱΛࣔ͠·ͨ͠ɻ͜ͷΞϓϩʔνʹΑ
    Γɺ؆୯ͳඍௐ੔ͱ൓సͷΈͰɺ࠷ઌ୺ͷ݁Ռ͕ಘΒΕΔ͜ͱΛఆੑత͓Αͼఆྔతʹࣔ͠·ͨ͠ɻ
    http://arxiv.org/abs/2110.11323v1
    w ໨తɿ4UZMF("/ͷ෼ੳͱԠ༻
    w ੒Ռɿ4UZMF("/ͷΫϩευϝΠϯͷసҠֶशͷੑೳΛվળͰ͖ͨ
    w ํ๏ɿ4UZMF("/Λશ͘ผυϝΠϯʹసҠֶशͨ࣌͠ͷજࡏۭؒΛ෼ੳ
    w ݻ༗໊ɿ4UZMF"MJHO
    w ஶऀॴଐɿϔϒϥΠେֶ ςϧΞϏϒେֶ "EPCF3FTFBSDI

    View full-size slide

  64. FFHQσʔληοτͰֶशͨ͠StyleGAN2Λ Mega, Dog
    ͰϑΝΠϯνϡʔχϯάͨ͠ޙɺॳظ஋Λ෦෼తʹFFHQ
    ͷॏΈʹϦηοτͨ࣌͠ͷ݁ՌͷมԽ
    ਌ͷॏΈʹϦηοτͯ͠΋Өڹ͕খ͍͞
    ʹ਌ͷॏΈ͔Β͋·ΓมΘ͍ͬͯͳ͍
    ਌ͷॏΈʹϦηοτ͢Δͱେ͖͘Өڹ
    ʹ਌ͷॏΈ͔Βେ͖͘มΘ͍ͬͯΔ
    ɹ ಛʹਓˠݘͷผυϝΠϯֶश࣌

    View full-size slide

  65. ਓͱڭձͷΑ͏ʹυϝΠϯ͕େ͖͘มΘͬͨͱͯ͠΋ɺ


    ϚοϐϯάɾΞϑΟϯ͸ྨࣅ͍ͯ͠Δ

    View full-size slide

  66. FFHQ -> Merface or Mega ΁ͷసҠֶश


    ਌ϞσϧͰͷηϚϯςΟοΫίϯτϩʔϧ͸ࢠϞσϧͰ΋Ҿ͖ܧ͕ΕΔɻ


    ʢ=જࡏۭؒWͱS͕͋·ΓมԽ͍ͯ͠ͳ͍)

    View full-size slide

  67. StyleAlign=StyleGAN2΍StyleGAN2-ADAΛ
    ඍௐ੔ͯ͠ϑΝΠϯνϡʔχϯάֶशͨ͠΋ͷ

    View full-size slide

  68. StyleAlign vs Others

    View full-size slide

  69. 8. AudacityͷͨΊͷਂ૚ֶशπʔϧɻݚڀऀ͕ΞʔςΟετͷπʔϧΩοτΛ֦ு͢Δͷʹ໾ཱͭ


    (ݪจ: Deep Learning Tools for Audacity: Helping Researchers Expand the


    Artist's Toolkit)
    ࢲͨͪ͸ɺΦʔϓϯιʔεͷਓؾΦʔσΟΦฤूιϑτAudacity
    ʹɺ࠷খݶͷ։ൃऀͷ࿑ྗͰχϡʔϥϧωοτϫʔΫΛ౷߹͢Δι
    ϑτ΢ΣΞϑϨʔϜϫʔΫΛ঺հ͠·͢ɻຊ࿦จͰ͸ɺΤϯυϢʔ
    βʔͱχϡʔϥϧωοτϫʔΫ։ൃऀͷ྆ํʹ޲͚ͯɺ͍͔ͭ͘ͷ
    ࢖༻ྫΛ঺հ͠·͢ɻ͜ͷݚڀ͕ɺਂ૚ֶशͷ࣮ફऀͱΤϯυϢʔ
    βʔͷؒͷ৽͍͠Ϩϕϧͷ૬ޓ࡞༻Λଅਐ͢Δ͜ͱΛظ଴͍ͯ͠·
    ͢ɻ
    http://arxiv.org/abs/2110.13323v1
    w ໨తɾ੒ՌɿԻ੠ฤूιϑτ"VEBDJUZΛχϡʔϥϧωοτϫʔΫʹରԠ͢Δ
    w ํ๏ɿΦϯϥΠϯαΠτ)VHHJOH'BDFͷެ։ϞσϧʹରԠ
    w ݻ༗໊ɿ"VEBDJUZ%JHJUBM"VEJP8PSLTUBUJPO
    w ஶऀॴଐɿϊʔε΢Σελϯେֶ "VEBDJUZ5FBN Ի੠ฤूιϑτ։ൃνʔϜ

    View full-size slide

  70. Ի੠ฤूιϑτ Audacity Digital Audio WorkstationͰ


    HuggingFaceʹެ։͞Ε͍ͯΔϞσϧΛϩʔΧϧPCͰ࣮
    ߦͯ͠Ի੠Ճ޻Ͱ͖ΔΑ͏ʹͳͬͨ

    View full-size slide

  71. ೚ҙͷϞσϧΛ࢖ͬͯɺ


    Ի੠ΛՃ޻ͨ͠ΓɺϥϕϧͳͲʹม׵Ͱ͖Δ

    View full-size slide

  72. 9. ECQx:௿ϏοτͰૄͳDNNͷͨΊͷઆ໌Մೳੑʹجͮ͘ྔࢠԽ


    (ݪจ: ECQx: Explainability-Driven Quantization for Low-Bit and


    Sparse DNNs)
    ༷ʑͳΞϓϦέʔγϣϯʹ͓͚ΔσΟʔϓχϡʔϥϧωοτϫʔΫʢDNNʣͷ໨֮·͍͠੒ޭ͸ɺωοτϫʔΫύϥϝʔλ΍ԋ
    ࢉྔͷେ෯ͳ૿ՃΛ൐͍ͬͯ·͢ɻ͜ͷΑ͏ͳϝϞϦ΍ܭࢉྔͷ૿Ճ͸ɺϞόΠϧػثͷΑ͏ͳϦιʔεʹ੍໿ͷ͋Δϋʔυ΢Σ
    ΞϓϥοτϑΥʔϜͰ͸ɺਂ૚ֶशΛߦ͏͜ͱ͕Ͱ͖·ͤΜɻ࠷ۙͰ͸ɺϞσϧͷੑೳΛՄೳͳݶΓҡ࣋ͭͭ͠ɺ͜ΕΒͷΦʔ
    όʔϔουΛ࡟ݮ͢Δ͜ͱΛ໨తͱͯ͠ɺύϥϝʔλ࡟ݮٕज़ɺύϥϝʔλྔࢠԽɺՄٯѹॖٕज़ͳͲ͕։ൃ͞Ε͍ͯΔɻ ຊষ
    Ͱ͸ɺDNNͷͨΊͷ৽͍͠ྔࢠԽύϥμΠϜΛ։ൃ͠ɺઆ໌͠·͢ɻຊख๏Ͱ͸ɺઆ໌ՄೳͳAI(XAI)ͷ֓೦ͱ৘ใཧ࿦ͷ֓೦Λ
    ׆༻͠ɺྔࢠԽΫϥελ΁ͷڑ཭ʹج͍ͮͯॏΈ஋ΛׂΓ౰ͯΔ୅ΘΓʹɺϨΠϠϫΠζؔ࿈ੑ఻೻(LRP)͔ΒಘΒΕΔॏΈͷؔ
    ࿈ੑͱΫϥελͷ৘ใྔ(Τϯτϩϐʔ࠷దԽ)Λ௥ՃͰߟྀ͠·͢ɻ࠷ऴతͳ໨ඪ͸ɺ࠷΋ؔ࿈ੑͷߴ͍ॏΈΛɺ࠷΋৘ใྔͷଟ
    ͍ྔࢠԽΫϥελʹอଘ͢Δ͜ͱͰ͢ɻ ࣮ݧͷ݁Ռɺ͜ͷ৽͍͠Τϯτϩϐʔ੍໿෇͖XAIௐ੔ྔࢠԽʢECQxʣ๏͸ɺϞσϧͷ
    ੑೳΛҡ࣋·ͨ͸޲্ͤ͞ͳ͕Βɺ௒௿ਫ਼౓ʢ2ʙ5ϏοτʣͰಉ࣌ʹεύʔεͳχϡʔϥϧωοτϫʔΫΛੜ੒͢Δ͜ͱ͕෼͔
    Γ·ͨ͠ɻ·ͨɼύϥϝʔλͷਫ਼౓͕௿͘ɼθϩཁૉͷ਺͕ଟ͍͜ͱ͔ΒɼϑΝΠϧαΠζͷ఺Ͱ΋ѹॖੑ͕ߴ͘ɼߴਫ਼౓ͷྔ
    ࢠԽ͞Ε͍ͯͳ͍DNNϞσϧͱൺֱͯ͠ɼ࠷େͰ 103ഒͷѹॖޮՌ͕ಘΒΕ·͢ɽզʑͷΞϓϩʔν͸ɺ༷ʑͳλΠϓͷϞσϧ
    ͱσʔληοτʢGoogle Speech Commands΍CIFAR-10ͳͲʣͰධՁ͞Εɺաڈͷݚڀͱൺֱ͞Ε·ͨ͠ɻ
    http://arxiv.org/abs/2109.04236v1
    w ໨తɿϞσϧͷੑೳΛҡ࣋ͭͭ͠ྔࢠԽͯ͠αΠζѹॖ͢Δ
    w ੒Ռɿ7((Ͱςετͯ͠࠷େͰഒͷѹॖޮՌ͕ಘΒΕͨ
    w ํ๏ɿઆ໌Մೳͳ"* 9"*
    ͷ֓೦ͱ৘ใཧ࿦ͷ֓೦Λ׆༻
    w ݻ༗໊ɿ&$2 &$2Y
    w ஶऀॴଐɿϑϥ΢ϯϗʔϑΝʔ))*ݚڀॴ υΠπ
    #*'0-% υΠπ

    View full-size slide

  73. VGG16ͱಠࣗMLPͷྔࢠԽ݁Ռ


    ECQxͰ4ϏοτྔࢠԽͨ͠VGG16Ϟσϧ͕ɺ-0.1ͷਫ਼౓ྼԽͰ102.59ഒͷѹॖഒ཰Λୡ੒ͨ͠

    View full-size slide

  74. 10. େن໛ͳࣄલτϨʔχϯάͷݶքΛ୳Δ


    (ݪจ: Exploring the Limits of Large Scale Pre-training)
    ॏෳ
    http://arxiv.org/abs/2110.02095v1

    View full-size slide

  75. DeepL Translator (deepl.com)
    https://www.deepl.com/en/translator

    View full-size slide