AI最新技術Update会 8月

イベント発表資料です。
https://deeplearning-b.connpass.com/event/181528/

Transcript

1. AI࠷৽ٕज़Updateձ 8݄
   ᷂tech vein ழມ ॆԝ
   

   View Slide

2. ࣗݾ঺հ
   ழມ ॆԝ (͍ͷ·ͨ ΈͭͻΖ)
   גࣜձࣾ tech vein ୅දऔక໾ ݉ σϕϩούʔ
   twitter: @ino2222
   IUUQTXXXUFDIWFJODPN
   

   View Slide

3. ΞδΣϯμ
   Archive Sanity (arxiv-sanity.com) ͔ΒϐοΫΞο
   ϓͨ͠ɺarxiv.org ͷաڈ1ϲ݄ؒͷ࿦จ঺հɻ
   ɾtop recentͷ࿦จτοϓ10
   ɾtop hype ͷ࿦จτοϓ10
   ɾҰ൪ؾʹͳͬͨ࿦จͷ঺հ
   

   View Slide

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

   View Slide

5. Arxiv Sanity
   Top recent: Best10
   

   View Slide

6. ᶃTransfromer͸RNNͰ͢ɻઢܗ஫ҙͷߴ଎ࣗݾճؼτϥϯεϑΥʔϚʔ
   (ݪจ: Transformers are RNNs: Fast Autoregressive Transformers with
   Linear Attention)
   Transformer͸͍͔ͭ͘ͷλεΫͰݦஶͳੑೳΛୡ੒͍ͯ͠Δ͕ɼ
   ೖྗͷ௕͞ʹରͯ͠ೋ࣍తͳෳࡶ͞Λ࣋ͭͨΊɼඇৗʹ௕͍γʔ
   έϯεͰ͸๏֎ʹ஗͍ɽ͜ͷ੍ݶʹରॲ͢ΔͨΊʹɼզʑ͸ࣗݾ
   ஫ҙΛΧʔωϧಛ௃ྔϚοϓͷઢܗ఺ੵͱͯ͠දݱ͠ɼߦྻੵͷ
   ࿈૝ੑͷੑ࣭Λར༻ͯ͠ɼෳࡶ͞Λ O(N^2) ͔Β O(N) ʹݮΒ͢
   ͜ͱΛࢼΈΔɽ͜ͷఆࣜԽʹΑΓɺࣗݾճؼܕTransformerΛܶత
   ʹߴ଎Խ͠ɺϦΧϨϯτɾχϡʔϥϧɾωοτϫʔΫͱͷؔ܎Λ
   ໌Β͔ʹ͢Δ൓෮࣮૷͕ՄೳͰ͋Δ͜ͱΛࣔ͢ɻզʑͷઢܗม׵
   ث͸όχϥม׵ثͱಉ༷ͷੑೳΛୡ੒͠ɺඇৗʹ௕͍γʔέϯε
   ͷࣗݾճؼత༧ଌʹ͓͍ͯ࠷େ4000ഒͷ଎౓Λୡ੒ͨ͠ɻ
   http://arxiv.org/abs/2006.16236v2
   

   View Slide

7. View Slide

8. View Slide

9. ᶄ ҟৗݕग़ͷͨΊͷσΟʔϓϥʔχϯά.ϨϏϡʔ
   (ݪจ: Deep Learning for Anomaly Detection: A Review)
   ҟৗ஋ݕग़ɺผ໊ʮ֎Ε஋ݕग़ʯ͸ɺ਺े೥લ͔Β༷ʑͳݚڀίϛϡχςΟ
   ʹ͓͍ͯɺӬଓతͰ͋Γͳ͕Β΋׆ൃͳݚڀ෼໺ͱͳ͍ͬͯ·͢ɻ·ͩ·
   ͩɺߴ౓ͳΞϓϩʔνΛඞཁͱ͢Δಠಛͷ໰୊ͷෳࡶ͞ͱ՝୊͕͋Γ·
   ͢ɻۙ೥ɺσΟʔϓϥʔχϯάΛར༻ͨ͠ҟৗ஋ݕग़ɺ͢ͳΘͪσΟʔϓ
   ҟৗ஋ݕग़͕ॏཁͳํ޲ੑͱͯ͠ු্͖͍ͯͯ͠Δɻຊ࿦จͰ͸ɺਂ૚ҟ
   ৗݕ஌ͷݚڀΛɺ3ͭͷߴϨϕϧΧςΰϦͱ11ͷࡉ෼ԽΧςΰϦʹ෼͚ͯ
   ໢ཏతͳλΫιϊϛΛ༻͍ͯϨϏϡʔ͢ΔɻຊߘͰ͸ɺ͜ΕΒͷख๏ͷओ
   ཁͳ௚؍ɺ໨తؔ਺ɺجૅͱͳΔԾఆɺ௕ॴͱ୹ॴΛϨϏϡʔ͠ɺલड़ͷ
   ՝୊ʹͲͷΑ͏ʹରॲ͍ͯ͠Δ͔Λٞ࿦͢Δɻ͞ΒʹɺকདྷͷՄೳੑͱ՝
   ୊ʹରॲ͢ΔͨΊͷ৽ͨͳࢹ఺ʹ͍ͭͯ΋ٞ࿦͢Δɻ
   http://arxiv.org/abs/2007.02500v2
   

   View Slide

10. ܽؕݕग़ͷྫ
    (MVTec Adσʔληοτ)
    https://www.mvtec.com/company/research/datasets/mvtec-ad/
    

    View Slide

11. View Slide

12. ᶅNVAE: ਂ૚֊૚తม෼ࣗಈΤϯίʔμʔ
    (ݪจ: NVAE: A Deep Hierarchical Variational Autoencoder)
    ਖ਼نԽϑϩʔɺࣗݾճؼϞσϧɺม෼ࣗಈΤϯίʔμʔʢVAEʣɺσΟʔϓΤωϧΪʔϕʔεϞ
    σϧ͸ɺσΟʔϓੜ੒ֶशͷͨΊͷڝ߹͢Δ໬౓ϕʔεͷϑϨʔϜϫʔΫͷҰͭͰ͋Δɻ͜Ε
    ΒͷதͰ͸ɺߴ଎Ͱѻ͍΍͍͢αϯϓϦϯάͱΞΫηε͠΍͍͢ූ߸ԽωοτϫʔΫͷར఺͕
    ͋Δɻ͔͠͠ɺݱࡏͷͱ͜Ζɺਖ਼نԽϑϩʔ΍ࣗݾճؼϞσϧͷΑ͏ͳଞͷϞσϧʹ͸ྼͬͯ
    ͍ΔɻVAEͷݚڀͷେ෦෼͸౷ܭతͳ՝୊ʹয఺Λ౰͍ͯͯΔ͕ɺզʑ͸֊૚Խ͞ΕͨVAEͷ
    ͨΊͷχϡʔϥϧΞʔΩςΫνϟΛ৻ॏʹઃܭ͢Δͱ͍͏௚ަ͢Δํ޲ੑΛ୳Δɻզʑ͸ɺਂ
    ͞ํ޲ͷ෼཭Մೳͳ৞ΈࠐΈͱόονਖ਼نԽΛ༻͍ͯը૾ੜ੒ͷͨΊʹߏங͞Εͨਂ૚֊૚ܕ
    VAEͰ͋ΔNouveau VAEʢ̣̫̖̚ʣΛఏҊ͢ΔɻNVAE͸ਖ਼ن෼෍ͷ࢒ࠩύϥϝʔλԽΛඋ
    ͓͑ͯΓɺֶश͸εϖΫτϧਖ਼ଇԽʹΑͬͯ҆ఆԽ͞Ε͍ͯΔɻMNIST, CIFAR-10, CelebA HQ
    ͷσʔληοτʹ͓͍ͯɺNVAE͕ඇࣗݾճؼత໬౓ϞσϧͷதͰ࠷ઌ୺ͷ݁ՌΛୡ੒͠ɺ
    FFHQͷڧྗͳϕʔεϥΠϯΛఏڙ͢Δ͜ͱΛࣔ͢ɻྫ͑͹ɺCIFAR-10Ͱ͸ɺNVAE͸1࣍ݩ͋
    ͨΓͷϏοτ਺Λ2.98͔Β2.91ʹԡ্͛͠ɺCelebA HQͰ͸ਤ1ʹࣔ͢Α͏ʹߴ඼࣭ͷը૾Λ
    ੜ੒͠·͢ɻࢲͨͪͷ஌ΔݶΓͰ͸ɺNVAE͸ɺ256x256ϐΫηϧͷେ͖͞ͷࣗવը૾ʹద༻
    ͞Εͨ࠷ॳͷ੒ޭͨ͠VAEͰ͢ɻ
    http://arxiv.org/abs/2007.03898v1
    

    View Slide

13. ෮शɿVAE
    https://qiita.com/shionhonda/items/e2cf9fe93ae1034dd771
    

    View Slide

14. ෮श:VAE(2)
    • ಛ௃ZΛͣΒ͍ͯ͘͠ͱ݁Ռ͕มΘ͍ͬͯ͘
    

    View Slide

15. NVAEͷੜ੒ྫ
    

    View Slide

16. ˡͷS͸3FTJEVSBM/FUXPSL
    ˣ3FTJEVSBM/FUXPSLͷߏ଄
    

    View Slide

17. ᶆϞσϧϕʔεͷڧԽֶश.ௐࠪ
    (ݪจ: Model-based Reinforcement Learning: A Survey)
    ϚϧίϑܾఆϓϩηεʢMDPʣ࠷దԽͱͯ͠ҰൠతʹܗࣜԽ͞Ε͍ͯΔஞ࣍తҙࢥܾఆ
    ͸ɺਓ޻஌ೳͷॏཁͳ՝୊Ͱ͋Δɻ͜ͷ໰୊ʹର͢Δ2ͭͷॏཁͳΞϓϩʔν͸ɺڧԽֶ
    श(RL)ͱϓϥϯχϯάͰ͋ΔɻຊߘͰ͸ɺϞσϧϕʔεͷڧԽֶशͱͯ͠ΑΓΑ͘஌ΒΕ
    ͍ͯΔ྆෼໺ͷ౷߹ʹ͍ͭͯͷௐࠪΛ঺հ͢ΔɻϞσϧϕʔεRLʹ͸2ͭͷओཁͳεςο
    ϓ͕͋ΔɻୈҰʹɺ֬཰ੑɺෆ࣮֬ੑɺ෦෼తͳ؍ଌՄೳੑɺ࣌ؒతͳந৅ԽͳͲͷ՝୊
    ΛؚΉྗֶϞσϧֶश΁ͷΞϓϩʔνΛମܥతʹѻ͏ɻୈೋʹɺܭըͱֶशͷ౷߹ͷମܥ
    తͳ෼ྨΛఏࣔ͠ɺͲ͔͜ΒܭըΛ։࢝͢Δ͔ɺܭըͱ࣮σʔλऩूʹͲͷΑ͏ͳ༧ࢉΛ
    ׂΓ౰ͯΔ͔ɺͲͷΑ͏ʹܭը͢Δ͔ɺֶशͱߦಈͷϧʔϓʹܭըΛͲͷΑ͏ʹ౷߹͢Δ
    ͔ɺͳͲͷଆ໘ΛؚΉɻ͜ΕΒ2ͭͷॏཁͳηΫγϣϯͷޙʹ͸ɺσʔλޮ཰ͷ޲্ɺ
    λʔήοτΛߜͬͨ୳ࡧɺ҆ఆੑͷ޲্ͳͲɺϞσϧϕʔεͷRLͷજࡏతͳར఺ʹ͍ͭͯ
    ΋ٞ࿦͢ΔɻௐࠪʹԊͬͯɺ֊૚తRL΍఻ୡͳͲͷؔ࿈͢ΔRL෼໺΍ɺߦಈ৺ཧֶͳͲͷ
    ଞͷݚڀ෼໺ͱͷؔ࿈ੑ΋͍ࣔͯ͠·͢ɻશମͱͯ͠ɺຊௐࠪͰ͸ɺMDP࠷దԽͷͨΊ
    ͷܭըֶशͷ૊Έ߹Θͤʹ͍ͭͯɺ෯޿͍֓೦తͳ֓ཁΛఏ͍ࣔͯ͠Δɻ
    http://arxiv.org/abs/2006.16712v2
    

    View Slide

18. ෮श:MDP(Ϛϧίϑܾఆաఔ)
    https://ja.wikipedia.org/wiki/%E3%83%9E%E3%83%AB%E3%82%B3%E3%83%95%E6%B1%BA%E5%AE%9A%E9%81%8E%E7%A8%8B
    4ঢ়ଶ
    "ߦಈ
    

    View Slide

19. View Slide

20. View Slide

21. ᶇ χϡʔϥϧωοτϫʔΫͷάϥϑߏ଄
    (ݪจ: Graph Structure of Neural Networks)
    χϡʔϥϧωοτϫʔΫ͸ɺχϡʔϩϯؒͷ઀ଓͷάϥϑͱͯ͠දݱ͞ΕΔ͜ͱ͕
    ଟ͍ɻ͔͠͠ɺ޿͘࢖ΘΕ͍ͯΔʹ΋͔͔ΘΒͣɺχϡʔϥϧωοτϫʔΫͷάϥ
    ϑߏ଄ͱ༧ଌੑೳͱͷؔ܎ʹ͍ͭͯ͸ɺ΄ͱΜͲཧղ͞Ε͍ͯͳ͍ͷ͕ݱঢ়Ͱ͢ɻ
    ͜͜Ͱ͸ɺχϡʔϥϧωοτϫʔΫͷάϥϑߏ଄͕༧ଌੑೳʹͲͷΑ͏ͳӨڹΛ༩
    ͑Δͷ͔Λܥ౷తʹௐ΂Δɻ͜ͷ໨తͷͨΊʹɺզʑ͸ؔ܎άϥϑͱݺ͹ΕΔ
    χϡʔϥϧωοτϫʔΫͷ৽͍͠άϥϑϕʔεͷදݱΛ։ൃͨ͠ɻ͜ͷදݱΛ༻͍
    ͯɺҎԼͷ͜ͱΛࣔ͢ɻ(2)χϡʔϥϧωοτϫʔΫͷੑೳ͸ɺͦͷؔ܎άϥϑͷΫ
    ϥελϦϯά܎਺ͱฏۉύε௕ͷ׈Β͔ͳؔ਺Ͱ͋Δ͜ͱɺ(3)զʑͷ஌ݟ͸ଟ͘ͷ
    ҟͳΔλεΫ΍σʔληοτʹ͓͍ͯҰ؏͍ͯ͠Δ͜ͱɺ(4)εΠʔτεϙοτ͸ޮ
    ཰తʹಛఆͰ͖Δ͜ͱɺ(5)τοϓύϑΥʔϚϯεͷχϡʔϥϧωοτϫʔΫ͸ɺ࣮
    ࡍͷੜ෺ֶతχϡʔϥϧωοτϫʔΫͷͦΕʹڻ͘΄Ͳࣅͨάϥϑߏ଄Λ͍࣋ͬͯ
    Δ͜ͱΛࣔ͢ɻզʑͷݚڀ͸ɺχϡʔϥϧΞʔΩςΫνϟͷઃܭͱχϡʔϥϧωο
    τϫʔΫҰൠͷཧղʹ৽ͨͳํ޲ੑΛ։͘΋ͷͰ͋Δɻ
    http://arxiv.org/abs/2007.06559v1
    

    View Slide

22. άϥϑߏ଄
    

    View Slide

23. View Slide

24. ᶈ PyTorch3DʹΑΔ3DσΟʔϓϥʔχϯάͷߴ଎Խ
    (ݪจ: Accelerating 3D Deep Learning with PyTorch3D)
    http://arxiv.org/abs/2007.08501v1
    ˒1JDL6Q
    

    View Slide

25. ᶉ ϥϕϧԽ͞Ε͍ͯͳ͍σʔλ͸͢΂ͯ౳͍͠Θ͚Ͱ͸ͳ͍ɻ൒ڭࢣ෇ֶ͖श
    ʹ͓͚ΔσʔλͷॏΈ෇ֶ͚श
    (ݪจ: Not All Unlabeled Data are Equal: Learning to Weight Data in
    Semi-supervised Learning)
    طଘͷ൒ڭࢣ෇ֶ͖शʢSSLʣΞϧΰϦζϜ͸ɺϥϕϧ෇͖ྫͱϥϕϧ
    ͳ͠ྫͷଛࣦͷόϥϯεΛͱΔͨΊʹ୯ҰͷॏΈΛ࢖༻͍ͯ͠·͢ɺ͢
    ͳΘͪɺ͢΂ͯͷϥϕϧͳ͠ྫ͕౳͘͠ॏΈ෇͚͞Ε͍ͯ·͢ɻ͔͠
    ͠ɺϥϕϧ෇͚͞Ε͍ͯͳ͍σʔλ͕͢΂ͯ౳͘͠ͳΔΘ͚Ͱ͸ͳ
    ͍ɻ͜ͷ࿦จͰ͸ɺϥϕϧ෇͚͞Ε͍ͯͳ͍ྫ͝ͱʹҟͳΔॏΈΛ࢖
    ༻͢Δํ๏Λݚڀ͍ͯ͠·͢ɻ͜Ε·ͰͷݚڀͰߦΘΕ͍ͯͨΑ͏ͳɺ
    ͢΂ͯͷॏΈΛखಈͰௐ੔͢Δ͜ͱ͸΋͸΍ෆՄೳͰ͋Δɻͦͷ୅Θ
    Γʹɺզʑ͸Өڹؔ਺ʹج͍ͮͨΞϧΰϦζϜΛ༻͍ͯॏΈΛௐ੔͢
    Δɻ͜ͷΞϓϩʔνΛޮ཰తʹ͢ΔͨΊʹɺզʑ͸Өڹؔ਺ͷߴ଎Ͱ
    ޮՌతͳۙࣅΛఏҊ͢Δɻ͜ͷख๏͕ɺ൒ڭࢣ෇͖ը૾͓Αͼݴޠ෼
    ྨλεΫʹ͓͍ͯɺ࠷ઌ୺ͷख๏ΑΓ΋༏Ε͍ͯΔ͜ͱΛ࣮ূ͢Δɻ
    http://arxiv.org/abs/2007.01293v1
    

    View Slide

26. ᶃ ᶄ ᶅ
    

    View Slide

27. ᶊϦΧϨϯτχϡʔϥϧωοτϫʔΫʹ͓͚Δτοϓμ΢ϯ৴߸ͱϘτϜΞοϓ
    ৴߸ͷ૊Έ߹ΘͤͷֶशͱϞδϡʔϧ্Ͱͷ஫ҙשى
    (ݪจ: Learning to Combine Top-Down and Bottom-Up Signals in Recurrent
    Neural
    ݎ࿚ͳ஌֮͸ɺϘτϜΞοϓͱτοϓμ΢ϯͷ྆ํͷ৴߸ʹґଘ͍ͯ͠·͢ɻϘτϜΞοϓ
    ৴߸͸ɺײ֮Λ௨ͯ͠௚઀؍࡯͞ΕΔ΋ͷͰ͋Δɻτοϓμ΢ϯͷγάφϧ͸ɺաڈͷܦݧ
    ΍୹ظهԱʹجͮ͘৴೦΍ظ଴ɺྫ͑͹ʮϐʔφοπόλʔͱʙ...ʯͱ͍͏ϑϨʔζ͕Ͳͷ
    Α͏ʹ׬੒͢Δ͔ͱ͍͏Α͏ͳ΋ͷͰ͢ɻϘτϜΞοϓ৘ใͱτοϓμ΢ϯ৘ใͷ࠷దͳ૊
    Έ߹Θͤ͸ະղܾͷ໰୊Ͱ͋Δ͕ɺͦͷ૊Έ߹Θͤํ͸ಈతͰ͋Γɺจ຺΍λεΫʹґଘ͠
    ͍ͯͳ͚Ε͹ͳΒͳ͍ɻར༻Մೳͳજࡏతͳτοϓμ΢ϯ৘ใͷ๛෋͞ΛޮՌతʹར༻͠ɺ
    ૒ํ޲ΞʔΩςΫνϟͰͷ৴߸ͷࠞ͟Γ߹͍ͷෆڠ࿨ԻΛ๷͙ͨΊʹ͸ɺ৘ใͷྲྀΕΛ੍ݶ
    ͢ΔϝΧχζϜ͕ඞཁͰ͋Δɻզʑ͸ɺϘτϜΞοϓͱτοϓμ΢ϯͷ৴߸͕஫ҙΛ࢖ͬͯ
    ಈతʹ݁߹͞ΕΔσΟʔϓϦΧϨϯτχϡʔϥϧωοτΞʔΩςΫνϟΛ୳ٻ͍ͯ͠ΔɻΞʔ
    ΩςΫνϟͷϞδϡʔϧੑ͸ɺ৘ใͷڞ༗ͱ௨৴Λ͞Βʹ੍ݶ͢Δɻ஫ҙͱϞδϡʔϧੑ͸
    ৘ใͷྲྀΕΛ༠ಋ͠ɺ஌֮ͱݴޠλεΫʹ͓͚Δ৴པੑͷߴ͍ੑೳ޲্Λ΋ͨΒ͠ɺಛʹ஫
    ҙࢄອ΍ϊΠζͷଟ͍σʔλʹର͢ΔϩόετੑΛ޲্ͤ͞ΔɻຊݚڀͰ͸ɺݴޠϞσϦϯ
    άɺஞ࣍ը૾෼ྨɺϏσΦ༧ଌɺڧԽֶशͳͲͷ༷ʑͳϕϯνϚʔΫʹ͓͍ͯɺʮ૒ํ޲ੑʯ
    ৘ใͷྲྀΕ͕ڧྗͳϕʔεϥΠϯΑΓ΋݁ՌΛ޲্ͤ͞Δ͜ͱΛ࣮ূͨ͠ɻ
    http://arxiv.org/abs/2006.16981v1
    

    View Slide

28. Bidirectional Recurrent
    Independent Mechanisms (BRIMs)
    

    View Slide

29. 47(-14UPDIBTUJD7JEFP(FOFSBUJPOXJUIB-FBSOFE1SJPS
    IUUQTBSYJWPSHQEGQEG
    

    View Slide

30. View Slide

31. ᶋ ϩόετੑͷଟ໘ੑɻ෼෍֎ҰൠԽͷ൷൑త෼ੳ
    (ݪจ: The Many Faces of Robustness: A Critical Analysis
    of Out-of-Distribution Generalization)
    ຊݚڀͰ͸ɺը૾ελΠϧɺ஍ཧతҐஔɺΧϝϥૢ࡞ͳͲͷࣗવൃੜత
    ͳ෼෍มԽ͔ΒͳΔ3ͭͷ৽͍͠ϩόετੑϕϯνϚʔΫΛ঺հ͢Δɻ͜
    ͷϕϯνϚʔΫΛ༻͍ͯɺҎલʹఏҊ͞Εͨ෼෍֎ϩόετੑʹؔ͢Δ
    ԾઆΛݕূ͠ɺͦΕΒΛݕূ͠·͢ɻ͜Ε·ͰͷݚڀͰͷओுʹ൓ͯ͠ɺ
    ΑΓେ͖ͳϞσϧͱ߹੒σʔλͷ૿ڧΛ࢖༻͢Δ͜ͱͰɺ࣮ੈքͷ෼෍
    γϑτʹର͢Δϩόετੑ͕վળ͞ΕΔ͜ͱ͕Θ͔Γ·ͨ͠ɻ͜ΕΛ
    ͖͔͚ͬʹɺզʑ͸࠷ઌ୺ͷٕज़Λਐาͤ͞ɺ1000ഒҎ্ͷϥϕϧ෇͚
    ͞ΕͨσʔλͰࣄલֶश͞ΕͨϞσϧΛ྇կ͢Δ৽͍͠σʔλ૿ڧ๏Λ
    ಋೖ͠·ͨ͠ɻͦͷ݁Ռɺ͍͔ͭ͘ͷख๏͸ςΫενϟ΍ہॴతͳը૾
    ౷ܭʹ͓͚Δ෼෍ͷมԽʹ͸Ұ؏ͯ͠༗ޮͰ͋Δ͕ɺ஍ཧతมԽͷΑ͏
    ͳଞͷ෼෍ͷมԽʹ͸༗ޮͰͳ͍͜ͱ͕Θ͔ͬͨɻࠓޙͷݚڀͰ͸ɺෳ
    ਺ͷ෼෍γϑτΛಉ࣌ʹݚڀ͠ͳ͚Ε͹ͳΒͳ͍ͱ݁࿦͚͍ͮͯ·͢ɻ
    http://arxiv.org/abs/2006.16241v1
    

    View Slide

32. ৽͍͠ϩόετωεࢦඪͷ঺հ
    

    View Slide

33. ImageNet-R
    ࣮෺͚ͩͰͳ͘૑࡞෺΋ؚΉ
    

    View Slide

34. • DeepFassion Remixed
    ద౰ͳΧϝϥઃఆʢζʔϜɾαΠζɾϑΥʔΧεͳ
    ͲʣͰࡱӨͨ͠ෳ਺ͷҥྉ඼ͷϚϧνϥϕϧ෼ྨλ
    εΫͷͨΊͷσʔληοτɻ
    • SVSF
    • ετϦʔτϏϡʔͷళͷ֎؍Λళฮͷۀछ͝ͱʹ
    ෼ྨͨ͠σʔλ(ඇެ։)
    

    View Slide

35. طଘ࿦จͷ7ͭͷϩόετੑԾઆͷݕূ(Ұ෦ൈਮ):
    ʮDiverse Data Augmentation͸ࣗવͳϩόετੑΛॿ͚ͳ͍આʯ
    DeepAugumentͨ͠σʔλͰݕূ → ImageNet-R͚ͩޮՌత
    

    View Slide

36. ᶌਆܦϓϩάϥϜʹ͓͚ΔڧྗͳҰൠԽͱޮ཰ੑ
    (ݪจ: Strong Generalization and Efficiency in Neural
    Programs)
    ຊݚڀͰ͸ɺχϡʔϥϧϓϩάϥϜ༠ಋͷ࿮૊ΈͷதͰɺڧྗʹҰൠԽ͢Δޮ཰త
    ͳΞϧΰϦζϜΛֶश͢Δ໰୊Λݚڀ͍ͯ͠·͢ɻχϡʔϥϧϞσϧͷೖग़ྗΠϯ
    λϑΣʔεΛ৻ॏʹઃܭ͠ɺ໛฿͢Δ͜ͱͰɺ೚ҙͷೖྗαΠζʹରͯ͠ਖ਼͍݁͠
    ՌΛग़͢ϞσϧΛֶश͠ɺڧྗͳҰൠԽΛ࣮ݱ͢Δɻ·ͨɺڧԽֶशΛར༻͢Δ͜
    ͱͰɺϓϩάϥϜͷޮ཰ࢦඪΛ࠷దԽ͠ɺ໛฿Ͱ࢖༻ͨ͠ڭࢣΛ௒͑Δ৽͍͠Ξϧ
    ΰϦζϜΛൃݟ͢Δ͜ͱ͕Ͱ͖Δɻ͜ΕʹΑΓɺզʑͷΞϓϩʔν͸ɺιʔτɺॱ
    ং෇͖ϦετͰͷݕࡧɺNP׬શͳ0/1φοϓβοΫ໰୊Ͱςετͨ͠Α͏ʹɺ༷ʑ
    ͳ໰୊ʹରͯ͠ΧελϜهड़͞ΕͨղΛ্ճΔੑೳΛֶश͢Δ͜ͱ͕Ͱ͖ɺχϡʔ
    ϥϧϓϩάϥϜ༠ಋͷ෼໺Ͱ஫໨͢΂͖ϚΠϧετʔϯΛઃఆ͍ͯ͠ΔɻϋΠϥΠ
    τͱͯ͠ɺզʑͷֶशϞσϧ͸ɺզʑ͕ςετͨ͠ͲͷΑ͏ͳೖྗσʔλαΠζͰ
    ΋ɺO(n log n)ͷෳࡶ͞ͰιʔτΛ׬ᘳʹ࣮ߦ͢Δ͜ͱ͕Ͱ͖ɺҰํͰɺֶशதʹ
    ݟΒΕͨϦεταΠζΛ͸Δ͔ʹ௒͑ΔϦεταΠζͰ΋ɺΫΠοΫιʔτΛؚΉ
    ϋϯυίʔυԽ͞ΕͨΞϧΰϦζϜΛૢ࡞ճ਺Ͱ্ճΔ͜ͱ͕Ͱ͖·ͨ͠ɻ
    http://arxiv.org/abs/2007.03629v2
    

    View Slide

37. a=f(s) ͷ f ΛҰൠԽֶͭͭ͠श͢Δ
    • ྫ) a=ιʔτ݁Ռσʔλɺs=ιʔτର৅σʔλ
    G
    
    T
    B
    ֶश
    

    View Slide

38. Arxiv Sanity
    Top hype: Best10
    

    View Slide

39. ᶃ GShard: ৚݅෇͖ܭࢉͱࣗಈγϟʔσΟϯάΛ༻͍ͨڊେϞσϧͷεέʔϦ
    ϯά
    (ݪจ: GShard: Scaling Giant Models with Conditional Computation and
    Automatic Sharding)
    χϡʔϥϧωοτϫʔΫͷεέʔϦϯά͸ɺ๲େͳྔͷֶशσʔλͱܭࢉྔΛ࣋ͭଟ͘
    ͷ࣮ੈքͷػցֶशΞϓϦέʔγϣϯʹ͓͍ͯɺϞσϧͷ඼࣭Λ޲্ͤ͞ΔͨΊʹॏཁ
    ͳ໾ׂΛՌ͖ͨͯ͠·ͨ͠ɻ͜ͷΑ͏ͳεέʔϦϯάͷ܏޲͸ɺϞσϧ඼࣭Λ޲্ͤ͞
    ΔͨΊͷ࣮֬ͳΞϓϩʔνͰ͋Δ͜ͱ͕֬ೝ͞Ε͍ͯ·͕͢ɺͦͷಓͷΓʹ͸ɺܭࢉί
    ετɺϓϩάϥϛϯάͷ༰қ͞ɺฒྻσόΠε্Ͱͷޮ཰తͳ࣮૷ͳͲͷ՝୊͕͋Γ·
    ͢ɻGShard͸ɺܰྔͳΞϊςʔγϣϯAPIͱXLAίϯύΠϥͷ֦ுػೳ͔ΒͳΔϞ
    δϡʔϧͰ͢ɻGShard͸ɺطଘͷϞσϧίʔυΛ࠷খݶʹมߋ͢Δ͚ͩͰɺ෯޿͍ฒྻ
    ܭࢉύλʔϯΛදݱ͢ΔΤϨΨϯτͳํ๏Λఏڙ͠·͢ɻGShard͸ɺࣗಈγϟʔσΟϯ
    άΛ༻͍ͯɺεύʔεϦʔɾήʔςουɾϛοΫενϟʔɾΦϒɾΤΩεύʔτʹΑΔ
    ଟݴޠχϡʔϥϧػց຋༁τϥϯεϑΥʔϚʔϞσϧΛ6000ԯݸͷύϥϝʔλΛ௒͑ͯ
    εέʔϧΞοϓ͢Δ͜ͱΛՄೳʹ͠·ͨ͠ɻ͜ͷΑ͏ͳڊେͳϞσϧΛ2048TPU v3Ξ
    ΫηϥϨʔλ্Ͱ4೔ؒͰޮ཰తʹֶश͠ɺ100ݴޠ͔Βӳޠ΁ͷ຋༁ʹ͓͍ͯɺैདྷ
    ͷٕज़ͱൺֱͯ͠͸Δ͔ʹ༏Εͨ඼࣭Λୡ੒Ͱ͖Δ͜ͱΛ࣮ূ͠·ͨ͠ɻ
    http://arxiv.org/abs/2006.16668v1
    

    View Slide

40. ᶄᷖճతMCMCɿ౷ҰతͳϑϨʔϜϫʔΫ
    (ݪจ: Involutive MCMC: a Unifying Framework)
    Ϛϧίϑ࿈࠯ϞϯςΧϧϩ๏ʢMCMCʣ͸ɺਪ࿦ɺ౷߹ɺ࠷దԽɺγϛϡϨʔ
    γϣϯͳͲͷجຊతͳ໰୊ʹର͢ΔܭࢉΞϓϩʔνͰ͢ɻ͜ͷ෼໺Ͱ͸ɺ͞·
    ͟·ͳΞϧΰϦζϜ͕։ൃ͞Ε͓ͯΓɺͦͷಈػɺద༻ํ๏ɺαϯϓϦϯάͷ
    ޮ཰ੑͳͲ͕ҟͳΔɻ͢΂ͯͷҧ͍ʹ΋͔͔ΘΒͣɺͦΕΒͷଟ͘͸ಉ͡ίΞ
    ݪཧΛڞ༗͓ͯ͠Γɺզʑ͸ͦΕΛInvolutive MCMC (iMCMC)ϑϨʔϜϫʔ
    Ϋͱͯ͠౷Ұ͍ͯ͠·͢ɻ͜Εʹج͍ͮͯɺզʑ͸iMCMCͷ؍఺͔Β޿ൣғͷ
    MCMCΞϧΰϦζϜΛهड़͠ɺ৽͍͠MCMCΞϧΰϦζϜΛ։ൃ͢ΔͨΊͷઃ
    ܭݪཧͱͯ͠࢖༻Ͱ͖Δ͍͔ͭ͘ͷʮτϦοΫʯΛఆࣜԽ͢Δɻ͜ͷΑ͏ʹɺ
    iMCMC͸ଟ͘ͷط஌ͷMCMCΞϧΰϦζϜΛ౷ҰతʹݟΔ͜ͱ͕Ͱ͖ɺڧྗ
    ͳ֦ுػೳͷ։ൃΛ༰қʹ͠·͢ɻզʑ͸ɺط஌ͷՄٯMCMCΞϧΰϦζϜΛ
    ΑΓޮ཰తͳෆՄٯMCMCΞϧΰϦζϜʹม׵͢Δ2ͭͷྫΛ༻͍ͯɺޙऀΛ
    ࣮ূ͠·͢ɻ
    http://arxiv.org/abs/2006.16653v1
    

    View Slide

41. ෮शɿMCMC?
    • Ϛϧίϑ࿈࠯
    • લͷঢ়ଶ͔Β࣍ͷঢ়ଶ͕ܾ·Δ
    • ϞϯςΧϧϩ๏
    • ֬཰తΞϧΰϦζϜɻ
    • ϥϯμϜʹσʔλΛूΊͨΒɺਖ਼͍͠ॴʹσʔ
    λ͕ଟ͘ͳΔɻ
    

    View Slide

42. https://qiita.com/kenmatsu4/items/55e78cc7a5ae2756f9da
    

    View Slide

43. iMCMC ͰఆࣜԽ͞Εͨ Trickྫ
    σʔλ෼෍ೱ౓͕ೱ͍ʢ֬཰͕ߴ͍ʣͱ͜Ζ͔Βബ
    ͍ʢ௿͍ʣͱ͜Ζ΁ͷҠಈΛڋ൱͢Δ
    αϯϓϧιʔε:
    https://github.com/necludov/iMCMC
    

    View Slide

44. ᶅ ݴޠϞσϧ͸গ਺ਫ਼Ӷͷֶशऀ
    (ݪจ: Language Models are Few-Shot Learners)
    ࠷ۙͷݚڀͰ͸ɺେن໛ͳςΩετͷίʔύεͰࣄલֶशΛߦͬͨޙɺಛఆͷλεΫͰඍௐ੔Λߦ͏͜ͱͰɺ
    ଟ͘ͷNLPλεΫ΍ϕϯνϚʔΫͰେ͖ͳ੒Ռ͕ಘΒΕ͍ͯΔ͜ͱ͕࣮ূ͞Ε͍ͯΔɻΞʔΩςΫνϟతʹ͸
    λεΫʹͱΒΘΕͳ͍ͷ͕ҰൠతͰ͕͢ɺ͜ͷํ๏Ͱ͸਺ઍ͔Β਺ສͷྫ୊ͷλεΫݻ༗ͷඍௐ੔σʔληο
    τ͕ඞཁͱͳΓ·͢ɻରরతʹɺਓؒ͸Ұൠతʹɺ৽͍͠ݴޠλεΫΛ਺ݸͷྫ΍؆୯ͳ໋ྩ͔Β࣮ߦ͢Δ͜
    ͱ͕Ͱ͖·͕͢ɺ͜Ε͸ݱࡏͷNLPγεςϜͰ͸͍·ͩʹࠔ೉ͳ͜ͱͰ͢ɻ͜͜Ͱզʑ͸ɺݴޠϞσϧΛε
    έʔϧΞοϓ͢Δ͜ͱͰɺλεΫʹґଘ͠ͳ͍ɺ਺γϣοτͷੑೳΛେ෯ʹ޲্ͤ͞ɺ࣌ʹ͸ઌߦ͢Δ࠷ઌ୺
    ͷඍௐ੔Ξϓϩʔνʹඖఢ͢Δੑೳʹୡ͢Δ͜ͱΛ͍ࣔͯ͠Δɻ۩ମతʹ͸ɺ1,750ԯݸͷύϥϝʔλΛ࣋ͭ
    ࣗݾճؼతݴޠϞσϧͰ͋ΔGPT-3Λֶशͤ͞ɺͦͷੑೳΛ਺γϣοτͷઃఆͰςετͨ͠ɻ͢΂ͯͷλεΫ
    ʹ͓͍ͯɺGPT-3͸ޯ഑ͷߋ৽΍ඍௐ੔ΛҰ੾ߦΘͣʹద༻͞ΕɺλεΫͱ਺γϣοτͷσϞ͸Ϟσϧͱͷς
    ΩετΠϯλϥΫγϣϯͷΈͰࢦఆ͞ΕͨɻGPT-3͸ɺ຋༁ɺ࣭໰Ԡ౴ɺΫϩʔδϯάͳͲͷଟ͘ͷNLPσʔ
    ληοτʹՃ͑ͯɺ୯ޠͷεΫϥϯϒϧղআɺจதͷ৽͍͠୯ޠͷ࢖༻ɺ3ܻͷԋࢉͳͲɺͦͷ৔Ͱͷਪ࿦΍
    ྖҬదԠΛඞཁͱ͢Δ͍͔ͭ͘ͷλεΫʹ͓͍ͯ΋ߴ͍ੑೳΛୡ੒͍ͯ͠Δɻಉ࣌ʹɺGPT-3ͷ਺ൃֶश͕ະ
    ͩʹۤઓ͍ͯ͠Δσʔλ΍ɺେن໛ͳ΢Σϒίʔύε্Ͱͷֶशʹؔ࿈ͯ͠ํ๏࿦తͳ໰୊Λ๊͍͑ͯΔσʔ
    λΛ͍͔ͭ͘ڍ͛ͨɻ࠷ޙʹɺGPT-3͸ਓ͕ؒॻ͍ͨهࣄͱਓ͕ؒॻ͍ͨهࣄΛ۠ผ͢Δͷ͕೉͍͠χϡʔε
    هࣄͷαϯϓϧΛੜ੒͢Δ͜ͱ͕Ͱ͖Δ͜ͱΛൃݟͨ͠ɻຊݚڀͰ͸ɺ͜ͷൃݟͱGPT-3ͷҰൠతͳࣾձతӨ
    ڹʹ͍ͭͯٞ࿦͢Δɻ
    http://arxiv.org/abs/2005.14165v4
    ઌ݄ɾઌʑ݄ͱॏෳ
    

    View Slide

45. ᶆઐ໳Ոͱͯ͠ͷࣄ࣮ɻ৅௃తͳ஌ࣝͷ্ʹదԠՄೳͰղऍՄೳͳਆܦه
    Ա
    (ݪจ: Facts as Experts: Adaptable and Interpretable Neural Memory
    over Symbolic Knowledge)
    େن໛ͳݴޠϞσϧ͸ݱ୅ͷNLPϞσϦϯάͷத֩Ͱ͋Γɺ๲େͳྔͷৗ
    ࣝతͰࣄ࣮ʹج͍ͮͨ৘ใΛΤϯίʔυ͢Δ͜ͱ͕ࣔ͞Ε͍ͯ·͢ɻ͔͠
    ͠ɺͦͷ஌ࣝ͸Ϟσϧͷજࡏతͳύϥϝʔλͷதʹ͔͠ଘࡏͤͣɺݕࠪ΍
    ղऍʹΞΫηε͢Δ͜ͱ͸Ͱ͖·ͤΜɻ·ͨɺύϥϝʔλͱͯ͠อଘ͞Ε
    ͨ஌ࣝ͸ɺඞવతʹݪࢿྉʹ಺ࡏ͢Δ͢΂ͯͷόΠΞεΛࣔ͢͜ͱʹͳ
    Δɻ͜ΕΒͷ໰୊ʹରॲ͢ΔͨΊʹɺզʑ͸ɺه߸తʹղऍՄೳͳࣄ࣮৘
    ใͱαϒγϯϘϧతͳਆܦ஌ࣝͱͷؒͷ໌ࣔతͳΠϯλʔϑΣʔεΛؚΉ
    ਆܦݴޠϞσϧΛ։ൃͨ͠ɻ͜ͷϞσϧ͕ɺ஌ࣝू໿తͳ2ͭͷ࣭໰Ԡ౴
    λεΫͷύϑΥʔϚϯεΛܶతʹ޲্ͤ͞Δ͜ͱΛࣔ͢ɻ͞Βʹڵຯਂ͍
    ͜ͱʹɺ͜ͷϞσϧ͸ɺͦͷه߸දݱΛૢ࡞͢Δ͜ͱͰɺ࠶܇࿅ͳ͠ʹߋ
    ৽͢Δ͜ͱ͕Ͱ͖Δɻಛʹ͜ͷϞσϧ͸ɺैདྷͷϞσϧͰ͸ෆՄೳͩͬͨ
    ৽͍͠ࣄ࣮ͷ௥Ճ΍طଘͷࣄ࣮ͷ্ॻ͖ΛՄೳʹ͢Δɻ
    http://arxiv.org/abs/2007.00849v1
    

    View Slide

46. Facts as Experts Ϟσϧߏ଄
    

    View Slide

47. ᶇFathomNet. ւ༸୳ࠪɾൃݟͷͨΊͷਫதը૾τϨʔχϯάσʔλϕʔε
    (ݪจ: FathomNet: An underwater image training database for ocean
    exploration and discovery)
    ԕִૢ࡞ंʢROVʣ΍ͦͷଞͷਫதࢿ࢈͔Β͸ɺ೥ؒ਺ઍ࣌ؒʹٴͿւ༸ϏσΦσʔλ͕ऩ
    ू͞Ε͍ͯ·͢ɻ͔͠͠ɺݱࡏͷख࡞ۀʹΑΔղੳํ๏Ͱ͸ɺROV΍େن໛ͳੜ෺ଟ༷ੑղ
    ੳͷͨΊͷϦΞϧλΠϜΞϧΰϦζϜͷͨΊʹऩूͨ͠σʔλΛे෼ʹ׆༻͢Δ͜ͱ͕Ͱ͖
    ·ͤΜɻFathomNet͸ɺ࠷৽ͷΠϯςϦδΣϯτͰࣗಈԽ͞Εͨਫதը૾ղੳͷ։ൃΛՃ଎
    ͢ΔͨΊʹ࠷దԽ͞Εͨɺ৽͍͠ϕʔεϥΠϯը૾τϨʔχϯάηοτͰ͢ɻࢲͨͪͷγʔ
    υσʔληοτ͸ɺ26,000࣌ؒҎ্ͷϏσΦςʔϓɺ680ສճͷΞϊςʔγϣϯɺ4,349ޠͷ
    ஌ࣝϕʔε͔ΒͳΔɺઐ໳ՈʹΑΔ஫ऍ෇͖ͷܧଓతͳσʔλϕʔε͔Βߏ੒͞Ε͍ͯ·͢ɻ
    FathomNet͸ɺ͜ͷσʔληοτΛ׆༻ͯ͠ɺػցֶशΞϧΰϦζϜͷ։ൃΛՄೳʹ͢Δͨ
    Ίʹɺਫதͷ֓೦ͷը૾ɺϩʔΧϦθʔγϣϯɺΫϥεϥϕϧΛఏڙ͍ͯ͠·͢ɻݱࡏ·Ͱ
    ʹɺதਫҬ΍ఈੜੜ෺ΛؚΉ233ͷҟͳΔΫϥεʹ͍ͭͯɺ80,000Ҏ্ͷը૾ͱ106,000Ҏ
    ্ͷϩʔΧϥΠθʔγϣϯ͕ఏڙ͞Ε͍ͯ·͢ɻզʑͷ࣮ݧͰ͸ɺऑڭࢣ෇͖ఆҐɺը૾ϥ
    ϕϦϯάɺ෺ମݕग़ɺ෼ྨͳͲͷΞϓϩʔνΛ༻͍ͯɺ༷ʑͳσΟʔϓϥʔχϯάΞϧΰϦζ
    ϜΛֶश͕ͤͨ͞ɺ͜ΕΒ͸༗๬Ͱ͋Δ͜ͱ͕Θ͔ͬͨɻ͜ͷ৽͍͠σʔληοτͰͷ༧ଌ
    ݁Ռ͸ྑ޷Ͱ͕͋ͬͨɺ࠷ऴతʹ͸ւ༸୳ࠪͷͨΊͷΑΓେ͖ͳσʔληοτ͕ඞཁͰ͋Δ
    ͜ͱΛ͍ࣔͯ͠Δɻ
    IUUQTBSYJWPSHBCTW
    

    View Slide

48. View Slide

49. View Slide

50. ᶈBlazePose.ΦϯσόΠεͰϦΞϧλΠϜʹମͷϙʔζΛτ
    ϥοΩϯά
    (ݪจ: BlazePose: On-device Real-time Body Pose tracking)
    զʑ͸ɺϞόΠϧσόΠε্ͰͷϦΞϧλΠϜਪ࿦ͷͨΊʹௐ੔͞Ε
    ͨɺਓؒͷϙʔζਪఆͷͨΊͷܰྔͳ৞ΈࠐΈχϡʔϥϧωοτϫʔ
    ΫΞʔΩςΫνϟͰ͋ΔBlazePoseΛ঺հ͠·͢ɻਪ࿦தɺ͜ͷωο
    τϫʔΫ͸Ұਓͷਓؒʹରͯ͠33ͷ਎ମΩʔϙΠϯτΛੜ੒͠ɺ
    Pixel 2ܞଳి࿩্Ͱຖඵ30ϑϨʔϜҎ্ͷ଎౓Ͱಈ࡞͠·͢ɻ͜ͷ
    ͨΊɺϑΟοτωετϥοΩϯά΍ख࿩ೝࣝͷΑ͏ͳϦΞϧλΠϜͷ
    Ϣʔεέʔεʹಛʹద͍ͯ͠·͢ɻզʑͷओͳߩݙ͸ɺ৽͍͠ମ੎τ
    ϥοΩϯάιϦϡʔγϣϯͱɺώʔτϚοϓͱΩʔϙΠϯτ࠲ඪ΁ͷ
    ճؼͷ྆ํΛ࢖༻͢Δܰྔͳମ੎ਪఆχϡʔϥϧωοτϫʔΫͰ͢ɻ
    http://arxiv.org/abs/2006.10204v1
    

    View Slide

51. View Slide

52. View Slide

53. طଘݚڀͱͷൺֱ
    BlazePose vs OpenPose
    

    View Slide

54. Ϟσϧߏ଄ʢখ͍͞ʂʣ
    

    View Slide

55. ᶉີͳςΩετݕࡧͷͨΊͷۙࣅ࠷ۙ๣ෛͷରরֶश
    (ݪจ: Approximate Nearest Neighbor Negative
    Contrastive Learning for Dense Text Retrieval)
    ֶश͞ΕͨີͳදݱۭؒΛ༻͍ͯςΩετݕࡧΛߦ͏͜ͱ͸ɺεύʔεݕࡧʹൺ΂ͯ
    ଟ͘ͷڵຯਂ͍ར఺͕͋Δɻ͔͠͠ɺີͳֶशදݱۭؒͰͷςΩετݕࡧͷ༗ޮੑΛ
    ߴΊΔͨΊʹ͸ɺεύʔεݕࡧͱͷ૊Έ߹Θ͕ͤඞཁͱͳΔ͜ͱ͕ଟ͍ɻຊ࿦จͰ
    ͸ɺֶशϝΧχζϜʹϘτϧωοΫ͕͋Δ͜ͱΛ໌Β͔ʹ͠ɺֶशʹ࢖༻͞ΕΔෛͷ
    Πϯελϯε͕ςετͰ͸ແؔ܎ͳจॻΛ୅ද͍ͯ͠ͳ͍͜ͱΛ໌Β͔ʹͨ͠ɻຊ࿦
    จͰ͸ɺίʔύεͷۙࣅ࠷ۙ๣ʢANNʣΠϯσοΫε͔ΒωΨςΟϒΛߏங͠ɺֶ
    शϓϩηεͱฒߦͯ͠ߋ৽͢Δ͜ͱͰɺΑΓݱ࣮తͳωΨςΟϒֶशΠϯελϯεΛ
    બ୒͢ΔֶशϝΧχζϜͰ͋Δۙࣅ࠷ۙ๣ωΨςΟϒɾίϯτϥετਪఆʢANCEʣ
    ΛఏҊ͢Δɻ͜ΕʹΑΓɺDRͷֶशͱςετͰ࢖༻͞ΕΔσʔλ෼෍ͷෆҰக͕ࠜ
    ຊతʹղܾ͞ΕΔɻզʑͷ࣮ݧͰ͸ɺANCE͸BERT-Siamese DRϞσϧΛϒʔετ
    ͠ɺڝ߹͢Δີͳݕࡧͱૄͳݕࡧͷ͢΂ͯͷϕʔεϥΠϯΛ্ճΔੑೳΛࣔͨ͠ɻ
    ANCEͰֶशͨ͠දݱۭؒʹ͓͚ΔυοτϓϩμΫτΛ༻͍ͨεύʔεݕࡧ͓Αͼ
    BERT࠶ϥϯΩϯάͷਫ਼౓ͱ΄΅Ұக͠ɺ΄΅100ഒͷߴ଎ԽΛ࣮ݱͨ͠ɻ
    http://arxiv.org/abs/2007.00808v1
    

    View Slide

56. View Slide

57. ᶊ޿͍ϕΠζχϡʔϥϧωοτϫʔΫͷਖ਼֬ͳࣄޙ෼෍
    (ݪจ: Exact posterior distributions of wide Bayesian
    neural networks)
    ࠷ۙͷݚڀͰ͸ɺσΟʔϓϕΠδΞϯχϡʔϥϧωοτϫʔΫ(BNN)ʹΑͬ
    ͯ༠ى͞Εͨؔ਺ͷࣄલ෼෍͸ɺશ૚ͷ෯͕େ͖͘ͳΔʹͭΕͯΨ΢εա
    ఔ(GP)ͷΑ͏ʹৼΔ෣͏͜ͱ͕ࣔ͞Ε͍ͯΔɻ͔͠͠ɺଟ͘ͷBNNΞϓϦ
    έʔγϣϯͰ͸ɺBNNؔ਺ۭؒͷࣄޙॲཧ͕໰୊ͱͳ͍ͬͯ·͢ɻNeal
    (1996)΍MatthewsΒ(2018)ͷΦϦδφϧͷݚڀͰ͸ࣄޙऩଋͷ͍͔ͭ͘ͷ
    ܦݧతূڌ͕ఏڙ͞Ε͍ͯ·͕͢ɺͦΕ͸BNNࣄޙۙࣅͷਖ਼֬͞Λऔಘ͠
    ͯݕূ͢Δ͜ͱͷѱ໊ߴ͍ࠔ೉͞ͷͨΊʹɺখ͞ͳσʔληοτ΍ΞʔΩ
    ςΫνϟʹݶఆ͞Ε͍ͯ·͢ɻզʑ͸ɺਖ਼֬ͳBNNࣄޙۙࣅ͕ɺࣄલ෼෍
    ͷGPۃݶʹΑͬͯ༠ى͞ΕΔ΋ͷʹ(ऑ͘)ऩଋ͢Δͱ͍͏ܽམͨ͠ཧ࿦త
    ূ໌Λఏڙ͢Δɻ࣮ূతݕূͷͨΊʹɺখ͞ͳσʔληοτ্ͷ༗ݶBNN
    ͔ΒڋઈαϯϓϦϯάΛ༻͍ͯਖ਼֬ͳαϯϓϧΛੜ੒͢Δํ๏Λࣔ͢ɻ
    http://arxiv.org/abs/2006.10541v1
    

    View Slide

58. d͕େ͖͘ͳΔ΄Ͳແݶͷ෯ݶքͷࣄલॲཧͱಉ͡໬౓ͷࣄޙॲཧʹऩଋ͢Δ͜ͱΛূ໌͠
    ͨɻແݶ෯ݶքࣄޙॲཧͷܭࢉ͕༰қͰ͋Ε͹ɺύϥϝʔλۭؒࣄޙॲཧͷධՁ͕ࠔ೉Ͱ͋ͬ
    ͯ΋ɺؔ਺ۭؒਪ࿦͕༰қʹͳΔಓ͕։͚Δɻ
    

    View Slide

59. ᶋֶश͞Εͨදݱͷઢܗࣝผੑʹ͍ͭͯ
    (ݪจ: On Linear Identifiability of Learned
    Representations)
    ࣝผՄೳੑ͸౷ܭϞσϧͷ๬·͍͠ಛੑͰ͋Γɺे෼ͳܭࢉࢿݯͱσʔλ͕
    ͋Ε͹ɺਅͷϞσϧύϥϝʔλΛ೚ҙͷਫ਼౓ͰਪఆͰ͖Δ͜ͱΛҙຯ͠·͢ɻ
    զʑ͸දݱֶशͷจ຺ͰࣝผՄೳੑΛݚڀ͍ͯ͠·͢ɿԼྲྀͷλεΫʹؔͯ͠
    ࠷దͳඇઢܗσʔλදݱΛൃݟ͢Δ͜ͱͰ͢ɻσΟʔϓχϡʔϥϧωοτϫʔ
    Ϋͱͯ͠ύϥϝʔλԽ͞Εͨ৔߹ɺͦͷΑ͏ͳදݱؔ਺͸ɺઃܭ্ύϥϝʔλ
    ͕ա৒ʹઃఆ͞Ε͍ͯΔͨΊɺҰൠతʹύϥϝʔλۭؒͰͷࣝผੑΛ͍͍ܽͯ
    ·͢ɻ͜ͷ࿦จͰ͸ɺ࠷ۙͷඇઢܗICAͷਐาʹج͍ͮͯɺେن໛ͳࣝผϞσ
    ϧͷϑΝϛϦʔ͕ɺઢܗෆ֬ఆੑ·Ͱؔ਺ۭؒͰ࣮ࡍʹࣝผՄೳͰ͋Δ͜ͱ
    Λࣔ͢͜ͱʹΑͬͯɺࣝผՄೳੑΛճ෮ͤ͞Δ͜ͱΛ໨తͱ͍ͯ͠Δɻදݱֶ
    शͷͨΊͷଟ͘ͷϞσϧ͸ɺςΩετɺը૾ɺԻ੠ͳͲɺ͞·͟·ͳྖҬͰ͜
    ͷҙຯͰࣝผՄೳͰ͋Γɺൃද࣌ʹ͸࠷ઌ୺ͷ΋ͷͰͨ͠ɻզʑ͸ɺઢܗࣝ
    ผՄೳੑͷͨΊͷे෼ͳ৚݅Λಋग़͠ɺγϛϡϨʔτ͞Εͨσʔλͱ࣮ੈք
    ͷσʔλͷ྆ํͰ͜ͷ݁ՌΛ࣮ূతʹࢧ࣋͢Δɻ
    http://arxiv.org/abs/2007.00810v3
    

    View Slide

60. ࣝผՄೳੑʁ
    • ࣝผՄೳੑ͸౷ܭϞσϧͷ๬·͍͠ಛੑͰ͋
    Γɺे෼ͳܭࢉࢿݯͱσʔλ͕͋Ε͹ɺਅͷ
    ϞσϧύϥϝʔλΛ೚ҙͷਫ਼౓ͰਪఆͰ͖Δ
    ͜ͱɻ
    

    View Slide

61. ࣝผՄೳੑΛٻΊΒΕΔϞσϧ
    ͷ৚݅(ಡ·ͳ͍͍ͯ͘Ͱ͢)
    

    View Slide

62. ࣝผՄೳੑΛٻΊΒΕΔͱخ͠
    ͍͜ͱ
    • ֶशͨ͠࠷దͳදݱ͕࠶ݱ͠΍͍͔͢Ͳ͏͔Λ
    ༧ଌ͢Δͷʹ໾ཱͭˠ࠷దͳ݁ՌΛ࠶ݱ͢Δͷ
    ʹ(ݪଇͱͯ͠)̍౓ֶ͚ͩश͢Ε͹Α͘ͳΔɻ
    • ূ໌Մೳͳ࠷దͳ܇࿅ϞσϧϥΠϒϥϦʹஔ͖
    ׵͑ΒΕΔΑ͏ʹͳΔɻ
    • ࣮ߦ͢ΔܭࢉϦιʔε͕ݮΔͷͰɺίετɾࢿ
    ݯͷ࡟ݮʹͳΔɻ
    

    View Slide

63. ᶌ ਆܦϓϩάϥϜʹ͓͚ΔڧྗͳҰൠԽͱޮ཰ੑ
    (ݪจ: Strong Generalization and Efficiency in Neural
    Programs)
    ຊݚڀͰ͸ɺχϡʔϥϧϓϩάϥϜ༠ಋͷ࿮૊ΈͷதͰɺڧྗʹҰൠԽ͢Δޮ཰త
    ͳΞϧΰϦζϜΛֶश͢Δ໰୊Λݚڀ͍ͯ͠·͢ɻχϡʔϥϧϞσϧͷೖग़ྗΠϯ
    λϑΣʔεΛ৻ॏʹઃܭ͠ɺ໛฿͢Δ͜ͱͰɺ೚ҙͷೖྗαΠζʹରͯ͠ਖ਼͍݁͠
    ՌΛग़͢ϞσϧΛֶश͠ɺڧྗͳҰൠԽΛ࣮ݱ͢Δɻ·ͨɺڧԽֶशΛར༻͢Δ͜
    ͱͰɺϓϩάϥϜͷޮ཰ࢦඪΛ࠷దԽ͠ɺ໛฿Ͱ࢖༻͢ΔڭࢣΛ௒͑Δ৽͍͠Ξϧ
    ΰϦζϜΛൃݟ͠·͢ɻ͜ΕʹΑΓɺզʑͷΞϓϩʔν͸ɺιʔτɺॱং෇͖Ϧε
    τͰͷݕࡧɺNP׬શͳ0/1φοϓβοΫ໰୊Ͱςετͨ͠Α͏ʹɺ༷ʑͳ໰୊ʹର
    ͯ͠ΧελϜهड़͞ΕͨղΛ্ճΔੑೳΛֶश͢Δ͜ͱ͕Ͱ͖ɺχϡʔϥϧϓϩά
    ϥϜ༠ಋͷ෼໺Ͱ஫໨͢΂͖ϚΠϧετʔϯΛઃఆ͍ͯ͠ΔɻϋΠϥΠτͱͯ͠ɺ
    զʑͷֶशϞσϧ͸ɺզʑ͕ςετͨ͠ͲͷΑ͏ͳೖྗσʔλαΠζͰ΋ɺ$O(n
    log n)$ͷෳࡶ͞ͰιʔτΛ׬ᘳʹ࣮ߦ͢Δ͜ͱ͕Ͱ͖ɺҰํͰɺֶशதʹݟΒΕͨ
    ϦεταΠζΛ͸Δ͔ʹ௒͑ΔϦεταΠζͰ΋ɺΫΠοΫιʔτΛؚΉϋϯυ
    ίʔυԽ͞ΕͨΞϧΰϦζϜΛૢ࡞ճ਺Ͱ্ճΔ͜ͱ͕Ͱ͖·ͨ͠ɻ
    http://arxiv.org/abs/2007.03629v2
    5PQSFDFOU/Pͱ
    ॏෳ
    

    View Slide

64. My favorite
    

    View Slide

65. ᶈ PyTorch3DʹΑΔ3DσΟʔϓϥʔχϯάͷߴ଎Խ
    (ݪจ: Accelerating 3D Deep Learning with PyTorch3D)
    σΟʔϓϥʔχϯά͸ɺ2Dͷը૾ೝࣝΛେ෯ʹվળ͖ͯͨ͠ɻ3Dʹ֦ு͢Δ͜ͱͰɺࣗ཯૸ߦंɺԾ૝
    ݱ࣮΍֦ுݱ࣮ɺ3DίϯςϯπͷΦʔαϦϯάɺ͞Βʹ͸2DೝࣝͷվળͳͲɺଟ͘ͷ৽͍͠ΞϓϦέʔ
    γϣϯ͕ਐల͢ΔՄೳੑ͕͋Γ·͢ɻ͔͠͠ɺؔ৺͕ߴ·͍ͬͯΔʹ΋͔͔ΘΒͣɺ3DσΟʔϓϥʔχϯ
    ά͸·ͩൺֱతະ։୓ͳঢ়ଶʹ͋Γ·͢ɻ͜ͷΑ͏ͳ֨ࠩ͸ɺҟछσʔλͷޮ཰తͳॲཧ΍ɺάϥϑΟο
    Ϋεૢ࡞ΛࠩผԽ͢ΔͨΊͷϦϑϨʔϛϯάͳͲɺ3DσΟʔϓϥʔχϯάʹؔΘΔ޻ֶతͳ՝୊ʹىҼ
    ͍ͯ͠Δͱߟ͑ΒΕ͍ͯ·͢ɻզʑ͸ɺ3DσΟʔϓϥʔχϯάͷͨΊͷϞδϡʔϧԽ͞Εͨޮ཰తͰඍ෼
    ՄೳͳԋࢉࢠͷϥΠϒϥϦͰ͋ΔPyTorch3DΛಋೖ͢Δ͜ͱͰɺ͜ΕΒͷ՝୊ʹରॲ͍ͯ͠·͢ɻ͜ͷϥ
    ΠϒϥϦʹ͸ɺϝογϡͱ఺܈ͷͨΊͷߴ଎ͰϞδϡʔϧࣜͷඍ෼ՄೳͳϨϯμϥʔؚ͕·Ε͓ͯΓɺ߹
    ੒ʹΑΔղੳΞϓϩʔνΛՄೳʹ͍ͯ͠·͢ɻଞͷඍ෼ՄೳͳϨϯμϥʔͱൺֱͯ͠ɺPyTorch3D͸ΑΓ
    ϞδϡʔϧԽ͞Ε͓ͯΓɺޮ཰తͰ͋ΔͨΊɺϢʔβʔ͸ΑΓ؆୯ʹ֦ு͢Δ͜ͱ͕Ͱ͖ɺେ͖ͳϝο
    γϡ΍ը૾΁ͷεέʔϦϯά΋༰қͰ͢ɻPyTorch3D ͷԋࢉࢠͱϨϯμϥʔΛଞͷ࣮૷ͱൺֱ͠ɺ଎౓ͱ
    ϝϞϦͷେ෯ͳվળΛ࣮ূ͍ͯ͠·͢ɻ·ͨɺPyTorch3DΛ࢖༻ͯ͠ɺShapeNet্ͷ2Dը૾͔Βͷڭ
    ࢣͳ͠3Dϝογϡͱ఺܈༧ଌͷͨΊͷ࠷৽ٕज़Λվળ͠·ͨ͠ɻPyTorch3D͸ΦʔϓϯιʔεͰ͋Γɺ
    3DσΟʔϓϥʔχϯάͷݚڀΛՃ଎ͤ͞ΔҰॿͱͳΔ͜ͱΛظ଴͍ͯ͠·͢ɻ
    http://arxiv.org/abs/2007.08501v1
    

    View Slide

66. 2Dը૾͔Βͷ3DγϧΤοτ༧૝
    

    View Slide

67. View Slide

68. https://pytorch3d.org/
    

    View Slide

69. PyTorch3Dͷओཁͳػೳ
    • ࡾ֯ܗϝογϡͷอଘͱૢ࡞ͷͨΊͷσʔλ
    ߏ଄
    • ࡾ֯ܗϝογϡͷޮ཰తͳૢ࡞ʢࣹӨม׵ɺ
    άϥϑ৞ΈࠐΈɺαϯϓϦϯάɺଛࣦؔ਺)
    • ඍ෼ՄೳͳϝογϡϨϯμϥʔ
    

    View Slide

70. PyTorch3D͸PyTorchͱ౷߹Ͱ͖
    ΔΑ͏ʹͳ͍ͬͯΔɻ
    • ͢΂ͯͷPyTorch3DΦϖϨʔλ͸ҎԼΛຬͨ͢ɻ
    • PyTorchςϯιϧΛ࢖࣮ͬͯ૷͞Ε͍ͯΔɻ
    • ҟछσʔλͷϛχόονΛѻ͏͜ͱ͕Ͱ͖Δɻ
    • ඍ෼Մೳɻ
    • ΞΫηϥϨʔγϣϯʹGPUΛར༻Ͱ͖Δɻ
    

    View Slide

71. ҟछσʔλͷόονྫ
    

    View Slide

72. ެ։3DֶशϞσϧ Mesh R-CNN
    https://github.com/facebookresearch/meshrcnn
    

    View Slide

73. Ұؾʹ̏DσΟʔϓϥʔχϯ
    άͷݚڀ͕Ճ଎ͦ͠͏
    

    View Slide

74. Special Thanks
    

    View Slide

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

    View Slide