Share
Embed
Transcript
86 slides
-
ࣗݾհ ழມ ॆԝ (͍ͷ·ͨ ΈͭͻΖ) גࣜձࣾ tech vein / DeepRad
גࣜձࣾ ֤දऔక ݉ σϕϩούʔ twitter: @ino2222 -
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 -
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 -
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 -
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 -
8. MobileViTɿܰྔɾ൚༻ɾϞόΠϧରԠͷϏδϣϯτϥϯεϑΥʔϚʔ (ݪจ: MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer)
ܰྔͷΈࠐΈχϡʔϥϧωοτϫʔΫʢCNNʣɺϞόΠϧɾϏδϣϯɾλεΫͷσϑΝΫτͱͳ͍ͬͯΔɻͦͷۭ ؒ༠ಋόΠΞεʹΑΓɺҟͳΔࢹ֮λεΫؒͰΑΓগͳ͍ύϥϝʔλͰදݱΛֶश͢Δ͜ͱ͕Ͱ͖Δɻ͔͠͠ɺ͜ΕΒͷ ωοτϫʔΫۭؒతʹϩʔΧϧͳͷͰ͋ΔɻάϩʔόϧͳදݱΛֶश͢ΔͨΊʹɺࣗݾҙϕʔεͷϏδϣϯτϥϯ εϑΥʔϚʔʢViTʣ͕࠾༻͞Ε͍ͯΔɻCNNͱҟͳΓɺViTsϔϏʔΣΠτͰ͋ΔɻຊߘͰɺCNNͱViTͷॴ ΛΈ߹ΘͤͯɺϞόΠϧɾϏδϣϯɾλεΫͷͨΊͷܰྔͰԆͷωοτϫʔΫΛߏங͢Δ͜ͱՄೳ͔ɺͱ͍͏ٙ Λ͔͚͍͛ͯΔɻ͜ͷతͷͨΊʹɺϞόΠϧσόΠε༻ͷܰྔͰ൚༻తͳϏδϣϯมثͰ͋ΔMobileViTΛհ ͠·͢ɻMobileViTɺมثΛ༻͍ͨใͷάϩʔόϧॲཧͷͨΊͷҟͳΔࢹɺ͢ͳΘͪɺΈࠐΈͱͯ͠ͷมث Λఏࣔ͠·͢ɻͦͷ݁ՌɺMobileViTɺ͞·͟·ͳλεΫσʔληοτʹ͓͍ͯɺCNNViTϕʔεͷωοτϫʔΫ Λେ෯ʹ্ճΔ͜ͱ͕Θ͔Γ·ͨ͠ɻImageNet-1kσʔληοτͰɼMobileViT600ສݸͷύϥϝʔλͰ78.4%ͷ top-1ਫ਼Λୡͨ͠ɽ͜ΕɼಉఔͷύϥϝʔλͷMobileNetv3ʢCNNϕʔεʣ͓ΑͼDeITʢViTϕʔεʣΑΓ 3.2%͓Αͼ6.2%ߴ͍ਫ਼Ͱ͋ΔɽMS-COCOͷମݕग़λεΫͰɺMobileViTಉͷύϥϝʔλͰMobileNetv3Α Γ5.7%ਫ਼͕ߴ͍ɻ http://arxiv.org/abs/2110.02178v1 w తɾՌɿ7J5ΛܰྔԽͨ͠৽Ϟσϧ.PCJMF7J5ͷ։ൃ w ํ๏ɿมثΛΈࠐΈͱ͍ͯ͠ɺάϩʔόϧͳදݱΛֶशͨ͠ w ݻ༗໊ɿ.PCJMF7J5 IUUQTHJUIVCDPNBQQMFNMDWOFUT w ஶऀॴଐɿ"QQMF -
ϞόΠϧͰಈ͔͍͢͠ߴਫ਼ͳϞσϧ ʢͨͩ͠ɺMobileNetv2ΑΓ͍ʣ • Mobile Friendly: ܰྔɺ൚༻తɺ(ൺֱత)ϨΠςϯγʔͰ͍͢ ͦ͏ɻ • ͨͩ͠ϞόΠϧػث(=iPhone12)ͰMobileNetv2ΑΓਪ͕8ഒ
͔ͬͨɻ →ϞόΠϧGPUʹCUDAΧʔωϧ͕ແ͍ͷͱɺCNNΈࠐΈͦ ͏ͱͷҰׅਖ਼نԽ༥߹ͳͲͷ࠷దԽ͕͋Δͷ͕ཧ༝ͱͷ͜ͱ -
1. ResNetͷٯऻɻtimmͰͷֶशखॱͷվળ (ݪจ: ResNet strikes back: An improved training procedure
in timm) HeΒʹΑͬͯઃܭ͞ΕͨӨڹྗͷ͋ΔResidual Networksɼଟ͘ͷՊֶจͰۚࣈౝతͳΞʔΩςΫ νϟͱͯ͠औΓ্͛ΒΕ͍ͯ·͢ɽ͜ΕΒͷΞʔΩςΫνϟ௨ৗɺݚڀʹ͓͚ΔσϑΥϧτͷΞʔΩ ςΫνϟͱͯ͠ɺ͋Δ͍৽͍͠ΞʔΩςΫνϟ͕ఏҊ͞ΕͨࡍͷϕʔεϥΠϯͱͯ͠ػೳ͍ͯ͠· ͢ɻ͔͠͠ɺ2015ʹResNetΞʔΩςΫνϟ͕ൃද͞ΕͯҎདྷɺχϡʔϥϧωοτϫʔΫͷτϨʔχ ϯάͷϕετϓϥΫςΟεʹ͍ͭͯେ͖ͳਐల͕͋Γ·ͨ͠ɻ৽ͨͳ࠷దԽˍσʔλΦʔάϝϯςʔ γϣϯʹΑΓɺτϨʔχϯάϨγϐͷ༗ޮੑ͕ߴ·͍ͬͯ·͢ɻຊจͰɺ͜ͷΑ͏ͳਐาΛ౷߹͠ ͨखॱͰτϨʔχϯάͨ͠߹ͷόχϥResNet-50ͷੑೳΛ࠶ධՁ͠·͢ɻզʑɺڝ૪ྗͷ͋Δֶश ઃఆͱࣄલʹֶश͞ΕͨϞσϧΛtimmΦʔϓϯιʔεϥΠϒϥϦͰڞ༗͠ɺকདྷͷݚڀͷͨΊͷΑΓྑ ͍ϕʔεϥΠϯͱཱͯͭ͜͠ͱΛظ͍ͯ͠·͢ɻྫ͑ɺզʑͷΑΓݫֶ͍͠शઃఆͰɺόχϥ ͷResNet-50ɺՃσʔλৠཹͳ͠ͰImageNet-valͷղ૾224x224Ͱ80.4%ͷτοϓ1ਫ਼Λ ୡ͍ͯ͠·͢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 -
2. େنͳࣄલτϨʔχϯάͷݶքΛ୳Δ (ݪจ: Exploring the Limits of Large Scale Pre-training)
ۙͷେنػցֶशͷൃలɺσʔλɺϞσϧαΠζɺֶश࣌ؒΛదʹεέʔϧΞοϓ͢Δ͜ͱͰɺ ࣄલֶशͷվળ͕΄ͱΜͲͷԼྲྀλεΫʹ༗རʹҠߦ͢Δ͜ͱΛ؍͢Δ͜ͱ͕Ͱ͖Δ͜ͱΛࣔࠦͯ͠ ͍ΔɻຊݚڀͰɺ͜ͷݱΛܥ౷తʹݚڀ͠ɺ্ྲྀͷਫ਼Λ্͛ΔͱԼྲྀͷλεΫͷੑೳ͕͢Δ ͜ͱΛূ໌͠·ͨ͠ɻ۩ମతʹɺVision TransformersɺMLP-MixerɺResNetsʹ͍ͭͯɺύϥϝʔλ ͕1000ສ͔Β100ԯͷൣғͰ4800ճҎ্ͷ࣮ݧΛߦ͍ɺ࠷େنͷը૾σʔλʢJFTɺ ImageNet21KʣͰֶश͠ɺ20Ҏ্ͷԼྲྀͷը૾ೝࣝλεΫͰධՁͨ͠ɻͦͷ݁ՌɺݱΛө ͠ɺ্ྲྀͱԼྲྀͷੑೳͷඇઢܗؔΛଊ͑ͨԼྲྀੑೳͷϞσϧΛఏҊͨ͠ɻ͞Βʹɺ͜ͷΑ͏ͳݱ͕ ൃੜ͢Δཧ༝Λ۷ΓԼ͛ͯཧղ͢ΔͨΊʹɺࢲ͕ͨͪ؍ͨ͠ݱɺϞσϧͷΛ௨ͯ͠දݱ͕ ਐԽ͢Δํ๏ͱີʹ͍ؔͯ͠Δ͜ͱΛࣔ͠·ͨ͠ɻ·ͨɺ͞Βʹۃͳྫͱͯ͠ɺΞοϓετϦʔ ϜͱμϯετϦʔϜͷύϑΥʔϚϯε͕૬͢Δ߹Λհ͠·͢ɻͭ·ΓɺμϯετϦʔϜͷੑ ೳΛ্ͤ͞ΔͨΊʹɺΞοϓετϦʔϜͷਫ਼Λམͱ͢ඞཁ͕͋Δͱ͍͏͜ͱͰ͢ɻ http://arxiv.org/abs/2110.02095v1 w తɿը૾ೝࣝϞσϧͷGFXTIPUֶशʹ͍ͭͯͷܥ౷తͳݚڀ w ՌɿGFXTIPUֶश༻ͷɺ൚༻తͳࣄલֶशϞσϧͷઃܭํ๏ͷఏࣔ w ํ๏ɿ7J5 .-1.JYFS 3FT/FUʹ͍ͭͯɺ6Q4USFBN%PXO4USFBNֶश࣌ͷݱͷ૬ؔΛ λεΫผɾύϥϝʔλผɾϞσϧαΠζผʹௐΔ w ݻ༗໊ɿͳ͠ w ஶऀॴଐɿ(PPHMF3FTFBSDI -
ཁ • ൚༻తͳfew shot ֶशλεΫ͚ͷࣄલֶशϞσϧΛ࡞Δ ͱ͖ͷઃܭࢦඪɾվળࡦΛఏ͍ࣔͯ͠Δɻ • ࣄલֶशͷਫ਼Λߴ͗͘͢͠ͳ͍ɾֶशΛམͱ͢ • ୯७ʹεέʔϧ͢Δ͜ͱ͕ͯ͢Λղܾ͢ΔͱݶΒͳ
͍ • ̍ͭͷԼྲྀλεΫͰಛԽ͢ΔͷͰͳ͘ɺ෯͍Լྲྀͷ λεΫͰύϑΥʔϚϯεΛ্ͤ͞ΔΑ͏ͳઃܭ্ͷબ Λ͖͢ -
3. σΟʔϓɾχϡʔϥϧɾωοτϫʔΫͱදܗࣜσʔλɻαʔϕΠ (ݪจ: Deep Neural Networks and Tabular Data: A
Survey) ෆۉ࣭ͳදܗࣜͷσʔλɺ࠷Ұൠతʹ༻͞Ε͍ͯΔσʔλܗࣜͰ͋Γɺଟ͘ͷॏཁ͔ͭܭࢉෛ ՙͷߴ͍ΞϓϦέʔγϣϯʹ͔ܽͤ·ͤΜɻಉछͷσʔληοτͰɺσΟʔϓχϡʔϥϧωοτ ϫʔΫ͕܁Γฦ͠༏ΕͨੑೳΛ͓ࣔͯ͠ΓɺͦͷͨΊ͘࠾༻͞Ε͍ͯ·͢ɻ͔͠͠ɺදܗࣜσʔλ ͷϞσϦϯάʢਪ·ͨੜʣͷద༻ɺґવͱͯ͠ඇৗʹࠔͰ͢ɻຊݚڀͰɺදܗࣜσʔ λʹର͢Δ࠷ઌͷਂֶशख๏ͷ֓ཁΛઆ໌͢Δɻ·ͣɺͦΕΒΛʮσʔλมʯʮಛघͳΞʔΩ ςΫνϟʯʮਖ਼ଇԽϞσϧʯͷ3ͭͷάϧʔϓʹྨ͠·͢ɻͦͷޙɺ֤άϧʔϓͷओཁͳΞϓϩʔν ͷแׅతͳ֓ཁΛఏڙ͠·͢ɻλϏϡϥʔσʔλΛੜ͢ΔͨΊͷਂֶशΞϓϩʔνͷٞɺλ Ϗϡϥʔσʔλ্ͰਂϞσϧΛઆ໌͢ΔͨΊͷઓུʹΑͬͯิ͞ΕΔɻࢲͨͪͷओͳߩݙɺ͜ ͷͷओͳݚڀͷྲྀΕͱطଘͷํ๏ΛऔΓ্͛ɺؔ࿈͢Δ՝ະղܾͷݚڀ՝Λ໌Β͔ʹ͢ Δ͜ͱͰ͢ɻզʑͷΔݶΓͰɺ͜ΕλϏϡϥʔσʔλʹର͢ΔਂֶशͷΞϓϩʔνΛৄࡉʹ ݕ౼ͨ͠ॳΊͯͷͷͰ͢ɻຊจɺදܗࣜσʔλΛ༻͍ͨਂֶशʹڵຯΛ࣋ͭݚڀऀ࣮Ո ʹͱͬͯɺوॏͳग़ൃͱͳΓɺࢦͱͳΔͰ͠ΐ͏ɻ http://arxiv.org/abs/2110.01889v1 w తɿදܗࣜσʔλʹର͢Δ࠷ઌͷਂֶशख๏ͷ֓ཁΛઆ໌͢ΔϨϏϡʔจ w ՌɿओͳݚڀͷྲྀΕͱطଘͷํ๏ɾϞσϧΛ·ͱΊͯɺະղܾͷݚڀ՝Λ໌Β͔ʹͨ͠ w ํ๏ɿදܗࣜͷσʔληοτΛάϧʔϓྨͯ͠ɺάϧʔϓ͝ͱͷΞϓϩʔνͷ֓ཁΛ·ͱΊͨ w ݻ༗໊ɿͳ͠ w ஶऀॴଐɿςϡʔϏϯήϯେֶ υΠπ 4$)6'")PMEJOH"( υΠπ -
4. ߴ࣍ݩͰͷֶͼɺৗʹ֎ԆతͳͷͰ͋Δɻ (ݪจ: Learning in High Dimension Always Amounts to
Extrapolation) ิؒͱ֎ૠͷ֓೦ɼਂֶश͔Βؔۙࣅ·Ͱ༷ʑͳͰجຊͱͳΔɽิؒɼ͋ Δαϯϓϧx͕ɼ༩͑ΒΕͨσʔληοτͷತแͷଆ·ͨڥք্ʹ͋Δͱ͖ʹߦΘ ΕΔɽ֎ૠɼx͕ͦͷತแͷ֎ଆʹ͋Δͱ͖ʹߦΘΕΔɽ1ͭͷجຊతͳʢޡͬͨʣೝ ࣝɺ࠷ઌͷΞϧΰϦζϜ͕͏·͘ػೳ͢ΔͷɺֶशσʔλΛਖ਼͘͠ิؒ͢Δೳྗ ͕͋Δ͔Βͩͱ͍͏ͷͰ͋Δɻ2ͭͷޡղɺิؒλεΫσʔληοτશମͰ ߦΘΕΔͱ͍͏ͷͰɺ࣮ࡍɺଟ͘ͷ؍ཧ͕͜ͷԾఆʹґଘ͍ͯ͠Δɻզʑܦ ݧతɺཧతʹ͜ΕΒͷ2ͭͷʹ͠ɺͲΜͳߴ࣍ݩʢ>100ʣͷσʔληοτͰ ɺ΄ͱΜͲ࣮֬ʹิؒى͜Βͳ͍͜ͱΛ࣮ূͨ͠ɻ͜ΕΒͷ݁ՌɺҰൠԽੑೳͷ ࢦඪͱͯ͠ͷݱࡏͷิؒ/֎ૠͷఆٛͷଥੑʹٙΛ͔͚͛ΔͷͰ͋Δɻ http://arxiv.org/abs/2110.09485v1 w తɿߴ࣍ݩۭؒ Ҏ্ ͰσʔληοτͰิ͕ؒى͜Βͳ͍ࣄΛཧతɾܦݧతʹ࣮ূ͢Δɻ w Ռɿ৽͍͠αϯϓϧʹର͢ΔิؒΛҡ࣋͢ΔͨΊʹɺσʔληοτͷαΠζ͕σʔλͷ࣍ݩʹର͠ ͯࢦؔతʹେ͖͘ͳΔ͜ͱΛ࣮ূ͠ɺطଘͷࢦඪΛ൱ఆͨ͠ɻ w ํ๏ɿ࣍ݩ͕૿͑ͨ࣌ʹิؒ͢ΔͨΊͷσʔληοτྔΛཧɾ࣮σʔληοτͷ྆໘Ͱܭࢉ w ݻ༗໊ɿ w ஶऀॴଐɿ'BDFCPPL"*3FTFBSDI -
5. ADOP: ۙࣅࠩҟԽ1ϐΫηϧɾϙΠϯτɾϨϯμϦϯά ݪจ: ADOP: Approximate Differentiable One-Pixel Point Rendering)
ຊݚڀͰɺγʔϯͷਫ਼ີԽͱ৽͍͠Ϗϡʔͷ߹ͷͨΊͷɺ৽͍͠ϙΠϯτϕʔεͷඍՄೳͳχϡʔϥϧϨϯμ ϦϯάύΠϓϥΠϯΛհ͠·͢ɻೖྗɺ܈ͷॳظਪఆͱΧϝϥͷύϥϝʔλͰ͢ɻग़ྗɺҙͷΧϝϥ ϙʔζ͔Β߹͞Εͨը૾Ͱ͢ɻ܈ͷϨϯμϦϯάɺඍՄೳͳϨϯμϥʔʹΑͬͯɺଟղ૾ͷ1ϐΫηϧ ϥελϥΠζΛ༻͍ͯߦΘΕ·͢ɻࢄతͳϥελϥΠζͷۭؒޯɺΰʔετδΦϝτϦͱ͍͏৽͍֓͠೦ʹ Αͬͯۙࣅ͞Ε·͢ɻϨϯμϦϯάޙɺχϡʔϥϧΠϝʔδϐϥϛουσΟʔϓχϡʔϥϧωοτϫʔΫʹ͞ ΕɺγΣʔσΟϯάܭࢉͱϗʔϧϑΟϦϯά͕ߦΘΕ·͢ɻͦͯ͠ɺඍՄೳͳཧϕʔεͷτʔϯϚούʔ͕ɺத ؒग़ྗΛλʔήοτը૾ʹม͠·͢ɻύΠϓϥΠϯͷͯ͢ͷεςʔδ͕ඍՄೳͰ͋ΔͨΊɺγʔϯͷͯ͢ͷ ύϥϝʔλʢΧϝϥϞσϧɺΧϝϥϙʔζɺϙΠϯτϙδγϣϯɺϙΠϯτΧϥʔɺڥϚοϓɺϨϯμϦϯάωο τϫʔΫͷॏΈɺϰΟωοτɺΧϝϥԠؔɺը૾͝ͱͷ࿐ग़ɺը૾͝ͱͷϗϫΠτόϥϯεʣΛ࠷దԽ͠·͢ɻ ຊγεςϜͰɺॳظ࠶ߏֶ͕शதʹվྑ͞ΕΔͨΊɺطଘͷΞϓϩʔνΑΓγϟʔϓͰҰ؏ੑͷ͋Δ৽͍͠ ϏϡʔΛ߹Ͱ͖Δ͜ͱΛ͍ࣔͯ͠·͢ɻ·ͨɺ1ϐΫηϧͷϙΠϯτΛޮతʹϥελϥΠζ͢Δ͜ͱͰɺҙͷ ΧϝϥϞσϧΛ༻͠ɺ100MϙΠϯτҎ্ͷγʔϯΛϦΞϧλΠϜͰදࣔ͢Δ͜ͱ͕Ͱ͖·͢ɻ http://arxiv.org/abs/2110.06635v2 w తɿΧϝϥը૾ɾ܈Λೖྗͱͨ͠ඍՄೳͳө૾ϨϯμϦϯάύΠϓϥΠϯͷհ w Ռɿҙͷඃࣸମͷ -J%"3ͷ Χϝϥը૾͔Βө૾ΛϦΞϧλΠϜͰඳըͰ͖ΔϞσϧɾπʔϧΛެ։ w ํ๏ɿ̏࣍ݩҐஔใ͖ͭͷը૾ɾ܈Λ̏࣍ݩతʹॲཧɺิؒɺ&YJGใΛͬͯ)%3ʹ࠶ߏ w ݻ༗໊ɿ"%01 IUUQTHJUIVCDPNEBSHMFJO"%01 w ஶऀॴଐɿΤΞϥϯήϯʹχϡϧϯϕϧΫେֶ υΠπ -
6. ਂχϡʔϥϧωοτϫʔΫʹΑΔྨͰɺΑ͘ྨ͞Εͨྫ͕աখධՁ͞ΕΔ (ݪจ: Well-classi fi ed Examples are Underestimated in
Classi fi cation with Deep Neural Networks) ैདྷͷਂྨϞσϧͷֶशͰɺྨͷѱ͍ྫʹ͠ɺܾఆڥք͔ΒΕͨྨͷྑ͍ྫΛແࢹ͢Δ͜ͱ ͕ৗࣝͰͨ͠ɻྫ͑ɺΫϩεΤϯτϩϐʔଛࣦΛ༻ֶ͍ͯश͢Δ߹ɺΑΓߴ͍Λ࣋ͭྫʢ͢ͳΘͪɺ Α͘ྨ͞ΕͨྫʣɺόοΫϓϩύήʔγϣϯʹ͓͍ͯখ͞ͳޯʹد༩͠·͢ɻ͔͠͠ɺ͜ͷҰൠతͳख๏ ɺදݱֶशɺΤωϧΪʔͷ࠷దԽɺϚʔδϯͷ૿ՃΛ͛Δ͜ͱΛཧతʹ͍ࣔͯ͠·͢ɻ͜ͷܽؕΛଧͪফ ͨ͢Ίʹɺզʑɺྨͷྑ͍ྫʹՃࢉϘʔφεΛ༩͑ͯɺֶशͷߩݙΛ෮׆ͤ͞Δ͜ͱΛఏҊ͢Δɻ͜ͷ ྫɺ͜ΕΒ3ͭͷΛཧతʹղܾ͢ΔͷͰ͋ΔɻຊจͰɺը૾ྨɺάϥϑྨɺػց༁ͳͲ ͷଟ༷ͳλεΫʹ͓͍ͯɺཧతͳ݁ՌΛݕূͨ͠ΓɺຊྫΛ༻͍ͯେ෯ͳੑೳ্Λ࣮ݱ͢Δ͜ͱ Ͱɺ͜ͷओுΛ࣮ূతʹࢧ࣋͢Δɻ͞ΒʹɺຊจɺզʑͷΞΠσΞ͕͜ΕΒ3ͭͷΛղܾͰ͖ΔͨΊɺ ෆۉߧͳྨɺOODݕग़ɺఢରత߈ܸԼͷΞϓϦέʔγϣϯͳͲͷෳࡶͳγφϦΦʹରԠͰ͖Δ͜ͱΛࣔͯ͠ ͍Δɻίʔυɺhttps://github.com/lancopku/well-classi fi ed-examples-are-underestimated ɻ http://arxiv.org/abs/2110.06537v2 w తɾՌɿΫϩεΤϯτϩϐʔϩεΛվྑͨ͠৽͍͠-PTTؔͷఏҊ w ํ๏ɿόοΫϓϩύήʔγϣϯʹΑΔ$SPTT&OUSPQZ $& ϩεͷΛ໌Β͔ʹ͢Δ w ݻ༗໊ɿ&ODPVSBHJOH-PTT &- w ஶऀॴଐɿژେֶ -
7. ByteTrack:ͯ͢ͷݕग़ϘοΫεΛؔ࿈͚Δ͜ͱʹΑΔෳମͷ (ݪจ: ByteTrack: Multi-Object Tracking by Associating Every Detection
Box) ϚϧνɾΦϒδΣΫτɾτϥοΩϯάʢMOTʣɼಈըͷΦϒδΣΫτͷόϯσΟϯάɾϘοΫεͱΞΠσϯςΟ ςΟΛਪఆ͢Δ͜ͱΛతͱ͍ͯ͠·͢ɽଟ͘ͷख๏ͰɺᮢΑΓߴ͍είΞΛ࣋ͭݕग़ϘοΫεΛؔ࿈͚Δ͜ ͱͰΞΠσϯςΟςΟΛಘ͍ͯ·͢ɻ͔͠͠ɺݕग़είΞͷ͍ΦϒδΣΫτʢྫ͑ɺӅ͞ΕͨΦϒδΣΫτʣ୯ ७ʹࣺͯΒΕͯ͠·͏ͨΊɺແࢹͰ͖ͳ͍ਅͷΦϒδΣΫτͷܽམɺஅยతͳي͕ੜͯ͡͠·͏ɻ͜ͷΛղܾ ͢ΔͨΊʹɺզʑBYTEͱݺΕΔγϯϓϧͰޮՌత͔ͭ൚༻తͳؔ࿈͚ํ๏Λఏࣔ͢ΔɻBYTEͱɺߴείΞͷ ݕग़ϘοΫε͚ͩͰͳ͘ɺͯ͢ͷݕग़ϘοΫεΛؔ࿈͚ͯ͢Δํ๏Ͱ͋ΔɻείΞͷݕग़ϘοΫεʹରͯ͠ ɼτϥοΫϨοτͱͷྨࣅੑΛར༻ͯ͠ਅͷΦϒδΣΫτΛ෮ݩ͠ɼഎܠݕग़ΛϑΟϧλϦϯά͢ΔɽBYTEΛ9ͭͷ ҟͳΔ࠷ઌͷτϥοΧʔʹద༻ͨ͠ͱ͜ΖɺIDF1είΞΛ1ʙ10ϙΠϯτͷൣғͰҰ؏ͯ͠վળ͢Δ͜ͱ͕Ͱ͖·͠ ͨɻMOTͷ࠷ઌͷੑೳΛ׆͔ͨ͢ΊʹɺByteTrackͱ໊͚ΒΕͨγϯϓϧͰڧྗͳτϥοΧʔΛઃܭ͠·ͨ͠ɻͦ ͷ݁ՌɺV100 GPUΛ༻͍ͨςετηοτʮMOT17ʯʹ͓͍ͯɺMOTA80.3ɺIDF1 77.3ɺHOTA63.1Λୡ͠ɺ30 FPSͷಈ࡞Λ࣮ݱ͠·ͨ͠ɻιʔείʔυɺࣄલֶशࡁΈϞσϧɺσϓϩΠόʔδϣϯɺଞͷτϥοΧʔͷద༻ʹ ؔ͢ΔνϡʔτϦΞϧɺhttps://github.com/ifzhang/ByteTrack Ͱެ։͍ͯ͠·͢ɻ http://arxiv.org/abs/2110.06864v2 w తɾՌɿϚϧνΦϒδΣΫττϥοΩϯάΞϧΰϦζϜ#:5&Ͱ4P5"Λୡͨ͠ w ํ๏ɿΧϧϚϯϑΟϧλʹΑΔΦϒδΣΫτҐஔਪఆ w ݻ༗໊ɿ#ZUF5SBDL IUUQTHJUIVCDPNJG[IBOH#ZUF5SBDL w ஶऀॴଐɿ՚தՊٕେֶɺ߳ߓେֶɺ#ZUF%BODF 5JL5PLͷձࣾ -
8. MobileViTɿܰྔɾ൚༻ɾϞόΠϧରԠͷϏδϣϯτϥϯεϑΥʔϚʔ (ݪจ: MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer)
Pickup http://arxiv.org/abs/2110.02178v1 -
9. εέʔϧͰͷߴϞσϧฤू (ݪจ: Fast Model Editing at Scale) ࣄલʹֶश͞ΕͨେنͳϞσϧɺ༷ʑͳμϯετϦʔϜͷλεΫͰૉΒ͍݁͠ՌΛ࣮ݱ͍ͯ͠·͕͢ɺطଘͷେن ͳϞσϧʹ·ͩΤϥʔ͕͋Γɺਖ਼֬ͳ༧ଌͰ͋ͬͯ࣌ؒͷܦաͱͱʹݹ͘ͳͬͯ͠·͏͜ͱ͕͋Γ·͢ɻ͜ͷΑ͏ͳࣦ
ഊΛֶͯ͢श࣌ʹݕग़͢Δ͜ͱෆՄೳͰ͋ΔͨΊɺ͜ͷΑ͏ͳϞσϧͷ։ൃऀͱΤϯυϢʔβͷ྆ํ͕ɺϞσϧΛͦͷ· ·ʹͯ͠ෆਖ਼֬ͳग़ྗΛमਖ਼Ͱ͖ΔΑ͏ʹ͢Δ͜ͱ͕·Ε·͢ɻ͔͠͠ɺେنͳχϡʔϥϧωοτϫʔΫֶ͕श͢Δදݱ ࢄ͓ͯ͠ΓɺϒϥοΫϘοΫεԽ͍ͯ͠ΔͨΊɺ͜ͷΑ͏ͳରΛߜͬͨฤूࠔͰ͢ɻͷ͋Δೖྗͱ৽ͨͳر ͷग़ྗ͕1͚ͭͩఏࣔ͞Εͨ߹ɺඍௐΞϓϩʔνΦʔόʔϑΟοτ͢Δ͕͋Δɻ·ͨɺଞͷฤूΞϧΰϦζϜɺඇ ৗʹେ͖ͳϞσϧʹద༻͢Δ߹ɺܭࢉ͕ෆՄೳͰ͋Δ͔ɺ୯ʹޮՌ͕ͳ͍ɻେنͳϞσϧͷϙετϗοΫฤूΛ༰қʹ͢ ΔͨΊʹɺզʑMEND (Model Editor Networks with Gradient Decomposition)ΛఏҊ͠·͢ɻMENDɺඪ४తͳඍௐ ʹΑͬͯಘΒΕͨޯΛɺޯͷϥϯΫղΛ༻͍ͯม͢Δ͜ͱΛֶश͠ɺ͜ͷมͷύϥϝʔλԽΛѻ͍ͯ͘͢͠ ͍·͢ɻMENDɺ100ԯҎ্ͷύϥϝʔλΛ࣋ͭϞσϧͰ͋ͬͯɺ1ͭͷGPU্Ͱ1Ҏʹֶश͢Δ͜ͱ͕Ͱ͖·͢ɻ· ͨɺҰֶशͨ͠MENDɺࣄલʹֶशͨ͠Ϟσϧʹ৽ͨͳฤूΛՃ͑Δ͜ͱ͕Ͱ͖·͢ɻT5ɺGPTɺBERTɺBARTϞσϧΛ ༻͍࣮ͨݧʹΑΓɺMENDɺઍສ͔Β100ԯҎ্ͷύϥϝʔλΛ࣋ͭϞσϧʹରͯ͠ޮՌతͳฤूΛߦ͏͜ͱ͕Ͱ͖Δ། ҰͷϞσϧฤूख๏Ͱ͋Δ͜ͱ͕Θ͔Γ·ͨ͠ɻ࣮ɺhttps://sites.google.com/view/mend-editing ɻ http://arxiv.org/abs/2110.11309v1 w తɿେنͳ5SBOTGPSNFSϞσϧͰͷՃमਖ਼ͷͨΊͷ࠶ֶश࣌ͷʹରॲ͢Δ w ՌɿߴʹɺΦʔόʔϑΟοτͳ͘ඍௐͰ͖ΔϞσϧ.&/%ͷఏҊ w ํ๏ɿϞσϧฤूࣗମΛֶशͱͯ͠ѻ͏ w ݻ༗໊ɿ.&/% IUUQTTJUFTHPPHMFDPNWJFXNFOEFEJUJOH w ஶऀॴଐɿελϯϑΥʔυେֶ -
࣌ͷྲྀΕʹΑ͕ͬͯ͑มΘΔ →࠶ֶश͕ඞཁ • ΠΪϦεͷट૬୭ʁ ✗ 76ट૬(ςϦʔβɾϝΠ) ◦ 77ट૬(ϘϦεɾδϣϯιϯ)
IUUQTTJUFTHPPHMFDPNWJFXNFOEFEJUJOH -
10. ࣗݾڭࢣֶ͖शσʔληοτͷෆۉߧʹରͯ͠ΑΓؤ݈Ͱ͋Δ (ݪจ: Self-supervised Learning is More Robust to Dataset
Imbalance) ࣗݾڭࢣֶ͖शʢSelf-Supervised Learning: SSLʣɼϥϕϧͳ͠Ͱֶश͢ΔͨΊɼҰൠతͳࢹ֮දݱΛֶश͢ΔͨΊͷ εέʔϥϒϧͳํ๏Ͱ͋Δɽ͔͠͠ɼେنͳϥϕϧͳ͠σʔληοτͰɼϥϕϧͷ͕ϩϯάςʔϧͰ͋Δ͜ͱ͕ଟ ͘ɼSSLͷಈ࡞ʹ͍ͭͯ΄ͱΜͲ͔͍ͬͯͳ͍ɽຊݚڀͰɼσʔληοτෆۉߧԼͰͷࣗݾڭࢣֶ͖शΛܥ౷తʹ ௐࠪ͢Δɽ·ͣɼେنͳ࣮ݧʹΑΓɼطͷڭࢣ͖දݱɼڭࢣ͖දݱΑΓΫϥεͷෆۉߧʹରͯ͠ΑΓؤ݈Ͱ͋ Δ͜ͱ͕Θ͔ͬͨɽSSLΛ༻͍ͨόϥϯεܕͱΞϯόϥϯεܕͷࣄલֶशͷੑೳࠩɺαϯϓϧαΠζʹؔΘΒͣɺυϝΠ ϯɺಛʹυϝΠϯ֎ͷධՁʹ͓͍ͯɺڭࢣֶ͖शͷੑೳࠩΑΓ༗ҙʹখ͘͞ͳ͍ͬͯ·͢ɻୈೋʹɼSSLͷؤ݈ੑΛ ཧղ͢ΔͨΊʹɼSSLසग़σʔλ͔ΒΑΓ๛͔ͳಛΛֶश͢Δͱ͍͏ԾઆΛཱͯͨɽͭ·ΓɼكͳΫϥεԼྲྀͷλε Ϋͷྨʹཱͭɼϥϕϧͱແ͕ؔͩୡՄೳͳಛΛֶश͢ΔͷͰͳ͍͔ͱߟ͑ΒΕΔɽରরతʹɼڭࢣֶ͖ शͰɼසग़͢Δྫ͔ΒϥϕϧͱແؔͳಛΛֶश͢Δಈػ͕ͳ͍ɽ͜ͷԾઆΛɺ୯७Խ͞ΕͨઃఆͰͷ߹࣮ݧ ͱཧతੳʹΑͬͯݕূ͢Δɻୈࡾʹɺཧతಎʹ৮ൃ͞Εͯɺ࠶ॏΈ͚ਖ਼ଇԽٕज़ΛߟҊͨ͠ɻ͜ͷٕज़ɺ͍͘ ͔ͭͷධՁج४ʹج͍ͮͯɺΞϯόϥϯεͳσʔληοτʹ͓͚ΔSSLදݱͷ࣭ΛҰ؏্ͯͤ͠͞ɺಉ͡ͷྫΛ࣋ͭ όϥϯεͷऔΕͨσʔληοτͱΞϯόϥϯεͳσʔληοτͷؒͷখ͞ͳΪϟοϓΛຒΊΔɻ http://arxiv.org/abs/2110.05025v1 w తɿΫϥεෆۉߧԼͰͷʮࣗݾڭࢣֶ͖ͭश 44- ʯͷදݱ࣭Λܥ౷తʹௐࠪ͢Δ w Ռɿ44-͕σʔληοτͷෆۉߧʹରͯ͠ؤڧͳࣄΛԠ༻ͯ͠SX4".ΛߟҊ w ݻ༗໊ɿ3FXFJHIUFE4". SX4". w ஶऀॴଐɿελϯϑΥʔυେֶɺτϤλɾϦαʔνɾΠϯεςΟςϡʔτ -
1. ADOP: Approximate Differentiable One-Pixel Point Rendering (ۙࣅࠩҟԽ1ϐΫη ϧɾϙΠϯτɾϨϯμϦϯά) ॏෳ
http://arxiv.org/abs/2110.06635v2 -
2. ࣮ɺσʔλαΠΤϯεɺΧΦεɿ͋ΒΏΔσʔληοτΛ୯ҰͷύϥϝʔλͰϑΟοτͤ͞Δํ๏ (ݪจ: Real numbers, data science and chaos: How
to fi t any dataset with a single parameter) ࣌ܥྻɺը૾ɺԻͳͲɺͲͷΑ͏ͳϞμϦςΟͷσʔλͰ͋ͬͯ ɺ୯Ұͷ࣮ͷύϥϝʔλΛ࣋ͭྑͳεΧϥʔؔʢ࿈ଓɺඍ Մೳ...ʣͰۙࣅͰ͖Δ͜ͱΛࣔ͠·͢ɻຊݚڀͰɺΧΦεཧͷ جຊతͳ֓೦ʹج͍ͮͯɺσʔλͷͯ͢ͷαϯϓϧʹҙͷਫ਼Ͱ ϑΟοτͤ͞ΔͨΊʹɺ͜ͷύϥϝʔλΛௐ͢Δํ๏Λࣔ͢ڭҭత ͳΞϓϩʔνΛ࠾༻͍ͯ͠·͢ɻح৺ԢͳσʔλαΠΤϯςΟε τΛରʹɺػցֶशϞσϧͷදݱྗͱҰൠԽʹؔ͢Δ͜Ε·Ͱͷಉ ༷ͷ؍݁ՌΛൃలͤͨ͞ͷͰ͢ɻ http://arxiv.org/abs/1904.12320v1 w తɿҙͷσʔληοτ9ͷͯ͢ͷαϯϓϧ͕ɼ୯७ͳඍํఔࣜʹΑͬͯ࠶ݱͰ͖Δ͜ͱΛࣔ͢͜ͱ w ՌɿҰͭͷ࣮ύϥϝʔλ͚ͩͰશͯͷԻɾࢹ֮σʔλΛੜͰ͖Δɺ୯७ͰඍՄೳͳఆࣜԽ w ํ๏ɿΧΦεཧΛݩʹͨ͠Իɾը૾σʔλͷҰൠԽ w ݻ༗໊ɿ4JOHMF1BSBNFUFS'JU IUUQTHJUIVCDPN3BOMPUTJOHMFQBSBNFUFS fi U w ஶऀॴଐɿ4"1-BCT υΠπͷιϑτΣΞاۀͷݚڀػؔ -
3. σϧϑΝΠػցͷྙཧɾنൣΛࢦͯ͠ (ݪจ: Delphi: Towards Machine Ethics and Norms) ػցʹྙཧతͳߦಈΛڭ͑ΔʹɺԿ͕ඞཁͰ͠ΐ͏͔ʁେ·͔ͳྙཧతϧʔϧɺʮೊɺࡴ͢ͳ͔Εʯͱ͍͏Α͏ʹ؆୯ʹड़Δ͜
ͱ͕Ͱ͖Δ͔͠Ε·ͤΜ͕ɺͦͷΑ͏ͳϧʔϧΛݱ࣮ͷঢ়گʹద༻͢Δ͜ͱΔ͔ʹෳࡶͰ͢ɻྫ͑ɺʮ༑ਓΛॿ͚Δʯ͜ͱ Ұൠతʹྑ͍͜ͱͰ͕͢ɺʮ༑ਓ͕ϑΣΠΫχϡʔεΛྲྀ͢ͷΛॿ͚Δʯ͜ͱྑ͍͜ͱͰ͋Γ·ͤΜɻࢲͨͪɺػցྙཧنൣ ʹର͢Δ4ͭͷࠜຊతͳ՝Λಛఆ͍ͯ͠·͢ɻ(1) ಓಙతڭ܇ͱࣾձతنൣͷཧղɺ(2) ݱ࣮ੈքͷঢ়گΛࢹ֮తʹɺ͋Δ͍ࣗવݴޠ ʹΑΔهड़ΛಡΈऔΔೳྗɺ(3) ҟͳΔจ຺ʹ͓͚Δସతͳߦಈͷ݁ՌΛ༧ଌ͢ΔͨΊͷৗࣝతͳਪɺ(4) ࠷ॏཁͳͷɺڝ߹͢ ΔՁͷ૬ޓ࡞༻ͱɺҟͳΔจ຺ʹ͓͚ΔͦΕΒͷࠜڌΛߟྀͯ͠ྙཧతஅΛԼ͢ೳྗʢྫɿදݱͷࣗ༝ͷݖརͱϑΣΠΫχϡʔε ͷ֦ࢄࢭʣɻ ຊߘͰɺ͜ΕΒͷٙΛਂֶशͷύϥμΠϜͷதͰղܾ͠Α͏ͱ͢ΔͷͰ͋ΔɻզʑͷϓϩτλΠϓϞσϧͰ͋ ΔDelphiɺݴޠϕʔεͷৗࣝతͳಓಙతਪʹڧ͍ظΛد͓ͤͯΓɺਓؒʹΑͬͯݕূ͞Εͨਫ਼࠷େͰ92.1%Ͱ͋Δɻ͜Ε ɺGPT-3ͷθϩγϣοτੑೳ͕52.3%Ͱ͋ΔͷͱରরతͰ͋Γɺେنͳεέʔϧ͚ͩͰɺࣄલʹ܇࿅͞ΕͨਆܦݴޠϞσϧʹਓ ؒͷՁΛ༩͢Δ͜ͱͰ͖ͳ͍͜ͱΛ͍ࣔࠦͯ͠Δɻͦ͜Ͱࢲͨͪɺػց༻ʹΧελϚΠζ͞ΕͨಓಙͷڭՊॻ ʮCommonsense Norm BankʯΛൃද͠·ͨ͠ɻ͜ͷڭՊॻʹɺৗͷ͞·͟·ͳঢ়گʹ͓͚Δਓʑͷྙཧతஅͷྫ͕170ສ݅ऩ ͞Ε͍ͯ·͢ɻຊݚڀɺࠓޙͷݚڀͷͨΊͷ৽ͨͳϦιʔεͱϕʔεϥΠϯͱͳΔੑೳʹՃ͑ͯɺਓؒͷීวతͳՁͱݸਓతͳՁ ͷ۠ผɺҟͳΔಓಙతΈͷϞσϧԽɺػցྙཧͷઆ໌ՄೳͰҰ؏ੑͷ͋ΔΞϓϩʔνͳͲɺ͍͔ͭ͘ͷॏཁͳະղܾͷݚڀ՝ ʹͭͳ͕Δ৽ͨͳಎΛఏڙ͍ͯ͠·͢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 ΞϨϯਓೳݚڀॴ -
4. ϚϧνλεΫϓϩϯϓτʹΑΔτϨʔχϯάͰθϩγϣοτλεΫͷҰൠԽΛ࣮ݱ (ݪจ: Multitask Prompted Training Enables Zero-Shot Task Generalization)
ۙɺେنͳݴޠϞσϧ͕ɺଟ༷ͳλεΫʹ͓͍ͯଥͳθϩγϣοτ൚ԽΛୡ͢Δ͜ͱ͕ࣔ͞Ε͍ͯ Δɻ͜ΕɺݴޠϞσϧͷֶशʹ͓͚Δ҉ͷϚϧνλεΫֶशͷ݁ՌͰ͋Δͱ͍͏Ծઆཱ͕ͯΒΕ͍ͯ· ͢ɻͰɼ໌ࣔతͳϚϧνλεΫֶशʹΑͬͯɼθϩγϣοτͷ൚Խ͕Ҿ͖ى͜͞ΕΔͷͰ͠ΐ͏͔ʁ͜ ͷٙΛେنʹݕূ͢ΔͨΊʹɺҰൠతͳࣗવݴޠλεΫΛਓ͕ؒಡΊΔϓϩϯϓτܗࣜʹ؆୯ʹϚοϐϯ ά͢ΔγεςϜΛ։ൃ͠·ͨ͠ɻେنͳڭࢣ͖σʔληοτΛม͠ɺͦΕͧΕͷσʔληοτʹ༷ʑ ͳࣗવݴޠΛ༻͍ͨෳͷϓϩϯϓτ͕༻ҙ͞Ε͍ͯΔɻ͜ΕΒͷϓϩϯϓτσʔληοτɺࣗવݴޠͰࢦ ఆ͞Εͨશ͘ݟͨ͜ͱͷͳ͍λεΫΛ࣮ߦ͢ΔϞσϧͷೳྗΛϕϯνϚʔΫ͢Δ͜ͱ͕Ͱ͖Δɻࣄલʹֶश͠ ͨΤϯίʔμʔͱσίʔμʔͷϞσϧΛɺ༷ʑͳλεΫΛؚΉϚϧνλεΫࠞ߹Ͱඍௐͨ͠ɻ͜ͷϞσϧ ɺ͍͔ͭ͘ͷඪ४తͳσʔληοτʹ͓͍ͯڧྗͳθϩγϣοτੑೳΛୡ͠ɺ͠͠16ഒͷαΠζͷϞ σϧΑΓ༏Ε͍ͯΔɻ͞ΒʹɺBIG-BenchϕϯνϚʔΫͷλεΫͷαϒηοτʹ͓͍ͯɺ6ഒͷϞσϧΑΓ ߴ͍ੑೳΛൃش͠·ͨ͠ɻͯ͢ͷϓϩϯϓτͱֶशࡁΈϞσϧɼgithub.com/bigscience-workshop/ promptsource/Ͱެ։͞Ε͍ͯ·͢ɽ http://arxiv.org/abs/2110.08207v1 w తɿେنͳݴޠϞσϧ͕θϩγϣοτ൚Խ͕Ͱ͖Δཧ༝ͷݚڀ w ՌɿݴޠϞσϧͷڧྗͳθϩγϣοτҰൠԽೳྗΛ࣮ݱͰ͖ΔπʔϧͱֶशࡁΈϞσϧΛެ։ w ํ๏ɿϚϧνλεΫϓϩϯϓττϨʔχϯά w ݻ༗໊ɿ5ϕʔεͷ5Ϟσϧ 1SPNQU4PVSDF HJUIVCDPNCJHTDJFODFXPSLTIPQ QSPNQUTPVSDF w ஶऀॴଐɿ)VHHJOH'BDF 🤗 ϒϥϯେֶ΄͔ -
5. ඇෛͷۭؒతҼղ (ݪจ: Nonnegative spatial factorization) ΨεաఔɺϊϯύϥϝτϦοΫͳॊೈੑͱෆ࣮֬ੑͷఆྔԽ͕ՄೳͰ͋Δ͜ͱ͔Βɺۭؒσʔλͷ ղੳʹ͘༻͍ΒΕ͓ͯΓɺ࠷ۙ։ൃ͞Εͨεέʔϥϒϧͳۙࣅ๏ʹΑΓɺେͳσʔληοτͷద ༻͕༰қʹͳ͍ͬͯΔɻଟมྔͷ݁Ռʹରͯ͠ɺۭؒ૬ؔΛར༻ͨ࣍͠ݩݮΛΈ߹ΘͤͨίΞϦ φΠζͷઢܗϞσϧ͕͋Δɻ͔͠͠ɺඇෛϞσϧͱҟͳΓɺύʔπϕʔεͷදݱΛճ෮͠ͳ͍ͨΊɺ
ͦͷ࣮જࡏҼࢠͱෛՙྔղऍ͕͍͠ɻຊݚڀͰɺۭؒΛߟྀͨ֬͠త࣍ݩݮϞσϧͰ͋Δ ඇෛͷۭؒҼࢠԽʢNSFʣΛఏҊ͢ΔɻNSFɺγϛϡϨʔγϣϯͱߴ࣍ݩۭؒτϥϯεΫϦϓτϛΫ εσʔλΛ༻͍ͯɺMEFISTOͷΑ͏ͳ࣮ۭؒҼࢠԽඇۭؒ࣍ݩݮ๏ͱൺֱͨ͠ɻNSFɺҨࢠ ൃݱͷҰൠԽՄೳͳۭؒύλʔϯΛಛఆ͠·͢ɻͯ͢ͷҨࢠൃݱύλʔϯ͕ۭؒతͰ͋ΔͱݶΒ ͳ͍ͨΊɺۭؒతͳཁૉͱඇۭؒతͳཁૉΛΈ߹ΘͤͨNSFͷϋΠϒϦου֦ுΛఏҊ͠ɺ؍ଌͱ ಛͷ྆ํʹ͍ۭͭͯؒతͳॏཁੑΛఆྔԽ͢Δ͜ͱΛՄೳʹ͍ͯ͠·͢ɻNSFͷTensorFlow࣮ɺ https://github.com/willtownes/nsf-paper ͔ΒೖखՄೳͰ͋Δɻ http://arxiv.org/abs/2110.06122v1 w తɿੜମ৫ͷݚڀʹ͓͚ΔۭؒతͳҨࢠൃݱͷଌఆ w ՌɿҨࢠൃݱͷҰൠԽՄೳͳۭؒύλʔϯΛಛఆ w ํ๏ɿΨεաఔΛ༻͍ͨσʔλղੳͰͷۭؒΛߟྀͨ֬͠త࣍ݩݮϞσϧΛߟҊ w ݻ༗໊ɿ/4' /POOFHBUJWF4QBUJBM'BDUPSJ[BUJPO /4') /4')ZCSJE w ஶऀॴଐɿϓϦϯετϯେֶ άϥουετʔϯݚڀॴ αϯϑϥϯγεί -
6. ߴ࣍ݩͰͷֶͼɺৗʹ֎ԆతͳͷͰ͋Δɻ (ݪจ: Learning in High Dimension Always Amounts to
Extrapolation) ॏෳ http://arxiv.org/abs/2110.09485v1 -
7. StyleAlignɿྻͨ͠StyleGANϞσϧͷੳͱԠ༻ (ݪจ: StyleAlign: Analysis and Applications of Aligned StyleGAN
Models) ຊจͰɺΞϥΠϝϯτ͞ΕͨੜϞσϧͷಛੑͱͦͷԠ༻ʹ͍ͭͯৄࡉʹݕ౼ͨ͠ɻ͜͜Ͱɺ2ͭͷϞσϧ ͕ಉ͡ΞʔΩςΫνϟΛڞ༗͠ɺҰํʢࢠʣ͕ଞํʢʣ͔ΒผͷυϝΠϯͷඍௐΛܦͯಘΒΕͨ߹ɺ ྻͨ͠ϞσϧͱݺͿ͜ͱʹ͢Δɻ͢Ͱʹ͍͔ͭ͘ͷ࡞ͰɺΞϥΠϝϯτ͞ΕͨStyleGANϞσϧͷجຊతͳಛ ੑΛར༻ͯ͠ɺը૾ؒͷ༁Λߦ͍ͬͯΔɻ͜͜ͰɺStyleGANʹযΛͯͯɺϞσϧͷΞϥΠϯϝϯτΛॳ Ίͯৄࡉʹௐࠪ͢Δɻ·ͣɺྻͨ͠ϞσϧΛܦݧతʹੳ͠ɺͦͷੑ࣭ʹؔ͢Δॏཁͳٙʹର͢Δ͑Λఏ ڙ͢ΔɻಛʹɺࢠϞσϧͷજࡏۭؒϞσϧͷજࡏۭؒͱҙຯతʹ߹͓ͯ͠ΓɺਓͷإڭձͳͲͷԕ͍ σʔλྖҬͰ͋ͬͯɺ৴͡ΒΕͳ͍΄Ͳ๛͔ͳҙຯΛܧঝ͍ͯ͠Δ͜ͱ͕Θ͔Γ·ͨ͠ɻ࣍ʹɺ͜ͷΑ͏ʹ͠ ͯಘΒΕͨཧղΛͱʹɺྻͨ͠ϞσϧΛ׆༻ͯ͠͞·͟·ͳ՝Λղܾ͠·͢ɻը૾༁ʹՃ͑ͯɺશʹ ࣗಈԽ͞ΕͨΫϩευϝΠϯͷը૾ϞʔϑΟϯάΛ࣮ূ͠·ͨ͠ɻ͞ΒʹɺྖҬͰͷࢹͷΈʹཔΓͳ͕Βɺ ࢠྖҬͰθϩγϣοτͷࢹ֮λεΫΛ࣮ߦͰ͖Δ͜ͱΛࣔ͠·͢ɻ͞ΒʹɺྖҬͷࢹͷΈʹґଘ͠ͳ͕ ΒɺࢠྖҬͰθϩγϣοτɾϏδϣϯɾλεΫΛ࣮ߦ͢Δ͜ͱ͕Ͱ͖Δ͜ͱΛࣔ͠·ͨ͠ɻ͜ͷΞϓϩʔνʹΑ Γɺ؆୯ͳඍௐͱసͷΈͰɺ࠷ઌͷ݁Ռ͕ಘΒΕΔ͜ͱΛఆੑత͓Αͼఆྔతʹࣔ͠·ͨ͠ɻ http://arxiv.org/abs/2110.11323v1 w తɿ4UZMF("/ͷੳͱԠ༻ w Ռɿ4UZMF("/ͷΫϩευϝΠϯͷసҠֶशͷੑೳΛվળͰ͖ͨ w ํ๏ɿ4UZMF("/Λશ͘ผυϝΠϯʹసҠֶशͨ࣌͠ͷજࡏۭؒΛੳ w ݻ༗໊ɿ4UZMF"MJHO w ஶऀॴଐɿϔϒϥΠେֶ ςϧΞϏϒେֶ "EPCF3FTFBSDI -
8. AudacityͷͨΊͷਂֶशπʔϧɻݚڀऀ͕ΞʔςΟετͷπʔϧΩοτΛ֦ு͢Δͷʹཱͭ (ݪจ: Deep Learning Tools for Audacity: Helping Researchers
Expand the Artist's Toolkit) ࢲͨͪɺΦʔϓϯιʔεͷਓؾΦʔσΟΦฤूιϑτAudacity ʹɺ࠷খݶͷ։ൃऀͷ࿑ྗͰχϡʔϥϧωοτϫʔΫΛ౷߹͢Δι ϑτΣΞϑϨʔϜϫʔΫΛհ͠·͢ɻຊจͰɺΤϯυϢʔ βʔͱχϡʔϥϧωοτϫʔΫ։ൃऀͷ྆ํʹ͚ͯɺ͍͔ͭ͘ͷ ༻ྫΛհ͠·͢ɻ͜ͷݚڀ͕ɺਂֶशͷ࣮ફऀͱΤϯυϢʔ βʔͷؒͷ৽͍͠Ϩϕϧͷ૬ޓ࡞༻Λଅਐ͢Δ͜ͱΛظ͍ͯ͠· ͢ɻ http://arxiv.org/abs/2110.13323v1 w తɾՌɿԻฤूιϑτ"VEBDJUZΛχϡʔϥϧωοτϫʔΫʹରԠ͢Δ w ํ๏ɿΦϯϥΠϯαΠτ)VHHJOH'BDFͷެ։ϞσϧʹରԠ w ݻ༗໊ɿ"VEBDJUZ%JHJUBM"VEJP8PSLTUBUJPO w ஶऀॴଐɿϊʔεΣελϯେֶ "VEBDJUZ5FBN Իฤूιϑτ։ൃνʔϜ -
9. ECQx:ϏοτͰૄͳDNNͷͨΊͷઆ໌Մೳੑʹجͮ͘ྔࢠԽ (ݪจ: ECQx: Explainability-Driven Quantization for Low-Bit and Sparse
DNNs) ༷ʑͳΞϓϦέʔγϣϯʹ͓͚ΔσΟʔϓχϡʔϥϧωοτϫʔΫʢDNNʣͷ֮·͍͠ޭɺωοτϫʔΫύϥϝʔλԋ ࢉྔͷେ෯ͳ૿ՃΛ͍ͬͯ·͢ɻ͜ͷΑ͏ͳϝϞϦܭࢉྔͷ૿ՃɺϞόΠϧػثͷΑ͏ͳϦιʔεʹ੍ͷ͋ΔϋʔυΣ ΞϓϥοτϑΥʔϜͰɺਂֶशΛߦ͏͜ͱ͕Ͱ͖·ͤΜɻ࠷ۙͰɺϞσϧͷੑೳΛՄೳͳݶΓҡ࣋ͭͭ͠ɺ͜ΕΒͷΦʔ όʔϔουΛݮ͢Δ͜ͱΛతͱͯ͠ɺύϥϝʔλݮٕज़ɺύϥϝʔλྔࢠԽɺՄٯѹॖٕज़ͳͲ͕։ൃ͞Ε͍ͯΔɻ ຊষ ͰɺDNNͷͨΊͷ৽͍͠ྔࢠԽύϥμΠϜΛ։ൃ͠ɺઆ໌͠·͢ɻຊख๏Ͱɺઆ໌ՄೳͳAI(XAI)ͷ֓೦ͱใཧͷ֓೦Λ ׆༻͠ɺྔࢠԽΫϥελͷڑʹج͍ͮͯॏΈΛׂΓͯΔΘΓʹɺϨΠϠϫΠζؔ࿈ੑ(LRP)͔ΒಘΒΕΔॏΈͷؔ ࿈ੑͱΫϥελͷใྔ(Τϯτϩϐʔ࠷దԽ)ΛՃͰߟྀ͠·͢ɻ࠷ऴతͳඪɺ࠷ؔ࿈ੑͷߴ͍ॏΈΛɺ࠷ใྔͷଟ ͍ྔࢠԽΫϥελʹอଘ͢Δ͜ͱͰ͢ɻ ࣮ݧͷ݁Ռɺ͜ͷ৽͍͠Τϯτϩϐʔ੍͖XAIௐྔࢠԽʢECQxʣ๏ɺϞσϧͷ ੑೳΛҡ࣋·্ͨͤ͞ͳ͕Βɺਫ਼ʢ2ʙ5ϏοτʣͰಉ࣌ʹεύʔεͳχϡʔϥϧωοτϫʔΫΛੜ͢Δ͜ͱ͕͔ Γ·ͨ͠ɻ·ͨɼύϥϝʔλͷਫ਼͕͘ɼθϩཁૉͷ͕ଟ͍͜ͱ͔ΒɼϑΝΠϧαΠζͷͰѹॖੑ͕ߴ͘ɼߴਫ਼ͷྔ ࢠԽ͞Ε͍ͯͳ͍DNNϞσϧͱൺֱͯ͠ɼ࠷େͰ 103ഒͷѹॖޮՌ͕ಘΒΕ·͢ɽզʑͷΞϓϩʔνɺ༷ʑͳλΠϓͷϞσϧ ͱσʔληοτʢGoogle Speech CommandsCIFAR-10ͳͲʣͰධՁ͞Εɺաڈͷݚڀͱൺֱ͞Ε·ͨ͠ɻ http://arxiv.org/abs/2109.04236v1 w తɿϞσϧͷੑೳΛҡ࣋ͭͭ͠ྔࢠԽͯ͠αΠζѹॖ͢Δ w Ռɿ7((Ͱςετͯ͠࠷େͰഒͷѹॖޮՌ͕ಘΒΕͨ w ํ๏ɿઆ໌Մೳͳ"* 9"* ͷ֓೦ͱใཧͷ֓೦Λ׆༻ w ݻ༗໊ɿ&$2 &$2Y w ஶऀॴଐɿϑϥϯϗʔϑΝʔ))*ݚڀॴ υΠπ #*'0-% υΠπ -
10. େنͳࣄલτϨʔχϯάͷݶքΛ୳Δ (ݪจ: Exploring the Limits of Large Scale Pre-training)
ॏෳ http://arxiv.org/abs/2110.02095v1
ecoopnet
institutdelelevage
mickey_0226
hayataka88
hanbao
.jpg){kind=avatar}
ktatsuya
nozakijcom
trycycle
takeofuture
.jpg){kind=avatar}
lychee1223
ykiyota
akasan
analokmaus
baasie
chriscoyier
jacobtomlinson
csswizardry
sergeychernyshev
jcasabona
rakko
techseoconnect
leimatthew05
rasagy
tmiket
reverentgeek
