AI最新論文読み会2022年まとめ

AI最新論文読み会2022年まとめ

Transcript

1. େࡕެཱେֶɹ২ాେथ
   AI࠷৽࿦จಡΈձ
   2022೥1೥·ͱΊ
   

   View Slide

2. 2022೥·ͱΊ
   AI࠷৽࿦จಡΈձ
   ɾϝΠϯ
   ConvNeXt (2݄ൃද): ࢖͍΍͍͢࠷ۙͷߴੑೳϞσϧ
   
   GLIDE (1݄ൃද)ɹςΩετtoը૾ੜ੒
   
   Imagic (11݄ൃද)ɹࡉ΍͔ͳमਖ਼
   
   AudioLM (10݄ൃද): Ի੠ੜ੒Ϟσϧ
   
   Socratic Models (5݄ൃද): ൚༻AI (AGI)
   
   ɾͦͷଞ
   Wav2Vec 2 (7݄ൃද): NeuroAI?Brain-inspired AI (AIͱਓؒͷ೴ͷؔ܎Λ୳Δ෼໺)
   
   Algorithmic Imprint (7݄ൃද): AI࡞੒ऀͷྙཧ
   

   View Slide

3. 2022೥·ͱΊ
   Text
   GLIDE, Imagic
   AGI
   ࣗવݴޠॲཧΛج൫ͱͨ͠൚ਓ޻஌ೳ(AGI)΁ͷνϟϨϯδͷ1೥ɻ
   CV
   ConvNeXt
   AudioLM
   Speech
   Socratic model
   Diffusion
   

   View Slide

4. 2022೥·Ͱ·ͱΊ
   Self-Attention
   2017೥ 2018೥
   BERT
   2020೥
   DETR
   ViT
   GPT3
   2021೥
   CLIP
   wav2vec 2
   w2v-BERT
   BigSSL
   2019೥
   GPT2
   SwinT
   DDPM ADM
   

   View Slide

5. 2022೥·ͱΊ
   Text
   GLIDE, Imagic
   AGI
   ࣗવݴޠॲཧΛج൫ͱͨ͠൚ਓ޻஌ೳ(AGI)΁ͷνϟϨϯδͷ1೥ɻ
   CV
   AudioLM
   Speech
   Socratic model
   Diffusion
   ConvNeXt
   

   View Slide

6. ConvNeXt: CNN x SwinTransformer
   ը૾෼ྨϞσϧͷstate-of-the-art
   

   View Slide

7. ConvNeXt: CNN x SwinTransformer
   ޻෉఺ͷ·ͱΊ: ϕʔε͸ResNet
   

   View Slide

8. ConvNeXt
   ·ͣॳΊʹɻ
   

   View Slide

9. ConvNeXt
   ֤εςʔδͷ܁Γฦ͠਺ΛSwinTʹ͚ۙͮΔ
   

   View Slide

10. ConvNeXt
    4×4 non-overlapping convolution
    ৞ΈࠐΈͷύονԽ
    

    View Slide

11. ConvNeXt
    Depthwise convolutionಋೖޙɺ෯Λ޿͛Δ
    

    View Slide

12. ConvNeXt
    Inverted bottleneck(Narrow→Wide→Narrow)ߏ଄ͷಋೖ
    TransformerͰ͸֦େ཰4ഒΛ࢖༻ɻ※MobileNetͰ͸֦େ཰͸6ഒɻ
    શମͱͯ͠ͷܭࢉྔ͸ݮΔ͕ɺConvͷԋࢉ͸૿Ճɻ
    SwinTͰ͸ίί
    

    View Slide

13. ConvNeXt
    Depthwise convolutionͷҠಈ
    ※Depthwise ConvolutionͰେ͖ͳΧʔωϧαΠζ࢖͏ͨΊ
    Ұ࣌తʹConvͷԋࢉྔݮগͰੑೳѱԽɻ
    SwinTͰ͸ίί
    MSAϒϩοΫ͕FFNΑΓ΋ઌ಄ʹ͋Δ
    

    View Slide

14. ConvNeXt
    SwinTransformerͷΧʔωϧαΠζ(7)ΛਅࣅΔ
    Depthwise convolutionͷ
    
    ΧʔωϧαΠζେ͖͍ͯ͘͘͠ɻ
    7Ͱੑೳ͕๞࿨(SwinTͱಉ͡)
    ↓
    

    View Slide

15. ConvNeXt
    ࡉ͔ͳSwinT or ViTͷ޻෉Λಋೖ
    ReLU→GELU
    NormalizationݮΒ͢
    BN→LN
    μ΢ϯαϯϓϧ૚Λ੾Γ཭͠
    

    View Slide

16. ConvNeXt
    ݁Ռ
    

    View Slide

17. ConvNeXt
    ResNetΛSwinTransformerԽͯ͠ɺ
    CNN͚ͩͰState-of-the-artग़ͨΑɻ
    

    View Slide

18. View Slide

19. 2022೥·ͱΊ
    Text AGI
    ࣗવݴޠॲཧΛج൫ͱͨ͠൚ਓ޻஌ೳ(AGI)΁ͷνϟϨϯδͷ1೥ɻ
    CV
    ConvNeXt
    AudioLM
    Speech
    Socratic model
    Diffusion
    GLIDE, Imagic
    

    View Slide

20. GLIDE
    Stable Di
    ff
    usionͷجૅϞσϧ
    

    View Slide

21. Diffusion model
    ੜ੒Ϟσϧ
    

    View Slide

22. Diffusion modelͷྺ࢙
    DDPM ADM GLIDE
    CLIP
    ↓
    ҆ఆԽɺߴղ૾౓Խ ݴޠΛѻ͏
    

    View Slide

23. Diffusion modelͷྺ࢙
    DDPM ADM GLIDE
    CLIP
    ↓
    

    View Slide

24. DDPM: diffusion modelͷ࢝·Γ
    DNN
    Image
    Noise
    Image
    

    +
    

    Noise
    ਪ࿦ͨ͠
    
    
    Noise
    ࣌ࠁ৘ใ
    ೋ৐ޡࠩ
    
    
    ࠷খԽ
    

    View Slide

25. Diffusion modelͷྺ࢙
    DDPM ADM GLIDE
    CLIP
    ↓
    

    View Slide

26. ADM: ϞσϧΛ2ͭʹ෼͚ͯɺߴղ૾౓Խʹ੒ޭɻ
    Base Upsampler
    ෼ྨ ߴղ૾
    Classi
    fi
    er guidance
    
    
    (CNN)
    

    View Slide

27. GLIDE = CLIP x Diffusion model
    Di
    ff
    usion modelͷྺ࢙
    DDPM ADM GLIDE
    CLIP
    ↓
    

    View Slide

28. CLIP: ը૾ͱςΩετͷڮ౉͠
    

    View Slide

29. CLIP: ը૾ͱςΩετͷڮ౉͠
    ը૾ͱςΩετΛ௚઀ൺֱͰ͖ΔΑ͏ʹಛ௃ม׵Ͱ͖ΔϞσϧ
    ViT: Image
    Transformer: Text
    ίαΠϯྨࣅ౓
    

    View Slide

30. ADM
    Base Upsampler
    ෼ྨ ߴղ૾
    Classi
    fi
    er guidance
    
    
    (CNN)
    

    View Slide

31. GLIDE = ADM-basedʹCNNΛCLIPʹมߋ
    ADM-basedʹCNNΛCLIPʹมߋ
    Base Upsampler
    ෼ྨ ߴղ૾
    Classi
    fi
    er guidance
    
    
    (CLIP)
    

    View Slide

32. Imagic: Stable DiffusionͷվྑςΫχοΫ
    Stable Di
    ff
    usionͷվྑςΫχοΫ
    

    View Slide

33. Imagic
    Overview
    

    View Slide

34. View Slide

35. 2022೥·ͱΊ
    Text
    GLIDE, Imagic
    AGI
    ࣗવݴޠॲཧΛج൫ͱͨ͠൚ਓ޻஌ೳ(AGI)΁ͷνϟϨϯδͷ1೥ɻ
    CV
    ConvNeXt
    Speech
    Socratic model
    Diffusion
    AudioLM
    

    View Slide

36. AudioLM
    Իͷੜ੒Ϟσϧ
    

    View Slide

37. AudioLM = w2v-BERT x SoundStream
    Overview
    ɾจষͱΦʔσΟΦͷؒʹ͸Ұରଟͷؔ܎͕͋Δɻ
    
    ɾΦʔσΟΦ͸ςΩετʹൺͯ͠σʔλྔ͕ଟ͍ɻ
    

    View Slide

38. SoundStream
    ԻΛྔࢠԽ͢Δ
    

    View Slide

39. w2v-BERT
    Contrastive LearningͱMasked Language Modelingͷ૊Έ߹Θͤ
    

    View Slide

40. View Slide

41. 2022೥·ͱΊ
    Text
    GLIDE, Imagic
    AGI
    ࣗવݴޠॲཧΛج൫ͱͨ͠൚ਓ޻஌ೳ(AGI)΁ͷνϟϨϯδͷ1೥ɻ
    CV
    ConvNeXt
    AudioLM
    Speech
    Diffusion
    Socratic model
    

    View Slide

42. Socratic models
    طଘֶशࡁΈϞσϧΛ૊Έ߹Θͤͨ(४ʁ)൚ਓ޻஌ೳϞσϧ
    

    View Slide

43. Socratic models
    Overview
    Language is an intermediate representation
    

    View Slide

44. Socratic models
    Overview
    طଘͷVLM (Visual Language Model)ɺLMs (Large Language Model) ɺ ALMs (Audio Language Model)ͷಉ࢜
    が
    ɺߏ଄Խ͞Εͨର࿩Λߦ͏ɻ
    
    ͦͯ͠ɺ
    ビデ
    ΦαʔνɺΩϟ
    プ
    γϣϯੜ੒ɺ
    ビデ
    ΦQ&A (ະ஌ͷλεΫ)ɺকདྷͷߦಈ༧ଌΛ͜ͷର࿩ۭؒ΁ͷ৽͍͠ࢀՃऀͱͯ͠ѻ͏ ɻ
    

    View Slide

45. Socratic models
    ྫࣔ̍ɿجຊฤ
    

    View Slide

46. Socratic models
    ྫࣔ̎ɿԠ༻ฤ
    

    View Slide

47. Socratic models
    ιΫϥςεର࿩ͱ͸ʁ
    

    View Slide

48. View Slide

49. Others: NeuroAIᶃ
    ೴ͷػೳͱݴޠϞσϧͷରԠΛ୳Δ
    

    View Slide

50. Others: NeuroAIᶃ
    શମ૾: Wav2Vec 2Λֶश͠ɺͦͷ݁Ռ͔ΒfMRIͷBOLDΛ༧ଌ͢ΔWΛ࡞੒ɾ݁Ռݕূ
    

    View Slide

51. Others: NeuroAIᶃ
    ฏۉԽͨ͠೴ͷ׆ੑͷදݱɻ
    

    View Slide

52. Others: NeuroAIᶃ
    ϞσϧͷϨΠϠʔͷਂ͞ͱ೴ͷ෦ҐʹରԠ͕͋ͬͨɻ
    

    View Slide

53. Others: NeuroAIᶄ
    ೴೾͔ΒݴޠΛੜ੒͢Δ
    

    View Slide

54. Others: NeuroAIᶄ
    ϞσϧͷτϨʔχϯάηογϣϯ
    81िؒʹΘͨΓ50ճͷηογϣϯ
    
    ݽཱޠλεΫͱจষλεΫ
    
    λʔήοτͷ୯ޠ΍จষ͕ը໘্ͷจࣈͱͯ͠
    
    ඃݧऀʹࢹ֮తʹఏࣔ͞Εඃݧऀ͸
    
    ͦͷ୯ޠ΍จষΛੜ੒͠Α͏ͱͨ͠ɻ
    
    ݽཱޠλεΫͰ͸ɺ50ݸͷӳ୯ޠηοτ͔Βݸʑͷ୯ޠΛੜ੒ɻ
    
    จষλεΫͰ͸ɺ50୯ޠηοτ͔ΒͳΔӳޠจ͔Β୯ޠྻΛੜ੒ɻ
    

    View Slide

55. Others: NeuroAIᶄ
    Ϟσϧͷ݁Ռ
    จষ͸75%ͷਫ਼౓
    
    ୯ޠ͸93%ͷਫ਼౓
    

    View Slide

56. View Slide

57. Others: AI Ethics
    ྙཧ
    

    View Slide

58. Algorithmic Imprint
    Ξϧ
    ゴ
    Ϧ
    ズ
    ϜʹΑΔ֐
    が
    ൃੜͨ͠৔߹ͷҰൠత
    で
    ߹ཧతͳରࡦͱͯ͠ɺͦͷ༗֐ͳӨڹ
    が
    ͞Βʹ఻ൖ͢ΔͷΛ๷
    ぐ
    ͨΊʹ
    Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝͷ࢖༻ఀࢭ
    が
    Α͘ߦΘΕΔ
    が
    ɺఀࢭ͔ͨ͠Βͱݴͬͯެฏੑɺઆ໌੹೚ɺಁ໌ੑɺྙཧͷ໰୊
    が
    ͳ͘ͳΔ
    Θ͚
    で
    ͸ͳ͍
    
    →͜ͷ༗֐ͳΞϧ
    ゴ
    Ϧ
    ズ
    ϜͷӨڹ͸ɺΞϧ
    ゴ
    Ϧ
    ズ
    Ϝ࡟আҎ߱΋௕͘Өڹ͠ଓ͚Δ(Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝͷࠟ੻)
    
    ྫ: ӳࠃΛڌ఺ͱ͢Δߴߍͷଔۀূॻࢼݧ
    で
    ͋ΔGCEࢼݧͷΞϧ
    ゴ
    Ϧ
    ズ
    ϜʹΑΔධՁΛऔΓר͘໰୊(2020)
    
    ■
    ど
    ͷΑ͏ͳࢼݧ͔?
    
    ɾ 160͔ࠃҎ্
    で
    ࣮ࢪ͞Ε͍ͯΔ(ͦͷଟ͘͸ӳࠃͷݩ২ຽ஍)ࠃࡍతʹೝΊΒΕͨࢼݧ
    
    ɾ AϨ
    ベ
    ϧͷ੒੷͸ඞવత
    で
    ͋Γɺେֶ΁ͷೖֶʹෆՄܽͳ໾ׂΛՌͨ͢
    
    ■ܦҢ
    
    ɾCOVID-19ͷେྲྀߦʹΑΓGCEࢼݧΛ؂ಜ͢ΔӳࠃʹຊڌΛஔ͘४੓෎ػؔ
    で
    ͋ΔOfqual͸ର໘ࢼݧΛதࢭͨ͠
    
    ɾࢼݧͷ୅ΘΓʹɺֶߍ
    で
    ͷੜెͷաڈͷ੒੷ɺڭࢣͷධՁΛ࢖༻ͯ͠Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝ
    で
    ੒੷Λ࡞੒ͨ͠
    
    →݁Ռɺੈքతͳ߅ٞߦಈ
    が
    ຄൃ͠ɺΞϧ
    ゴ
    Ϧ
    ズ
    Ϝ͸࡟আ͞Εͨ
    
    ɹڭࢣଆ: ͦ΋ͦ΋աڈͷੜెͷධՁΛه࿥͍ͯ͠ͳ͍
    
    ɹੜెଆ: ੒੷ʹରͯ͠ਅ݋ʹऔΓ૊Μ
    で
    ͍ͳ͔ͬͨ(ࢼݧ
    が
    શͯͳͷ
    で
    ௚લͷ30~60೔ʹ໠ษڧ͢Δੜె
    が
    ଟ͍)
    
    ɾΞϧ
    ゴ
    Ϧ
    ズ
    Ϝ͸࡟আ͞Εͨ
    が
    ɺֶੜͷ࠶ධՁ͸ߦΘΕͳ͔ͬͨɻ
    
    ͢ͳΘͪɺ࠾఺ํ๏͸มΘͬͨ
    が
    ɺΞϧ
    ゴ
    Ϧ
    ズ
    ϜͷӨڹΛେ͖͘ड͚͍ͯͨ(Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝͷࠟ੻)
    

    View Slide

59. Algorithmic Imprint
    ■Algorithmic Imprint(Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝͷࠟ੻)Λҙࣝͨ͠Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝͷઃܭ
    
    ʮΞϧ
    ゴ
    Ϧ
    ズ
    Ϝͷࠟ੻ʯΛҙࣝͨ͠ઃܭͷߟ͑ํʹΑΓɺΞϧ
    ゴ
    Ϧ
    ズ
    Ϝ։ൃ
    プ
    ϩηεΛΑΓެฏ
    で
    ࣾձٕज़తͳ
    
    ৘ใʹج
    づ
    ͍ͨ΋ͷʹ͢Δ͜ͱ
    がで
    ͖Δɻ
    
    (1)Ξϧ
    ゴ
    Ϧ
    ズ
    ϜͷӨڹ Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝ͸࡟আͨ͠ޙʹ΋ར֐ؔ܎ऀʹӨڹΛٴ
    ぼ
    ͢ɻ։ൃऀͱӡӦऀ͸Ξϧ
    ゴ
    Ϧ
    ズ
    ϜΛ࡟আ
    ͢Δ
    だ
    ͚
    で
    ͳ͘ɺΞϧ
    ゴ
    Ϧ
    ズ
    ϜʹΑΔة֐Λੋਖ਼͠ɺઆ໌੹೚
    が
    ࣋ଓͯ͠ཁٻ͞ΕΔɻ
    
    (2)Ξϧ
    ゴ
    Ϧ
    ズ
    Ϝઃܭͷઆ໌੹೚ ։ൃऀ͸ʮΞϧ
    ゴ
    Ϧ
    ズ
    Ϝͷࠟ੻ʯͷӨڹΛड͚Δਓʹ֐ΛΑΓೝࣝ
    で
    ͖ΔΑ͏ʹ͢Δ
    べ
    ͖
    で
    ͋Δɻ
    
    (3)AIྙཧ
    ガ
    バ
    φϯε
    で
    ิڧ͢Δ

    ٕज़తͳհೖ
    だ
    ͚
    で
    ͸֐Λ࡟ݮ͢Δ͜ͱ͸
    で
    ͖ͳ͍ɻ ʮΞϧ
    ゴ
    Ϧ
    ズ
    Ϝͷࠟ੻ʯΛҙࣝͨ͠Ξϧ
    ゴ
    Ϧ
    ズ
    ϜઃܭΛ ద੾ͳAI
    ྙཧ
    ガ
    バ
    φϯε
    で
    ิ׬͢Δɻ
    

    View Slide

60. View Slide

61. 2023೥ʹ͍ͭͯ
    ʮզʑͷݚڀࣨʹ͔͠Ͱ͖ͳ͍͜ͱʯΛɻ
    Ҿ͖ଓ͖ษڧձ։࠵͢Δɻ
    ҩֶ΁ͷൺॏΛॏ͘͢Δɻ
    ҩྍը૾ݚڀ༻ϞσϧͷνϡʔτϦΞϧɾϋϯζΦϯ
    

    View Slide