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

【動画あり】Transformer論文解説

 【動画あり】Transformer論文解説

下記YouTube動画で使用したスライド資料です。
https://youtu.be/6tcjwdanedU

Avatar for 数理の弾丸

数理の弾丸

July 16, 2024
Tweet

More Decks by 数理の弾丸

Other Decks in Technology

Transcript

  1. 

  2. 5SBOTGPSNFSͷԠ༻ൣғ  ※: https://blog.google/products/search/search-language-understanding-bert/ FUD 5SBOTGPSNFS ςΩετ຋༁Λओ؟ͱͯ͠ఏҊ #&35 (15 ςΩετ෼ྨFUD

    ςΩετੜ੒ ෦෼ΞʔΩςΫνϟͷ࠾༻ ը૾΁ͷద༻ 7J5 %JGGVTJPO 5SBOTGPSNFS ը૾෼ྨFUD ը૾ੜ੒ $IBU(15 -MBNB 4UBCMF%JGGVTJPO 4PSB (PPHMFݕࡧ˞ $-*1 ۃΊͯ޿ൣғʹج൫ٕज़ͱͯ͠׆༂
  3. ܥྻϞσϦϯά  ॱংͷ͋Δཁૉͷ࿈ͳΓͱΈͳͤΔ΋ͷΛܥྻʢTFRVFODFʣͱݺͼɺ ͜ΕΛର৅ͱ͢ΔϞσϦϯάΛܥྻϞσϦϯάͱݺͿ ྫ ຋༁ ஫ҙػߏ͑͋͞Ε͹े෼ "UUFOUJPOJTBMMZPVOFFE ྫ ςΩετੜ੒

    Ͳ͏ͧΑΖ͓͘͠ئ͍͠·͢ɻ Կ͔࣭͝໰΍ϦΫΤετ͕͋Ε͹ڭ͍͑ͯͩ͘͞ɻ ΑΖ͘͠པΉ ྫ ߏจղੳ 4 /1ΑΖ͘͠  71པΉ ΑΖ͘͠པΉ
  4. ܥྻϞσϦϯά  ॱংͷ͋Δཁૉͷ࿈ͳΓͱΈͳͤΔ΋ͷΛܥྻʢTFRVFODFʣͱݺͼɺ ͜ΕΛର৅ͱ͢ΔϞσϦϯάΛܥྻϞσϦϯάͱݺͿ ྫ ຋༁ ஫ҙػߏ͑͋͞Ε͹े෼ "UUFOUJPOJTBMMZPVOFFE ྫ ςΩετੜ੒

    Ͳ͏ͧΑΖ͓͘͠ئ͍͠·͢ɻ Կ͔࣭͝໰΍ϦΫΤετ͕͋Ε͹ڭ͍͑ͯͩ͘͞ɻ ΑΖ͘͠པΉ ྫ ߏจղੳ 4 /1ΑΖ͘͠  71པΉ ΑΖ͘͠པΉ TPVSDF
  5. ܥྻϞσϦϯά  ॱংͷ͋Δཁૉͷ࿈ͳΓͱΈͳͤΔ΋ͷΛܥྻʢTFRVFODFʣͱݺͼɺ ͜ΕΛର৅ͱ͢ΔϞσϦϯάΛܥྻϞσϦϯάͱݺͿ ྫ ຋༁ ஫ҙػߏ͑͋͞Ε͹े෼ "UUFOUJPOJTBMMZPVOFFE ྫ ςΩετੜ੒

    Ͳ͏ͧΑΖ͓͘͠ئ͍͠·͢ɻ Կ͔࣭͝໰΍ϦΫΤετ͕͋Ε͹ڭ͍͑ͯͩ͘͞ɻ ΑΖ͘͠པΉ ྫ ߏจղੳ 4 /1ΑΖ͘͠  71པΉ ΑΖ͘͠པΉ UBSHFU
  6. ର߅അ͸Կ͔  ࠶ؼܕχϡʔϥϧωοτϫʔΫ ৞ΈࠐΈχϡʔϥϧωοτϫʔΫ -45.<)PDISFJUFS > (36<$IVOH > FUD #ZUF/FU<,BMDICSFOOFS

    > $POW44<(FISJOH > FUD ܥྻͷཁૉΛॱʹೖྗ͍ͯ͘͠ ฒྻܭࢉ͕Ͱ͖ͳ͍ ཁૉؒڑ཭ʹԠͨ͡ܭࢉྔ૿Ճ͕ݦஶ ௕ڑ཭ґଘͷֶश͕ࠔ೉
  7. ର߅അ͸Կ͔  ࠶ؼܕχϡʔϥϧωοτϫʔΫ ৞ΈࠐΈχϡʔϥϧωοτϫʔΫ -45.<)PDISFJUFS > (36<$IVOH > FUD #ZUF/FU<,BMDICSFOOFS

    > $POW44<(FISJOH > FUD ܥྻͷཁૉΛॱʹೖྗ͍ͯ͘͠ ฒྻܭࢉ͕Ͱ͖ͳ͍ ཁૉؒڑ཭ʹԠͨ͡ܭࢉྔ૿Ճ͕ݦஶ ௕ڑ཭ґଘͷֶश͕ࠔ೉ ฒྻԽ͕Մೳ͔ͭ௕ڑ཭ґଘΛֶशͰ͖ΔϞσϧͱͯ͠ 5SBOTGPSNFSΛఏҊʢ4FD ʣ
  8. طଘݚڀ͔ΒҾ͖ܧ͙΋ͷ  Τϯίʔμɾσίʔμػߏ FODPEFSEFDPEFSNFDIBOJTN ࣗݾ஫ҙػߏ TFMGBUUFOUJPONFDIBOJTN Τϯίʔμ σίʔμ ೖྗ ग़ྗ

    ಛ௃நग़ɾܥྻੜ੒ͷೋஈߏ͑ ࢲ ͸ ٢ా ࢲ ͸ ٢ా ࣗܥྻؒͰͷॏΈ෇͚ Τϯίʔμɾσίʔμͷ༗༻ੑΛੜ͔ͭͭ͠ ࣗݾ஫ҙػߏͰ׬݁͢ΔॳΊͯͷϞσϧʢ4FDʣ
  9. ·ͱΊ  ର߅അ͸Կ͔ ओఏҊ͸Կ͔ ධՁͷ8IBU3FTVMU ࠶ؼܕωοτϫʔΫ ৞ΈࠐΈωοτϫʔΫ ࣗݾ஫ҙͰ׬݁ͨ͠ ΤϯίʔμɾσίʔμΛఏҊ w

    ฒྻԽ͕༰қ w ௕ڑ཭ґଘΛଊ͑Δ .PEFM"SDIJUFDUVSF 8IZ4FMG"UUFOUJPO 5SBJOJOH 3FTVMUT
  10. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ Ґஔූ߸Խ Ґஔූ߸Խ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒
  11. Ґஔූ߸Խ ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ

    ϚεΫ෇͖ 
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ Τϯίʔμ Ґஔූ߸Խ
  12. Ґஔූ߸Խ ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ

    ϚεΫ෇͖ 
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ σίʔμ Ґஔූ߸Խ
  13. Ґஔූ߸Խ ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ

    ϚεΫ෇͖ 
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ σίʔμ Ґஔූ߸Խ ࣗݾճؼ BVUPSFHSFTTJPO ࣌ࠁ ͷग़ྗ͕ ͷೖྗʹͳΔػߏ t t + 1
  14. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ Ґஔූ߸Խ Ґஔූ߸Խ
  15. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ Ґஔූ߸Խ Ґஔූ߸Խ
  16. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ ͔͜͜Βઌͷॲཧ͸ೖྗςΩετͷ ޠॱΛೝࣝͰ͖ͳ͍ Ґஔූ߸Խ Ґஔූ߸Խ
  17. ຒΊࠐΈɾҐஔූ߸Խ  ς Ω ε τ τ Ϋ ϯ Խ

    ࢲ ͸ ٢ా ʜ ʜ ʜ ຒΊࠐΈ e1 e2 e3 ࣍ݩͷϕΫτϧ dmodel
  18. ຒΊࠐΈɾҐஔූ߸Խ ς Ω ε τ τ Ϋ ϯ Խ ࢲ

    ͸ ٢ా ʜ ʜ ʜ ຒΊࠐΈ e1 e2 e3 ʜ ʜ ʜ Ґஔූ߸Խ p1 p2 p3 ppos [2i] = sin ( pos 10000 2i dmodel ) ppos [2i + 1] = cos ( pos 10000 2i dmodel )
  19. ຒΊࠐΈɾҐஔූ߸Խ  ς Ω ε τ τ Ϋ ϯ Խ

    ࢲ ͸ ٢ా ʜ ʜ ʜ ຒΊࠐΈ e1 e2 e3 ʜ ʜ ʜ Ґஔූ߸Խ p1 p2 p3 ppos [2i] = sin ( pos 10000 2i dmodel ) ppos [2i + 1] = cos ( pos 10000 2i dmodel ) x1 x2 x3 = de1 + p1 = de2 + p2 = de3 + p3
  20. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ ࣗݾ஫ҙ TFMGBUUFOUJPO w ϕΫτϧྻΛจ຺Λߟྀ͠ͳ͕Βม׵ w 5SBOTGPSNFSͷ࠷΋ॏཁͳ෦඼ Ґஔූ߸Խ Ґஔූ߸Խ
  21. ࣗݾ஫ҙ  x1 x2 x3 Q = [ q1 ]

    [ q2 ] [ q3 ] K = [ k1 ] [ k2 ] [ k3 ] V = [ v1 ] [ v2 ] [ v3 ] qi = Wq xi , Wq ∈ ℝdk ×dmodel ki = Wk xi , Wk ∈ ℝdk ×dmodel vi = Wv xi , Wv ∈ ℝdv ×dmodel
  22. ࣗݾ஫ҙ  x1 x2 x3 Q = [ q1 ]

    [ q2 ] [ q3 ] K = [ k1 ] [ k2 ] [ k3 ] V = [ v1 ] [ v2 ] [ v3 ] h1 h2 h3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ͜Ε͔Βܭࢉ͍ͨ͠΋ͷ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K
  23. ࣗݾ஫ҙ  h1 h2 h3 = ◯v1 + ◯v2 +

    ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ͜Ε͔Βܭࢉ͍ͨ͠΋ͷ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K QK⊤ = [ q1 ] [ q2 ] [ q3 ] [ k1 ] [ k2 ] [ k3 ] = q1 ⋅ k1 q1 ⋅ k2 q1 ⋅ k3 q2 ⋅ k1 q2 ⋅ k2 q2 ⋅ k3 q3 ⋅ k1 q3 ⋅ k2 q3 ⋅ k3
  24. ࣗݾ஫ҙ  h1 h2 h3 = ◯v1 + ◯v2 +

    ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ͜Ε͔Βܭࢉ͍ͨ͠΋ͷ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K QK⊤ = [ q1 ] [ q2 ] [ q3 ] [ k1 ] [ k2 ] [ k3 ] = q1 ⋅ k1 q1 ⋅ k2 q1 ⋅ k3 q2 ⋅ k1 q2 ⋅ k2 q2 ⋅ k3 q3 ⋅ k1 q3 ⋅ k2 q3 ⋅ k3 ʮ٢ాʯ͔Βݟͨʮࢲʯͷॏཁ౓
  25. ࣗݾ஫ҙ  h1 h2 h3 = ◯v1 + ◯v2 +

    ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ͜Ε͔Βܭࢉ͍ͨ͠΋ͷ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K QK⊤ = [ q1 ] [ q2 ] [ q3 ] [ k1 ] [ k2 ] [ k3 ] = q1 ⋅ k1 q1 ⋅ k2 q1 ⋅ k3 q2 ⋅ k1 q2 ⋅ k2 q2 ⋅ k3 q3 ⋅ k1 q3 ⋅ k2 q3 ⋅ k3 ʮ٢ాʯ͔Βݟͨʮ͸ʯͷॏཁ౓
  26. ࣗݾ஫ҙ  h1 h2 h3 = ◯v1 + ◯v2 +

    ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ͜Ε͔Βܭࢉ͍ͨ͠΋ͷ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K QK⊤ = [ q1 ] [ q2 ] [ q3 ] [ k1 ] [ k2 ] [ k3 ] = q1 ⋅ k1 q1 ⋅ k2 q1 ⋅ k3 q2 ⋅ k1 q2 ⋅ k2 q2 ⋅ k3 q3 ⋅ k1 q3 ⋅ k2 q3 ⋅ k3 ʮ٢ాʯ͔Βݟͨʮ٢ాʯͷॏཁ౓
  27. ࣗݾ஫ҙ  h1 h2 h3 = ◯v1 + ◯v2 +

    ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ͜Ε͔Βܭࢉ͍ͨ͠΋ͷ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K softmax ( QK⊤ dk ) = a11 a12 a13 a21 a22 a23 a31 a32 a33 εέʔϦϯά ิ ͜ͷߦྻΛ஫ҙߦྻ BUUFOUJPONBUSJY ͱݺͿ ֤੒෼ ͸஫ҙॏΈ BUUFOUJPOXFJHIU ͱݺͿ aij
  28. ࣗݾ஫ҙ  h1 h2 h3 = a21 v1 + a22

    v2 + a23 v3 ࣗݾ஫ҙͷ࠷ऴग़ྗ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K softmax ( QK⊤ dk ) = a11 a12 a13 a21 a22 a23 a31 a32 a33 εέʔϦϯά = a11 v1 + a12 v2 + a13 v3 = a31 v1 + a32 v2 + a33 v3
  29. ࣗݾ஫ҙ  h1 h2 h3 = a21 v1 + a22

    v2 + a23 v3 ࣗݾ஫ҙͷ࠷ऴग़ྗ पลจ຺ͷࠞͥ߹Θͤ۩߹Λ˓ͷ஋͕ܾΊΔ ͜ΕΛ Λ༻͍ͯٻΊΔ Q, K softmax ( QK⊤ dk ) = a11 a12 a13 a21 a22 a23 a31 a32 a33 εέʔϦϯά = a11 v1 + a12 v2 + a13 v3 = a31 v1 + a32 v2 + a33 v3 ࣗݾ஫ҙ ʹपลจ຺Λߟྀ͢Δػߏ
  30. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ Ґஔූ߸Խ Ґஔූ߸Խ
  31. ϚεΫ෇͖ࣗݾ஫ҙ  softmax ( QK⊤ dk ) = a11 a12

    a13 a21 a22 a23 a31 a32 a33 [ 1 0 0 1 1 0 1 1 1 ] ϚεΫߦྻ
  32. ϚεΫ෇͖ࣗݾ஫ҙ  softmax ( QK⊤ dk ) = a11 a12

    a13 a21 a22 a23 a31 a32 a33 [ 1 0 0 1 1 0 1 1 1 ] ϚεΫߦྻ a11 a12 a13 a21 a22 a23 a31 a32 a33 ⊙ [ 1 0 0 1 1 0 1 1 1 ] = a11 0 0 a21 a22 0 a31 a32 a33
  33. ϚεΫ෇͖ࣗݾ஫ҙ  h1 h2 h3 = a21 v1 + a22

    v2 + 0v3 ࣗݾ஫ҙͷ࠷ऴग़ྗ softmax ( QK⊤ dk ) = a11 a12 a13 a21 a22 a23 a31 a32 a33 = a11 v1 + 0v2 + 0v3 = a31 v1 + a32 v2 + a33 v3 [ 1 0 0 1 1 0 1 1 1 ] ϚεΫߦྻ a11 a12 a13 a21 a22 a23 a31 a32 a33 ⊙ [ 1 0 0 1 1 0 1 1 1 ] = a11 0 0 a21 a22 0 a31 a32 a33
  34. ࣗݾ஫ҙ  h1 h2 h3 = ◯v1 + ◯v2 +

    ◯v3 = ◯v1 + ◯v2 + ◯v3 = ◯v1 + ◯v2 + ◯v3 ܭࢉ͍ͨ͠΋ͷ QK⊤ = [ q1 ] [ q2 ] [ q3 ] [ k1 ] [ k2 ] [ k3 ] = q1 ⋅ k1 q1 ⋅ k2 q1 ⋅ k3 q2 ⋅ k1 q2 ⋅ k2 q2 ⋅ k3 q3 ⋅ k1 q3 ⋅ k2 q3 ⋅ k3 softmax ( QK⊤ dk ) = a11 a12 a13 a21 a22 a23 a31 a32 a33
  35. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ Ґஔූ߸Խ Ґஔූ߸Խ
  36. 1PTJUJPOXJTF'FFE'PSXBSE/FUXPSLT  ࣗ ݾ ஫ ҙ ࢲ ͸ ٢ా ʜ

    ʜ ʜ จ຺ԽຒΊࠐΈ h1 h2 h3 ReLU(h1 W1 + b1 )W2 + b2 ReLU(h2 W1 + b1 )W2 + b2 ReLU(h3 W1 + b1 )W2 + b2 ϕΫτϧͦΕͧΕʹରͯ͠'FFEGPSXBSE/FUXPSLΛద༻
  37. 1PTJUJPOXJTF'FFE'PSXBSE/FUXPSLT  ࣗ ݾ ஫ ҙ ࢲ ͸ ٢ా ʜ

    ʜ ʜ จ຺ԽຒΊࠐΈ h1 h2 h3 ReLU(h1 W1 + b1 )W2 + b2 ReLU(h2 W1 + b1 )W2 + b2 ReLU(h3 W1 + b1 )W2 + b2 ϕΫτϧͦΕͧΕʹରͯ͠'FFEGPSXBSE/FUXPSLΛద༻ 5SBOTGPSNFSʹ͓͚Δ''/ͷ໾ׂʹ͍ͭͯ͸ͦͷޙ͞·͟·ͳٞ࿦͕͋Δ w (FWB .PS FUBM5SBOTGPSNFSGFFEGPSXBSEMBZFSTBSFLFZWBMVFNFNPSJFTBS9JWQSFQSJOUBS9JW   w ;IBOH ;IFOHZBO FUBM.PF fi DBUJPO5SBOTGPSNFSGFFEGPSXBSEMBZFSTBSFNJYUVSFTPGFYQFSUTBS9JWQSFQSJOUBS9JW   w (FWB .PS FUBM5SBOTGPSNFSGFFEGPSXBSEMBZFSTCVJMEQSFEJDUJPOTCZQSPNPUJOHDPODFQUTJOUIFWPDBCVMBSZTQBDFBS9JWQSFQSJOUBS9JW   w FUD
  38. ఏҊϞσϧͷΞʔΩςΫνϟ  Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ 'FFE'PSXBSE 
 /FUXPSL Ճࢉਖ਼نԽ ϚεΫ෇͖

    
 Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ϛϧνϔου ࣗݾ஫ҙ Ճࢉਖ਼نԽ Ճࢉਖ਼نԽ ೖྗςΩετ ग़ྗςΩετ 'FFE'PSXBSE 
 /FUXPSL ઢܗม׵ ιϑτϚοΫεؔ਺ ֬཰෼෍ 🌟 🌟 🌟 /ʷ ʷ/ ˞ਤ͸7BTXBOJ ͷ'JHVSFΛϕʔεͱͯ͠࡞੒ Ґஔූ߸Խ Ґஔූ߸Խ
  39. 8IBU  λεΫ ػց຋༁ σʔληοτ 8.5 w χϡʔε຋༁ͷֶशɾධՁσʔληοτ w FOEFNJMMJPOTFOUFODFQBJST

    w FOGSNJMMJPOTFOUFODFQBJST ධՁࢦඪ w #-&6ʢ຋༁ੑೳʣ w '-01Tʢܭࢉྔʣ
  40. ·ͱΊ  ର߅അ͸Կ͔ ओఏҊ͸Կ͔ ධՁͷ8IBU3FTVMU ࠶ؼܕωοτϫʔΫ ৞ΈࠐΈωοτϫʔΫ ࣗݾ஫ҙͰ׬݁ͨ͠ ΤϯίʔμɾσίʔμΛఏҊ w

    ฒྻԽ͕༰қ w ௕ڑ཭ґଘΛଊ͑Δ 5SBOTGPSNFSͷػߏ Ґஔූ߸Խ 8IBUػց຋༁ 3FTVMU405" ଞϞσϧʹඖఢ͢ΔੑೳΛ ΑΓগͳֶ͍शίετͰ࣮ݱ
  41. ֶशύϥμΠϜͷมભ    ݱࡏ εΫϥονֶश ϑΝΠϯνϡʔχϯά *ODPOUFYUMFBSOJOH ಛఆͷλεΫʹಛԽͨ͠ϞσϧΛ ϥϯμϜͳ஋ΛͱΔύϥϝλ͔Βֶश

    ࣄલֶशࡁΈϞσϧΛ ݸผλεΫ޲͚ʹඍௐ੔ Ϟσϧͦͷ΋ͷΛௐ੔ͤͣ ࢦࣔʹै༷ͬͯʑͳλεΫΛ͜ͳ͢ w ࠶ؼܕωοτϫʔΫ w ৞ΈࠐΈωοτϫʔΫ w 5SBOTGPSNFS w #&35 w (15  w 3FT/FU w (15 w -MBNB w 1B-.
  42. ֶशύϥμΠϜͷมભ    ݱࡏ εΫϥονֶश ϑΝΠϯνϡʔχϯά *ODPOUFYUMFBSOJOH ಛఆͷλεΫʹಛԽͨ͠ϞσϧΛ ϥϯμϜͳ஋ΛͱΔύϥϝλ͔Βֶश

    ࣄલֶशࡁΈϞσϧΛ ݸผλεΫ޲͚ʹඍௐ੔ Ϟσϧͦͷ΋ͷΛௐ੔ͤͣ ࢦࣔʹै༷ͬͯʑͳλεΫΛ͜ͳ͢ w ࠶ؼܕωοτϫʔΫ w ৞ΈࠐΈωοτϫʔΫ w 5SBOTGPSNFS w #&35 w (15  w 3FT/FU w (15 w -MBNB w 1B-. #&35ʹΑΔϑΝΠϯνϡʔχϯά(15ʹΑΔ*ODPOUFYUMFBSOJOH ͕ಛʹΤϙοΫϝΠΩϯά
  43. ςΩετΛ௒͑ͨ׆༂  %PTPWJUTLJZ   'JH 3BEGPSE   'JH

    7JTJPO5SBOTGPSNFS $-*1 ࣗݾ஫ҙͷՄೳੑΛ୳Δޙଓݚڀ͕ଟൃ
  44. ·ͱΊ  ର߅അ͸Կ͔ ओఏҊ͸Կ͔ ධՁͷ8IBU3FTVMU ࠶ؼܕωοτϫʔΫ ৞ΈࠐΈωοτϫʔΫ ࣗݾ஫ҙͰ׬݁ͨ͠ ΤϯίʔμɾσίʔμΛఏҊ w

    ฒྻԽ͕༰қ w ௕ڑ཭ґଘΛଊ͑Δ 5SBOTGPSNFSͷػߏ Ґஔූ߸Խ 8IBUػց຋༁ 3FTVMU405" ଞϞσϧʹඖఢ͢ΔੑೳΛ ΑΓগͳֶ͍शίετͰ࣮ݱ