アニーリングマシンを活用したエッジAIにおける 生成モデルの学習効率化のためのアーキテクチャ

Transcript

1. ͘͞ΒΠϯλʔωοτגࣜձࣾ
   (C) Copyright 1996-2019 SAKURA Internet Inc
   ͘͞ΒΠϯλʔωοτݚڀॴ
   ΞχʔϦϯάϚγϯΛ׆༻ͨ͠ΤοδAIʹ͓͚Δ
   ੜ੒Ϟσϧͷֶशޮ཰ԽͷͨΊͷΞʔΩςΫνϟ
   2019/12/06
   ϚϧνϝσΟΞɺ෼ࢄɺڠௐͱϞόΠϧʢDICOMO2020ʣγϯϙδ΢Ϝ
   ɹ௽ా തจɼদຊ ྄հ
   ͘͞ΒΠϯλʔωοτגࣜձࣾ ͘͞ΒΠϯλʔωοτݚڀॴ
   

   View Slide

2. 2
   1. എܠͱ໨త
   2. ΤοδAIʹ͓͚Δੜ੒Ϟσϧͷ׆༻ͱ՝୊
   3. ఏҊ͢ΔΞʔΩςΫνϟ
   4. ධՁͱߟ࡯
   5. ·ͱΊ
   ໨࣍
   

   View Slide

3. 1.
   എܠͱ໨త
   

   View Slide

4. 4
   ΤοδAI
   • IoTσόΠε͕ීٴʹ൐͍ɼσόΠε͔Βੜ੒͞ΕΔσʔλྔ͕ٸ଎ʹ૿Ճ͍ͯ͠Δɽ
   • IoTσόΠε͕ੜ੒͢Δଟछଟ༷ͳσʔλΛ׆༻ͨ͠՝୊ղܾͷΞϓϩʔνͱͯ͠ɼ
   ਓ޻஌ೳ(AI)ٕज़Λ༻͍ͯΤϯυσόΠε্Ͱ஌తͳλεΫΛ࣮ߦ͢ΔΤοδAI͕
   ஫໨͞Ε͍ͯΔɽ
   ࣗಈӡసंɼ΢ΣΞϥϒϧ୺຤ʢεϚʔτ΢ΥονɼεϚʔτάϥεʣɼ
   ࢈ۀϩϘοτɼ؂ࢹΧϝϥɼetc.
   ظ଴͞ΕΔԠ༻෼໺
   

   View Slide

5. 5
   ΤοδAI࣮ݱͷͨΊͷγεςϜܗଶ
   ΤοδAIͷ࣮ݱํ๏͸ɼܭࢉίετ͕ߴ͍AIͷֶशͱਪ࿦ͷϓϩηεΛͲͷܭࢉࢿݯ্Ͱ
   ࣮ߦ͢Δ͔ʹΑΓଟذʹ౉͓ͬͯΓɼ໨త΍ཁٻʹԠͯ͡ద੾ʹબ୒͢Δɽ
   Ϋϥ΢υͰॲཧ͢Δ৔߹
   ΤοδAIͷγεςϜܗଶͷྫ
   Cloud
   ֶश
   ਪ࿦
   Edge
   ֶश΍ਪ࿦ͷҰ෦ΛσόΠεΑΓ५୔ͳॲཧೳྗ
   Λ༗͢ΔΫϥ΢υͰߦ͏
   σόΠεɾΤοδͰॲཧ͢Δ৔߹
   Cloud
   ֶश
   ਪ࿦
   ֶश΍ਪ࿦ͷҰ෦ΛσόΠε্ɼ͋Δ͍͸σόΠε
   ʹ෺ཧతʹ͍ۙωοτϫʔΫ্ͷܭࢉࢿݯͰߦ͏
   ֶश
   ਪ࿦
   ྫ͑͹ɼσόΠε͕ਓؒʹର͢ΔΠϯλʔϑΣʔεΛ୲͏΋ͷΛ૝ఆͨ͠৔߹ɼਓ͕ؒ஗ԆΛײ͡ͳ͍
   ਺ेmsΦʔμʔͷԠ౴ੑ͕ཁٻ͞ΕΔͨΊɼ௨৴஗ԆͷΈͰཁٻ͞ΕΔԠ౴࣌ؒΛ௒͑Δɽ
   ͜ͷ৔߹ɼΫϥ΢υΛհ͞ͳ͍γεςϜܗଶ͕๬·ΕΔɽ
   ਺ඦmsΦʔμʔͷ
   ௨৴஗Ԇ͕ൃੜ
   ਺msΦʔμʔͷ
   ௨৴஗ԆΛ཈͑Δ
   ਪ࿦
   ਪ࿦
   

   View Slide

6. 6
   ΤοδAIʹ͓͚Δੜ੒Ϟσϧ
   • ؍ଌσʔλ͸֬཰తʹੜ੒͞Ε͍ͯΔͱԾఆ͠ɼͦͷੜ੒աఔΛ
   ֬཰෼෍ͱͯ͠ϞσϧԽ͢ΔΞϓϩʔν (=֬཰తੜ੒Ϟσϧ)
   • ֶशޙͷϞσϧ͔Β৽ͨʹσʔλΛੜ੒Ͱ͖Δ
   ੜ੒Ϟσϧͱ͸
   • ଟ͘ͷ֬཰తੜ੒Ϟσϧͷֶशɾਪ࿦ʹ͸ɼҰൠʹܭࢉίετ͕ߴ͍֬཰෼෍(໬౓ؔ਺)͔Βͷ
   αϯϓϦϯά͕ඞཁ
   • σόΠε͸Ұൠʹॲཧೳྗ͕௿͍ͨΊɼσόΠε্Ͱֶशɾਪ࿦Λ׬݁͢Δͷ͸೉͍͠ɽ
   ੜ੒Ϟσϧͷֶश
   [1] R. Salakhutdinov and G. Hinton, “Deep Boltzmann Machines”, in Proceedings of the International Conference on Artificial Intelligence and Statistics, 2009, pp. 448-455.
   • ΤϯυσόΠεʹը૾ɾจষੜ੒౳ͷ૑଄తλεΫ΍ҟৗݕ஌ɼ
   ऩूσʔλͷܽଛ஋ิ׬ɾϊΠζআڈͳͲ෯޿͍Ԡ༻ੑΛ΋ͨΒ͢ɽ ਂ૚ϘϧπϚϯϚγϯʹΑΔը૾ੜ੒[1]
   pθ
   (x)
   pdata
   (x)
   ਅͷ෼෍(σʔλ෼෍) ੜ੒Ϟσϧ
   ੜ੒Ϟσϧ Λσʔλ෼෍
   ʹ͚ۙͮΔΑ͏ʹֶश͢Δ
   pθ
   (x) pdata
   (x)
   

   View Slide

7. 7
   ຊݚڀͷ໨త
   • ΤοδAIʹ͓͍ͯɼߴ͍Ԡ༻ੑΛ࣋ͭੜ੒ϞσϧΛΫϥ΢υΛհͣ͞ʹΤϯυ
   σόΠε্Ͱ׆༻͢ΔͨΊͷΞʔΩςΫνϟͷఏҊ͢Δɽ
   • ൚༻త͔ͭ௿εϖοΫͳσόΠε্Ͱੜ੒Ϟσϧͷ׆༻ΛՄೳʹ͢Δɽ
   • Ϋϥ΢υΛհͣ͞ʹੜ੒ϞσϧΛޮ཰తʹֶश͠ɼ͔ͭਪ࿦࣌ͷσόΠε
   ͷԠ౴ੑΛ޲্ͤ͞Δɽ
   

   View Slide

8. 2.
   ΤοδAIʹ͓͚Δੜ੒Ϟσϧͷ
   ׆༻ͱ՝୊
   

   View Slide

9. 9
   ΤοδAIʹ͓͚Δੜ੒Ϟσϧͷ׆༻
   • Ϋϥ΢υΛհ͞ͳ͍ੜ੒Ϟσϧͷֶशɾਪ࿦ͷͨΊͷΞϓϩʔν
   • σόΠε͔ΒͷΞϓϩʔν
   • ߴ͍ॲཧೳྗΛ༗͢ΔσόΠεͷ։ൃ
   ྫʣEdge TPUɼJetson Nano
   • ֶशϞσϧ͔ΒͷΞϓϩʔν
   • αϯϓϦϯά͕ܰྔͳֶशϞσϧɼϞσϧѹॖٕज़ͷ։ൃͳͲɽ
   • ֶशɾਪ࿦ख๏͔ΒͷΞϓϩʔν
   • ޮ཰తͳֶशɾਪ࿦ख๏ͷ։ൃͳͲ
   

   View Slide

10. 10
    ΤοδAIʹ͓͚Δੜ੒Ϟσϧͷ׆༻
    • Ϋϥ΢υΛհ͞ͳ͍ੜ੒Ϟσϧͷֶशɾਪ࿦ͷͨΊͷΞϓϩʔν
    • σόΠε͔ΒͷΞϓϩʔν
    • ߴ͍ॲཧೳྗΛ༗͢ΔσόΠεͷ։ൃ
    ྫʣEdge TPUɼJetson Nano
    • ֶशϞσϧ͔ΒͷΞϓϩʔν
    • αϯϓϦϯά͕ܰྔͳֶशϞσϧɼϞσϧѹॖٕज़ͷ։ൃͳͲɽ
    • ֶशɾਪ࿦ख๏͔ΒͷΞϓϩʔν
    • ޮ཰తͳֶशɾਪ࿦ख๏ͷ։ൃͳͲ
    

    View Slide

11. 11
    ྔࢠΞχʔϦϯάͷԠ༻ઌ
    ྔࢠΞχʔϦϯάΛ༻͍ͨ৽ͨͳΞϓϩʔν
    1ɽ૊߹ͤ࠷దԽ໰୊
    2ɽαϯϓϦϯά
    αϯϓϦϯάͱ͸ɼ֬཰෼෍͔Βඪຊ(αϯϓϧ)Λநग़͢Δ͜ͱɽ
    ֬཰෼෍ͷظ଴஋ܭࢉʹ༻͍ΒΕΔɽ
    D-Wave Systems͕ࣾ։ൃ͢ΔྔࢠΞχʔϦϯάϚγϯʢҎԼɼD-WaveϚγϯʣ͸ɼҎԼͷ
    ੜ੒Ϟσϧͷֶशɾਪ࿦ͷߴ଎Խ΍ߴਫ਼౓Խʹ੒ޭ͍ͯ͠Δɽ
    [1] S. H. Adachi and M. P. Henderson, “Application of Quantum Annealing to Training of Deep Neural Networks”, arXiv:1510.06356 2015.
    [2] D. Korenkevych et al., “Benchmarking Quantum Hardware for Training of Fully Visible Boltzmann Machines”, arXiv:1611.04528 2016.
    [3] W. Vinci et al., “A Path Towards Quantum Advantage in Training Deep Generative Models with Quantum Annealers”, arXiv:1912.02119 2019.
    • ϘϧπϚϯϚγϯ
    • ੍ݶϘϧπϚϯϚγϯ[1]
    • ՄࢹϊʔυͷΈͷϘϧπϚϯϚγϯ[2]
    • ม෼ΦʔτΤϯίʔμʢVAEʣ[3]
    

    View Slide

12. 12
    ΤοδAI΁ͷద༻ͱ՝୊
    [1] http://dwavejapan.com/system/
    [2] D. Korenkevych et al., “Benchmarking Quantum Hardware for Training of Fully Visible Boltzmann Machines”, arXiv:1611.04528 2016.
    [3] Z. Chen et al., “An empirical study of latency in an emerging class of edge computing applications for wearable cognitive assistance”, in Proceedings of the Second ACM/IEEE
    Symposium on Edge Computing, Oct. 2017, pp. 1-14.
    D-Wave 2000Q[1]
    • ྔࢠΞχʔϦϯάΛ׆༻ͨ͠ੜ੒Ϟσϧͷֶशɾਪ࿦ͷ࢓૊ΈΛΤοδAI
    ʹద༻͢Δ৔߹ɼD-WaveϚγϯ͸අ༻΍ۃ௿Թͷಈ࡞؀ڥΛཁ͢ΔͳͲͷ
    ཧ༝͔ΒΤοδྖҬ΁ͷ഑ஔ͸ࠔ೉Ͱ͋Δɽ
    • D-WaveϚγϯΛར༻͢Δࡍ͸ɼωοτϫʔΫΛܦ༝ͨ͠Ϋϥ΢υαʔϏε
    ͱͯ͠ར༻͢Δɽ
    • Ϋϥ΢υΛհͣ͞ΤοδଆͰֶशΛߦ͏͜ͱ͕Ͱ͖ͳ͍ͨΊɼσόΠεͱΫϥ΢υؒͷ෺ཧత
    ͳڑ཭ʹىҼͨ͠௨৴஗Ԇ͕ൃੜ͢Δɽ
    • D-WaveϚγϯͷαϯϓϦϯάʹ͔͔Δ͕࣌ؒ֓ͶμsΦʔμʔ[2]Ͱ͋Δ͜ͱɼ͓ΑͼΫϥ΢υ
    ͱͷ௨৴஗Ԇ͕֓Ͷ਺ඦmsΦʔμʔ[3]Ͱ͋Δ͜ͱΛߟ͑Δͱɼੜ੒Ϟσϧͷֶश΍ਪ࿦ʹ͓͍
    ͯ௨৴஗Ԇ͕ϘτϧωοΫʹͳΔɽ
    

    View Slide

13. 13
    D-WaveϚγϯͱΞχʔϦϯάϚγϯ
    ΠδϯάϞσϧ
    [1] M. Yamaoka et al., “CMOS Annealing Machine: an In-memory Computing Accelerator to Process Combinatorial Optimization Problems”, in IEEE Custom Integrated Circuits
    Conference (CICC), Apr. 2019, pp. 1-8.
    [2] M. Aramon et al., “Physics-Inspired Optimization for Quadratic Unconstrained Problems Using a Digital Annealer”, Front. Phys., vol. 7, no. 48, Apr. 2019.
    [3] E. Crosson and A. W. Harrow, “Simulated Quantum Annealing Can Be Exponentially Faster Than Classical Simulated Annealing”, in IEEE 57th Annual Symposium on
    Foundations of Computer Science (FOCS), Oct. 2016, pp. 714-723.
    • D-WaveϚγϯͷΑ͏ͳྔࢠϏοτͷৼΔ෣͍Λσδλϧճ࿏্ͰγϛϡϨʔτ͢Δ΋ͷͱͯ͠ɼ
    ΞχʔϦϯάϚγϯ͕͋Δɽ
    • ΞχʔϦϯάϚγϯ͸͜Ε·ͰʹɼݱߦͷD-WaveϚγϯͰ͸ղ͚ͳ͍໰୊αΠζΛ࣋ͭ૊߹ͤ
    ࠷దԽ໰୊ͷղ๏ͱͯ͠ɼͦͷ༗ޮੑ͕ࣔ͞Ε͍ͯΔɽ
    • ΞχʔϦϯάϚγϯ͸ɼD-WaveϚγϯͱൺ΂ͯখܕ͔ͭ௿ίετͰ͋ΓɼৗԹͰಈ࡞͢Δ
    ͱ͍͏ಛ௃Λ༗͍ͯ͠Δɽ
    ※ ຊൃදʹ͓͚ΔΞχʔϦϯάϚγϯͷఆٛ͸ɼΠδϯάϞσϧͷجఈঢ়ଶ୳ࡧΛ໨త
    ɹͱͨ͠΋ͷͰ͋ΓɼFPGA΍ASICΛ༻͍ͨϋʔυ΢ΣΞ࣮૷[1,2]ʹՃ͑ɼGPU΍CPU
    ɹ্Ͱಈ࡞͢Διϑτ΢ΣΞ࣮૷[3]΋ؚΉ
    

    View Slide

14. 3.
    ఏҊ͢ΔΞʔΩςΫνϟ
    

    View Slide

15. 15
    ΞʔΩςΫνϟ֓ཁ
    ੜ੒Ϟσϧͷֶशɾਪ࿦ʹඞཁͳαϯϓϦϯάॲཧͷΞΫηϥϨʔλͱͯ͠ɼ
    ΞχʔϦϯάϚγϯΛσόΠεͷۙ๣ʹ഑ஔ͢ΔΞʔΩςΫνϟ
    Edge server
    Cloud
    WAN
    Edge
    End devices
    Annealing
    machine
    AI
    AI
    AI
    AI
    AI
    ΤοδྖҬʹ഑ஔͨ͠ΞχʔϦϯάϚγϯ͸ɼαϯϓϦϯάॲཧͷΞΫηϥϨʔλͱͯ͠
    ෳ਺୆ͷσόΠεʹڞ༗͞ΕΔɽ
    

    View Slide

16. 16
    ΞχʔϦϯάϚγϯΛ׆༻ֶͨ͠शɾਪ࿦
    ΞχʔϦϯάϚγϯ͸ɼྔࢠྗֶతͳΏΒ͗Λར༻͠ͳ͍ͳͲಈ࡞ݪཧ͸D-WaveϚγϯͱҟ
    ͳΔ͕ɼର৅ͱ͢Δ໰୊(ΠδϯάϞσϧͷجఈঢ়ଶ୳ࡧ)͸D-WaveϚγϯͱಉ༷Ͱ͋Δɽ
    ͦͷͨΊɼΞχʔϦϯάϚγϯΛੜ੒Ϟσϧͷֶशɾਪ࿦ʹ׆༻Ͱ͖ͳ͍͔ɽ
    ධՁ݁ՌΛޙड़͢Δ
    ظ଴Ͱ͖Δ͜ͱ
    • αϯϓϦϯάॲཧͷߴ଎Խ
    ΠδϯάϞσϧͷجఈঢ়ଶ୳ࡧʹର͢Δߴ଎ੑΛαϯϓϦϯάʹ׆༻
    • ֶशਫ਼౓ͷ޲্
    D-WaveϚγϯͰࣔ͞ΕͨΑ͏ͳੜ੒Ϟσϧͷֶशਫ਼౓޲্
    

    View Slide

17. 17
    σόΠεͷԠ౴ੑͱ௨৴஗Ԇ
    • σόΠεɾΞχʔϦϯάϚγϯؒͷ௨৴஗ԆΛ਺msΦʔμʔʹ཈͑Δɽ
    • σόΠεͷԠ౴࣌ؒʹର͢Δ௨৴஗ԆͷӨڹ͕খ͘͞ɼΞχʔϦϯάϚγϯͷαϯϓϦϯ
    ά࣌ؒΛߴ଎ԽͰ͖Ε͹ɼͦΕʹ൐͍σόΠεͷԠ౴ੑΛ޲্Ͱ͖Δɽ
    ΞχʔϦϯάϚγϯ͸ɼD-WaveϚγϯͱൺ΂ͯখܕ͔ͭ௿ίετͰ͋ΓɼৗԹͰಈ࡞͢ΔͳͲ
    ಛघͳಈ࡞؀ڥΛཁ͠ͳ͍ͨΊɼσόΠεʹ͍ۙΤοδྖҬʹ഑ஔ͢Δ͜ͱʹద͍ͯ͠Δɽ
    ͦͷͨΊɼΞχʔϦϯάϚγϯΛσόΠεۙ๣ʹ഑ஔ͢ΔΞʔΩςΫνϟΛͱΔɽ
    ධՁ݁ՌΛޙड़͢Δ
    ظ଴Ͱ͖Δ͜ͱ
    

    View Slide

18. 4.
    ධՁͱߟ࡯
    

    View Slide

19. 19
    ఏҊΞʔΩςΫνϟͷධՁ
    ҎԼͷ2఺ʹ͓͍ͯɼఏҊΞʔΩςΫνϟͷγεςϜͱͯ͠ͷར఺ΛධՁ͢Δɽ
    ̍ɽΞχʔϦϯάϚγϯΛ༻͍ͯੜ੒ϞσϧͷҰͭͰ͋ΔϘϧπϚϯϚγϯΛޮ཰త
    ɹɹʹֶशͰ͖Δ͔ʹ͍ͭͯͷධՁɽ
    ɹɹΞχʔϦϯάϚγϯΛʮΞΫηϥϨʔλͱͯ͠ར༻͢Δʯ͜ͱͷར఺
    ̎ɽΞχʔϦϯάϚγϯΛ௨৴஗Ԇ͕খ͍͞σόΠεۙ๣ʹ഑ஔ͢Δ͜ͱͰਪ࿦࣌ͷ
    ɹɹσόΠεͷԠ౴ੑΛ޲্Ͱ͖Δ͔ʹ͍ͭͯධՁ
    ɹɹΞχʔϦϯάϚγϯΛʮσόΠεͷۙ๣ʹ഑ஔ͢Δʯ͜ͱͷར఺
    

    View Slide

20. ࣮ݧ؀ڥ
    ֤ϩʔϧͷઆ໌
    • σόΠε
    • ੜ੒Ϟσϧͷֶशɾਪ࿦Λߦ͏໾ׂɽ
    ֶशɾਪ࿦ͰඞཁͳαϯϓϦϯάΛΞχʔϦϯάϚγϯʹཁٻ͢ΔΫϥΠΞϯτͱͯ͠ಈ࡞
    • ੜ੒Ϟσϧͱͯ͠ɼ੍ݶϘϧπϚϯϚγϯ(RBM)Λ༻͍Δɽ
    • ֶशσʔλͱͯ͠MNISTσʔληοτΛ༻͍ͯɼRBMͰը૾ͷੜ੒աఔΛֶश͢Δɽ
    • ΞχʔϦϯάϚγϯ
    • σόΠε͔ΒͷαϯϓϦϯάཁٻΛड͚ɼܭࢉ݁ՌΛฦ͢໾ׂ
    • αϯϓϦϯάͷॲཧʹ͸ɼSimulated Quantum Annealing(SQA)Λ༻͍Δɽ
    • SQAΛ࣮૷ͨ͠Sqaod[1]Λ༻͍ͯWeb APIΛߏஙͨ͠ɽ
    v
    1
    v
    2
    v
    3
    v
    n
    h
    1
    h
    2
    h
    m
    RBMͷάϥϑߏ଄
    ӅΕ૚
    m = 100
    Մࢹ૚
    n = 784
    ΞχʔϦϯάϚγϯ
    σόΠε
    CPU1ίΞ
    ϝϞϦ1GB
    CPU8ίΞ
    ϝϞϦ32GB
    αϯϓϦϯάཁٻ
    ܭࢉ݁ՌΛฦ͢
    20
    Web API
    ֶशσʔλ
    [1] https://github.com/shinmorino/sqaod
    

    View Slide

21. 1. ΞχʔϦϯάϚγϯΛ༻ֶ͍ͨश
    RBMΛҎԼͷೋͭͷֶशख๏Ͱֶश͠ɼൺֱ͢Δɽ
    ɹطଘͷֶशख๏ɿίϯτϥεςΟϒɾμΠόʔδΣϯε(CD)๏[1]
    ɹఏҊख๏ɿΞχʔϦϯάϚγϯ (ΞϧΰϦζϜ͸SQA)
    ֶशํ๏
    αϯϓϦϯάͱύϥϝʔλߋ৽͕Ұճ࣮ͣͭߦ
    21
    1ΤϙοΫ=600ΠςϨʔγϣϯ
    CD๏͸ɼશͯͷॲཧΛσόΠε্Ͱ࣮ߦ͢Δɽ
    ҰํɼSQAͰ͸αϯϓϦϯάͷॲཧͷΈΛWeb APIܦ༝ͯ͠ΞχʔϦϯάϚγϯͰ࣮ߦ͢Δɽ
    MNISTͷ60000ຕͷ܇࿅ը૾σʔλΛόοναΠζ100ͰϛχόονֶशΛߦ͏
    ɾ
    ɾ
    ɾ ɾ
    ɾ
    ɾ ɾ
    ɾ
    ɾ ɾ
    ɾ
    ɾ
    1ΠςϨʔγϣϯ
    [1] G. Hinton, “Training Products of Experts by Minimizing Contrastive Divergence”, Neural computation, vol. 14, no. 8, pp. 1771-1800, Aug. 2002.
    

    View Slide

22. 22
    1. ΞχʔϦϯάϚγϯΛ༻ֶ͍ͨश
    • ޿ൣͳ໰୊ઃఆʹରͯ͠ಉ༷ͷ݁Ռ͕ಘΒΕΔ͔ɽ
    • ݫີʹ͸࠷খ஋୳ࡧͱظ଴஋ܭࢉ͸ҟͳΔॲཧͰ͋ΔͨΊɼ࠷খ஋୳ࡧͷฏۉ஋Λظ଴஋
    ͱͯۙ͠ࣅ͢Δ͜ͱʹΑΔӨڹ͸Ͳ͏͔ɼetc.
    SQA͸CD๏ʹൺ΂ͯɼগͳ͍ΤϙοΫ਺Ͱ௿͍࠶ߏ੒ޡࠩʹ
    ౸ୡ͓ͯ͠Γɼऩଋޙͷ࠶ߏ੒ޡࠩ΋௿͍஋Λ͍ࣔͯ͠Δɽ
    ࠓճͷ৚݅ʹ͓͍ͯɼSQA͸CD๏ΑΓ΋RBMΛޮ཰తʹֶश
    Ͱ͖͍ͯΔͱݴ͑Δɽ
    ֶशޙͷϞσϧ͕σʔλͷੜ੒աఔΛͲΕ͚ͩଊ͑ͯ
    ͍Δ͔ͷࢦඪͱͳΓɼ௿͍஋΄Ͳྑ͍Ϟσϧͱݴ͑Δ
    ධՁࢦඪ
    ߋͳΔݕূ͕ඞཁ
    ΤϙοΫ͋ͨΓͷฏۉֶश࣌ؒ
    CD๏ɿ5.21 sɼSQAɿ282 s
    ΤϙοΫ਺ʹର͢Δ࠶ߏ੒ޡࠩ
    Reconstruction Error =
    ∑
    i
    (image(i) ≠ reconstructedImage(i))
    total pixels
    

    View Slide

23. 23
    2. ௨৴஗ԆͷӨڹ
    ҎԼͷೋͭͷ؀ڥͰRBMͷਪ࿦ʹཁ͢Δ࣌ؒΛܭଌ͠ɼൺֱ͢Δɽ
    ๺ւಓੴङࢢͷDC
    ΞϝϦΧ߹ऺࠃ
    όʔδχΞभͷDC
    ΞχʔϦϯάϚγϯ
    σόΠε
    σόΠε
    ᶃ RTT: 0.612 ms
    ᶄ RTT: 176 ms ᶃ ఏҊख๏
    σόΠεͱΞχʔϦϯάϚγϯΛಉDC಺ʹ഑ஔ͢Δ
    ᶄ Ϋϥ΢υ؀ڥ
    σόΠεͱΞχʔϦϯάϚγϯΛ෺ཧతʹڑ཭͕
    ɹ཭ΕͨDC಺ʹͦΕͧΕ഑ஔ͢Δɽ
    ਪ࿦ͷॲཧͷྲྀΕ
    σόΠε ΞχʔϦϯάϚγϯ
    ਪ࿦࣌ؒʹ͸ɼ
    ɾσόΠεͱΞχʔϦϯάϚγϯؒͷ௨৴஗Ԇ
    ɾΞχʔϦϯάϚγϯ্ͰͷαϯϓϦϯά࣌ؒ
    ɾWeb APIͷԠ౴࣌ؒ
    ͳͲؚ͕·Ε͍ͯΔɽ
    ᶃαϯϓϦϯάཁٻ
    ᶅܭࢉ݁ՌΛฦ͢
    ᶄαϯϓϦϯά࣮ߦ
    ※ RTT: pingίϚϯυΛ༻͍ͯଌఆͨ͠ϥ΢ϯυτϦοϓλΠϜ
    Web API
    

    View Slide

24. 24
    2. ௨৴஗ԆͷӨڹ
    RBMͷਪ࿦࣌ؒͷܭଌ݁Ռ
    • ਪ࿦࣌ؒ͸ɼΫϥ΢υ؀ڥ͕ఏҊख๏ʹൺ΂ͯ350 ms
    ஗͍݁Ռͱͳͬͨɽ
    • ͜ͷࠩ෼͸ɼΫϥ΢υ؀ڥ͕ఏҊख๏ʹൺ΂ͯσόΠε
    ͱΞχʔϦϯάϚγϯؒͷ෺ཧతͳڑ཭͕ԕ͘ͳΔ͜ͱ
    Ͱੜ͡Δ௨৴஗ԆʹىҼ͍ͯ͠Δͱߟ͑ΒΕΔɽ
    ࣮ݧ݁Ռ
    ※ਪ࿦࣌ؒʹ͸αϯϓϦϯάؚ͕࣌ؒ·ΕΔ
    

    View Slide

25. 25
    2. ௨৴஗ԆͷӨڹ
    RBMͷਪ࿦࣌ؒͷܭଌ݁Ռ
    • ਪ࿦࣌ؒ͸ɼΫϥ΢υ؀ڥ͕ఏҊख๏ʹൺ΂ͯ350 ms
    ஗͍݁Ռͱͳͬͨɽ
    • ͜ͷࠩ෼͸ɼΫϥ΢υ؀ڥ͕ఏҊख๏ʹൺ΂ͯσόΠε
    ͱΞχʔϦϯάϚγϯؒͷ෺ཧతͳڑ཭͕ԕ͘ͳΔ͜ͱ
    Ͱੜ͡Δ௨৴஗ԆʹىҼ͍ͯ͠Δͱߟ͑ΒΕΔɽ
    • D-WaveϚγϯΛ༻͍ͨ৔߹ɼαϯϓϦϯά࣌ؒ͸ΞχʔϦϯάϚγϯΑΓߴ଎Ͱ͋Δ͜ͱ͕༧૝͞Ε
    Δ͕ɼྫ͑αϯϓϦϯά͕࣌ؒμsΦʔμʔͰ͋ͬͯ΋ɼ௨৴஗Ԇ͕ϘτϧωοΫͱͳΓɼਪ࿦࣌ؒ͸ఏ
    Ҋख๏ΑΓ΋஗͍400 msఔ౓Ͱ͋Δ͜ͱ͕༧૝͞ΕΔɽ
    • ҰํɼఏҊख๏Ͱ͸ɼਪ࿦࣌ؒʹରͯ͠αϯϓϦϯά͕࣌ؒࢧ഑తͰ͋ΔͨΊɼΑΓαϯϓϦϯά࣌ؒ
    ͕୹͍ΞχʔϦϯάϚγϯΛ༻͍Δ͜ͱͰਪ࿦࣌ؒΛେ෯ʹ୹ॖͰ͖Δɽ
    ྫ͑͹ɼαϯϓϦϯά࣌ؒΛ30 msఔ౓·Ͱߴ଎ԽͰ͖Ε͹ɼਪ࿦͕࣌ؒ100 msΛԼճΔɽ
    Ϋϥ΢υܦ༝ͰD-WaveϚγϯΛ༻͍Δ৔߹ͱͷൺֱ
    ࣮ݧ݁Ռ
    ※ਪ࿦࣌ؒʹ͸αϯϓϦϯάؚ͕࣌ؒ·ΕΔ
    

    View Slide

26. 5.
    ·ͱΊ
    

    View Slide

27. 27
    ·ͱΊ
    • ຊݚڀͰ͸ɼੜ੒Ϟσϧͷֶशɾਪ࿦ʹඞཁͳαϯϓϦϯάॲཧͷΞΫηϥϨʔλͱͯ͠ɼ
    ΞχʔϦϯάϚγϯΛσόΠεͷۙ๣ʹ഑ஔ͢Δ͜ͱͰɼΤοδAIʹ͓͍ͯੜ੒Ϟσϧͷֶश
    ΛΫϥ΢υΛհͣ͞ʹޮ཰Խ͠ɼ͔ͭਪ࿦࣌ͷσόΠεͷԠ౴ੑΛ޲্ͤ͞ΔΞʔΩςΫ
    νϟΛఏҊͨ͠ɽ
    • ఏҊΞʔΩςΫνϟʹΑΓɼσόΠε͸൚༻త͔ͭ௿εϖοΫͰ͋Γͳ͕Βɼܭࢉίετ͕
    ߴ͍αϯϓϦϯάͷॲཧΛΞχʔϦϯάϚγϯ͕୲͏͜ͱͰɼߴ͍Ԡ༻ੑΛ࣋ͭੜ੒Ϟσϧ
    ΛσόΠε্Ͱ׆༻͢Δ͜ͱΛՄೳʹ͢Δɽ
    • ࣮ݧ͔ΒɼΞχʔϦϯάϚγϯʢSQAʣ༻͍ͯੜ੒ϞσϧͷҰͭͰ͋ΔRBMΛޮ཰తʹֶश
    Ͱ͖Δ͜ͱɼ͓ΑͼΞχʔϦϯάϚγϯΛσόΠεͷۙ๣ʹ഑ஔ͢Δ͜ͱͰɼਪ࿦࣌ؒͷߴ
    ଎Խ͕ظ଴Ͱ͖Δ͜ͱΛࣔͨ͠ɽ
    

    View Slide