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

AlphaGo에서 시작하는 인공지능

AlphaGo에서 시작하는 인공지능

AlphaGo, Alpha Zero에서 시작해서 여러 인공지능의 개념을 살펴봅니다.

Leonardo YongUk Kim

December 11, 2021
Tweet

More Decks by Leonardo YongUk Kim

Other Decks in Technology

Transcript

  1. ALPHAGOোഄ AlphaG o Fan Goח ੌࠄয۽ ߄قਸ ੄޷ 2015 2016.

    3. 9 AlphaG o Lee 2017. 5. 17 AlphaG o M aster AlphaG o Zero 2017. 10. 19 2018.12.7 Alpha Zero ౸ റ੉ (2ױ)җ ؀Ѿ೧ࢲ थܻ ੉ࣁج (9ױ)җ ؀Ѿ೧ࢲ थܻ 
 ੷੘ӂ ೧Ѿػ 16݅ ӝࠁ ೟ण. 48ѐ੄ TPU ৡۄੋ ؀Ҵ 60োथ ழઁীѱ थܻ 4ѐ੄ TPU. 10ߓ ীց૑ ബਯ ੋр੄ ૑धহ੉ (ӝࠁ X) ߄ق ੿ࠂ ঌ౵Ҋ ݃झఠ৬ ऱਕ 89थ 11ಁ ࣳӝ, ୓झ, झఋ 2 ੿ࠂ Goܳ ੉ܴীࢲ ઁ৻
  2. TIC TAC TOE O O O X X O X

    O X X O X O - ݢ੷ 3ѐܳ աۆ൤ ֬਷ ࢎۈ੉ ੉ӝח ѱ੐ - ੉ ੿ب ࠂ੟ೠ ѱ੐਷ ݽٚ ҃਋੄ ࣻܳ ࠅ ࣻ ੓਺. - 9 x 8 x 7 x 6 x 5 x 4 x 3 x 2 = 9! O X O O X X O
  3. O O O O O O O O O ROOT

    O X O X O X O X O X O X O X O X O X O O X O O X O O X O O X O O X O O X O ֢٘ 1 0 -1 0 -1 -1 ઺ۚ ܖ౟ ੉ӝݶ 1੼, ࠺ӝݶ 0੼,૑ݶ -1੼ ROOT੄ ੑ੢ীࢲ 
 ੗ध੉ۄҊب ೤פ׮. ݽٚ ֢٘ח ੗ध੉੗ ࠗݽੑפ׮.
  4. O O O O O O O O O ROOT

    O X O X O X O X O X O X O X O X O X O O X O O X O O X O O X O O X O O X O ֢٘ 1 0 -1 0 -1 -1 ઺ۚ ܖ౟ MY TURN ௾ ੼ ࣻܳ ࢶఖೞҊ ੼ࣻо ੘਷ ࠳ے஖ܳ ઁѢ
  5. O O O O O O O O O ROOT

    O X O X O X O X O X O X O X O X O X O O X O O X O O X O O X O O X O O X O ֢٘ ઺ۚ ܖ౟ YOUR TURN ੘਷ ੼ࣻܳ ࢶఖ 
 ௾ ੼ࣻо ੓ח ࠳ے஖ח ઁѢ
  6. O O O O O O O O O ROOT

    O X O X O X O X O X O X O X O X O X O O X O O X O O X O O X O O X O O X O ઺ۚ MY TURN ௾ ੼ࣻܳ ࢶఖ
  7. - (19 x 19)! - 19 x 19 = 361

    - 361 x 360 x 359 x 358 …. x 2 - 26744876149564427899473201526425013452390919904351815721084971068304474 7437531294143149639831010372677443849403182318969228741381559487197927737 64930851408087543453474101182344879484162985721534603948370802204778391 45379274006646833128661312942336287321284636912937632439789397222224742 52826712518506072707918591157844247991603554375217925635775598044364577 67819229829195896785070533331329604395837235880245012197523337773352603 746540435758711323413067205097510873318696774954051195138779582025728045 717997197429383169516478847881483048003766654327470766455887103023601081 7570107837589904730596477443151082000948524919032642496288011069869044 42993333787797164945029657423253487692054233010201128993815319944149127 636942433924747935483481769568213401600000000000000000000000000000000 0000000000000000000000000000000000000000000000000000000000000000 ߄ق੄ ҃਋੄ ࣻ
  8. MONTE CARLO METHOD - ےؒೞѱ ੼ਸ ନ਻द׮. - ਗ উ੄

    ੼੄ іࣻо 314ѐ, ࢎпഋ উ੄ ੼੄ іࣻо 400ѐۄ о੿೧ ࠇद׮. - 4(R ^ 2) : π(R ^ 2) = 400 : 314, π = 3.14 - ୽࠙൤ ੼ਸ ݆੉ ନਸ ࣻ۾ ؊ ੿޻ೠ ਗ઱ਯਸ ঌѱ ؾפ׮.
  9. O O O O O O O O O ROOT

    ୐ ࣻ ੉റ ےؒਵ۽ 10౸ فয пп थܫਸ ҳೣ. 40% PLAYOUT 60% 30% 20% 80% 30% 20% 50% 10% - п ࣻ ݃׮ ੉റ, ےؒೞѱ 10౸ਸ فয (Playout) थܫਸ ҳೣ. (9 x 10 = 90౸) - थܫ੉ о੢ ֫਷ 5ߣ૩ ࣻ(80%)о ୭Ҋ৓׮Ҋ ౸ױೣ. - दр੉ ؊ ݆੉ ੓׮ݶ Playout੄ പࣻܳ ט۰ࢲ न܉بܳ ૐоदఆ ࣻ ੓਺. (100౸ فӝ) - Playout੄ പࣻо טযաݶ טযզ ࣻ۾ MiniMax੄ Ѿҗ৬ ਬࢎ೧૕ Ѫ੐. - Playout, Rollout, Simulation ١੄ ױযо ࢎਊ ؽ.
  10. O O O O O O O O O ROOT

    40% PLAYOUT 60% 30% 20% 80% 30% 20% 50% 10% - ےؒೞѱ فযࢲ थܫ੉ ֫਷ ࣻܳ ݢ੷ Ҋܵפ׮.
  11. O O O O O O O O O ROOT

    - ੉ઁ थܫ੉ ֫਷ 5ߣ૩ ࣻ ਤ઱۽ ےؒ ؀Ѿਸ ೤פ׮. (70%) - աݠ૑ ٜࣻب ےؒ ؀Ѿਸ оՔ೤פ׮. (30%) - ֬஘ ࣻо ੓ӝ ٸޙੑפ׮.
  12. O O O O O O O O O ROOT

    - 5ߣ૩ ࣻী ؀೧ যו ੿ب ےؒ ؀Ѿਸ ೮ਵݶ, 5ߣ૩ ࣻ੄ ੗धب ౟ܻী ನೣ೤פ׮. - ੉ઁ ےؒೞѱ ف૑݅ ୊਺ ف ࣻח Ҋ੿ೞҊ ೤פ׮. - ੼ର ৌ۰ ੓ח ֢٘о ݆ই ૘פ׮. X O X O X O O X X O O X O X O X
  13. - ୓झח 1996֙ 2ਘ 10ੌ MiniMax ۽ ੿ࠂ. (गಌ ஹೊఠ

    ٩ ࠶ܖ) - ୭Ӕীח Monte Carlo Tree Search۽ ോ؀ಪਵ۽ب Ӓے٘݃झ ఠܳ ੉ӣ. - ೞ૑݅ ߄ق਷ ࠛоמ೮਺. - ഛܫ੸ਵ۽ ೞӝূ ڜࣻо ցޖ ݆ ਺. ࣻ ੍ӝ݅ਵ۽ ୽ ࠙ೠоਃ?
  14. X1 X2 Y W1 W2 Y = W1X1 + W2X2

    ࣻ ݆਷ ׏۠੉ Ѿ೤೧ (֎౟ਕ௼) ૑מਸ ٜ݅যմ Ѫ ୊ۢ, ࣻ рױೠ о઺஖ ো࢑(ੋҕ ׏۠)ਸ ֎౟ਕ௼ܳ ٜ݅ݶ যڌѱ ؼөਃ?
  15. - ೙ӝ୓ 0~9ܳ ޷Ҵ NISTо ࣻ૘. - о۽ 28, ࣁ۽

    28੄ ೗ࣄ۽ ҳࢿ. - ೠ Ӗ੗׼ 784 ೗ࣄ (28 x 28) MNIST
  16. - ࡈр࢝਷ ੼੉ ੓ਵݶ উػ׮ח ڷ. - ౵ۆ࢝਷ ੼੉ ੓ਵݶ

    જ׮ח ڷ. - 0੄ о਍ؘח ੼੉ ੓ਵݶ উػ׮. - 1੄ о਍ؘח ੼੉ ੓ਸ оמࢿ੉ ֫ ਺. - ࡈъ਷ -1, ౵ی਷ 1۽ ࠁ੗. Ѩ਷ ࢝ ਷ 0 ೙ӝ୓ ੿ࠂೞ੗.
  17. X1 X784 W1 W784 W1X1 + … + W784X784 =

    Y Yо ௼ݶ ೙ӝ୓о 5ۄח ڷ.
  18. - W1X1 + … + W784X784 = Y - X1ࠗఠ

    X784ө૑ ੓֎ਃ? ੑ۱੉ 784ѐ - W1ࠗఠ W784ө૑ ੓֎ਃ? о઺஖о 784ѐ - ইۄ࠺ই ं੗о 0ࠗఠ 9ө૑ਗ਼ইਃ? о઺஖о 7840ѐ (784 x 10) न҃ݎ
  19. X1 X2 Y1 W1 W2 X784 … Y2 Y10 …

    784ѐ ੑ۱ 0ੋоਃ? 1ੋоਃ? 9ੋоਃ? ୐ߣ૩ ೗ࣄ W10 W7840 7840ѐ о઺஖ 10ѐ ୹۱
  20. - ӝ҅о ҕࠗೞӝ ٸޙী ӝ҅ ೟ण - ӝ҅о ҕࠗೡ ࣻ

    ੓ѱ ޙઁ৬ ੿׹ਸ ળ࠺೤פ׮. - MNISTۄח ೙ӝ୓ ࣁ౟ب ੉޷૑৬ ੿׹੉ э੉ ઓ੤೤פ׮. - ف ઙܨ੄ ؘ੉ఠܳ ৮੹൤ ܻ࠙ೡ ӝ਎ӝ৬ ੺ಞਸ 
 ӝ҅о ଺ইմ׮Ҋ ࠁݶ ؾפ׮. - ೟ण ؘ੉ఠܳ о૑Ҋ ࣻ হ੉ ޷࠙ਸ ߈ࠂ ೞݴ о઺஖ܳ ઑӘঀ 
 Ҋ୛оݴ ৢ߄ܲ ݽ؛ਸ ٜ݅যцפ׮. - ੿׹ਸ ઱Ҋ ೟णਸ दఃח Ѫਸ ૑ب ೟ण੉ۄҊ ೤פ׮. ӝ҅೟ण
  21. - Ӓې೗ਸ ਤ೧ GPUח ࣻ হ੉ Ӓܿ੗৬ ࡄ੉ ై ৔ؼ

    ੉޷૑ܳ ҅࢑ೞҊ ژ ҅࢑೤פ׮. - GPU੄ ఌਗೡ ҅࢑ מ۱਷ ݠन۞׬җ ঐഐ ചತ ী ੸೤೤פ׮. - ߈ݶ GPUח ਬোೠ ౸ױਸ ޅ೤פ׮. - 1950֙ ࠗఠ োҳػ AIо ੜ উغ঻؍ ੉ਬ ઺ ೞ ա۽ ো࢑ מ۱ ࠗ઒੉ ੉ঠӝ ؽ. - 1970֙ AI ѹ਎ - 1980֙ AI ѹ਎ - 2012֙ীঠ AIо ࡄਸ ࠆ. ݠन۞׬ب GPUо ೤פ׮.
  22. - ೦࢚ ৘৻о ੓ਸ ࣻ ੓णפ׮. (Ҋন੉৬ ъই૑ ࠙ܨب ೐۽Ӓې߁੉

    য۰਑) - ੋр੄ ঱যח ݽഐೞҊ ਋ܻо ਗೞח Ѿҗܳ ঳ӝ য۰਎ ࣻ ੓णפ׮. - ৘৻о ࢤѹب ৘৻੸ੋ ؘ੉ఠب ӝ҅ ೟णਸ दఃݶ ೧Ѿؾפ׮. - ݠन ۞׬਷ ѾҴ ਋ܻо ૒ҙਵ۽ ׮ܖ؍ ࠗ࠙ਸ ೧Ѿ೧ સפ׮. ૒ҙ਷ ঌҊ્ܻਵ۽ ಽӝ য۵णפ׮
  23. - য়ܲଃ੄ ޙઁח ࢶਸ Ӓযࢲ ޙઁܳ ೧Ѿೡ ࣻ হ਺. -

    ࠂ੟ೠ ޙઁח ׮க੄ ۨ੉যо ೙ਃ. ੑ۱கҗ ୹۱கਵ۽ח উغח ޙઁ
  24. X1 X2 X784 … … ੑ۱க Y1 Y2 Y10 …

    ୹۱க ਷ץக - ࠂ੟ೠ ޙઁח 1ѐ ੉࢚੄ ਷ץக(Hidden Layer)ਸ ٟ݅פ׮. - 2ѐ ੉࢚੄ ਷ץகਸ ೟णೞח ҃਋ Deep Learning੉ۄҊ ೞݴ बக न҃ݎ (Deep Neural Network)ۄח ݺடਸ ࢎਊ೤פ׮. - Microsoftח 152கਸ о૓ ResNetਸ ٜ݅׮ ੉റ 1001ѐ੄ கө૑ ٜ݅঻णפ׮.
  25. CNN

  26. Ҋন੉ VS ѐ - बகݎਵ۽ח ੉޷૑ ࠙ܨ ޙઁܳ ੿ࠂೞӝ য۰ਛणפ׮.

    - CNN (Convolutional Neural Network)੄ AlexNet੉ աয়ݶࢲ ੉޷૑ ޙઁܳ ೧Ѿ. - 2012֙ 9ਘ 30ੌࠗఠ ஹೊఠо ૒ҙ੄ ৔৉ਸ ֈযࢲӝ द੘೮णפ׮.
  27. 0 1 2 3 4 5 6 7 8 KERNEL੉

    ೨ब ੑ۱ 0 1 2 3 ழօ 19 25 37 43 X = - 0 x 0 + 1 x 1 + 3 x 2 + 4 + 3 = 19 - “ழօ”਷ “੉޷૑ ೙ఠ”ۄҊ ࢤпೞݶ ؾפ׮. ੉޷૑ ೙ఠ۽ ؘ੉ఠܳ оҕೞח Ѫ. - ೙ఠ݃׮ ౠࢿਸ ъചೠ׮Ҋ ࢤпೞҊ ੓णפ׮. - যڃ ೙ఠח ਮҘਸ ъച? - যڃ ೙ఠח ӈա ௏ܳ ؊ ੜ ࠁ੉ѱ ъച? э਷ ࢚࢝ՙܻ ғೞҊ ؊೤פ׮.
  28. X1 X2 X784 … … ੑ۱க Y1 Y2 Y10 …

    ୹۱க ਷ץக ழօ 3ѐ ೙ఠ
  29. ъച೟ण - ഒ੗ ೒ۨ੉ (Self Play)ܳ ೞݴ प۱ਸ ט۰х. -

    (࢚క, ঘ࣌, थಁ ৈࠗ)ܳ о૑Ҋ ೟ण, ੉റ ࢚కী ؀೧ ࣻ೯೧ঠ ೡ ঘ࣌ਸ ঌ۰ષ. - ঌ౵Ҋח झझ۽ ؊ ъ೧૗. (ࢎ੹ ೟णೠ ӝࠁח ੷੘ӂ ೧Ѿػ 16݅ ӝࠁ. ؀ࠗ࠙ അ؀ ೐۽о ইש.) - CNN (੿଼ݎ, о஖ݎ) + ހప ஠ܳ۽ ౟ܻ ࢲ஖ + ъച ೟ण - ঌ౵ ઁ۽ח ӝࠁ ೟णب ೞ૑ ঋҊ ъച ೟ण݅ਵ۽ ֎౟ਕ௼ܳ ҳ୷. X1 X2 X784 … … ࢚క Y1 Y2 Y10 … ঘ࣌ ਷ץக ழօ
  30. MINIMAX۽ ೧ب ؾפ׮. ъച೟ण + न҃ݎ + ހప஠ܳ۽ ౟ܻ ࢲ஖۽

    ٜ݅঻णפ׮. 
 ALPHAZEROی э਷ ߑधਵ۽ਃ.
  31. ੋҕ૑מ੄ ࢎਊ ৘ ܻ࠭ ಣо ߣ৉ӝ ੸؀੸ ࢤࢿ न҃ݎ ੗ਯ

    ઱೯ ର۝ GPT-3 न࠙ૐ ੋध ୁࠈ नਊಣоݽ؛
  32. - “ঘ࣌੉ ഴܯೞҊ ࠺઱঴੉ ഴܯ೮णפ׮.” -> ଵ (୶ୌ) - “ցޖ

    ੘ਤ੸੉Ҋ ઴Ѣܻо ࡞೮णפ׮.” -> Ѣ૙ (࠺୶) - CNNਵ۽ ೟णदெ ࢜۽਍ ܻ࠭ী ؀೧ ୶ୌੋ૑ ࠺୶ੋ૑ ഛੋ. ܻ࠭ ಣо
  33. - RNNਸ ࢎਊ. CNNҗ੄ ର੉੼਷ ਷ץக੄ ؘ੉ఠܳ ׮द ਷ץகਵ۽ (ӝর۱)

    - ҳӖ੉ न҃ݎ ӝ߈੄ ߣ৉ӝܳ ݅ٚ ੉റ ֎੉ߡ ౵౵Ҋ ١ ؀ࠗ࠙ न҃ݎਵ۽. - ҳӖ਷ ঱য ೟੗ܳ Ҋਊೞ૑ ঋח׮Ҋ ೣ. ߣ৉ӝ RNN Y X RNN I RNN LOVE RNN YOU RNN դ RNN օ RNN ࢎی೧
  34. CNNਸ ੉ਊ೧ࢲ о૞ ੉޷૑ܳ ٜ݅য ղח GENERATOR৬ о૞ ੉޷૑ܳ ౸ݺೞח

    DISCRIMINATORо ؀݀ೣ. थܫ 50%о ؼ ٸө૑ ߈ࠂೞݶ ੉޷૑о ࢤࢿ.
  35. - openAIо ݅ٚ ੋҕ૑מ. - Generative Pre-trained Transformer 3 -

    ߣ৉җ ؀ച, ੘ޙ੉ оמ - ࠺૑ب ೟ण - “I love you so much”ী ؀೧ ׮਺ਵ۽ ೟ण. - I -> Love - I love -> you - I love you -> so - I love you so -> much GPT-3
  36. - ࢚׸ਗ੉ ؀ച ࢸ҅. ୁࠈ ؀਽ റ पઁ ࢎۈ੉ ؀਽.

    - ୡӝ ୁࠈ਷ RNNਵ۽ ҳഅ - ௑బஎ, ೟ण, QA ౱ਵ۽ ؀਽. - ஠஠য়ח 89.7%, ஠஠য়ߛ௼ח 34.1% ࢚׸ ୁࠈਵ۽ ؀਽ ৮ܐ. - ੌ߈਷೯਷ 10% ੿بо ୁࠈ ؀਽. ୁࠈ