Lock in $30 Savings on PRO—Offer Ends Soon! ⏳
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
AlphaGo에서 시작하는 인공지능
Search
Leonardo YongUk Kim
December 11, 2021
Technology
1
290
AlphaGo에서 시작하는 인공지능
AlphaGo, Alpha Zero에서 시작해서 여러 인공지능의 개념을 살펴봅니다.
Leonardo YongUk Kim
December 11, 2021
Tweet
Share
More Decks by Leonardo YongUk Kim
See All by Leonardo YongUk Kim
Recap: Kotlin Language Features in 2.0 and Beyond (Michail Zarečenskij)
dalinaum
0
580
Compose Multiplatform 101
dalinaum
3
620
Kotlin 2.0을 통해 알아보는 코틀린의 미래
dalinaum
1
2.5k
실리콘밸리 스타트업에서 일어난 일
dalinaum
0
110
Zip: Data compression (20분만에 배우는 압축 알고리즘)
dalinaum
1
2.3k
안드로이드 빌드: 설탕없는 세계
dalinaum
0
140
Obfuscation 101 @ Naver Tech Concert
dalinaum
4
530
Realm은 어떻게 효율적인 데이터베이스를 만들었나?
dalinaum
1
560
MVC부터 MVVM, 단방향 데이터 흐름까지
dalinaum
5
850
Other Decks in Technology
See All in Technology
SLMをエッジAIとして検証してみて分かったこと
iotcomjpadmin
0
300
Entra ID の多要素認証(Japan Microsoft 365 コミュニティ カンファレンス 2024 )
murachiakira
0
1.7k
SONY AITRIOSによるAIエッジセンシングの新たな可能性(仮)
iotcomjpadmin
0
310
Engineer Recruting Deck
siva_official
PRO
1
3.2k
日本全国・都市3D化プロジェクト「PLATEAU」とデータ変換OSS「PLATEAU GIS Converter」の公開
nokonoko1203
2
290
Amazon Forecast亡き今、我々がマネージドサービスに頼らず時系列予測を実行する方法
sadynitro
0
220
iOS 18 から追加された SwiftUI の傾向について調べてみる
swiftty
2
110
Raspberry Pi 秋の新製品をチェックしてみよう / 20231202-rpi-jam-tokyo
akkiesoft
0
160
情シスの引継ぎが大変という話
miyu_dev
2
530
累計2500万着電を支える大規模 電話自動応答サービスのアーキテクチャ / Architecture of a Large-Scale Automated Phone Response Service Supporting 25 Million Cumulative Calls
ymachida
8
4.3k
システムリプレイスプロジェクト発足から7年、改めてコスト最適化に向き合う / replace and cost optimization
takumi
1
150
OpenLLMetry-Hands-On 生成AIアプリを観測してみよう!OpenLLMetryハンズオン編
tkhresk
1
150
Featured
See All Featured
VelocityConf: Rendering Performance Case Studies
addyosmani
326
24k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
44
2.2k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
229
52k
The Cult of Friendly URLs
andyhume
78
6.1k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
226
22k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
250
21k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
42
9.3k
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
27
4.3k
Raft: Consensus for Rubyists
vanstee
136
6.7k
Teambox: Starting and Learning
jrom
133
8.8k
Building Your Own Lightsaber
phodgson
103
6.1k
Learning to Love Humans: Emotional Interface Design
aarron
273
40k
Transcript
LEONARDO YONGUK KIM
[email protected]
ALPHAGOীࢲ दೞח ੋҕמ
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 ঌҊ ݃झఠ৬ ऱਕ 89थ 11ಁ ࣳӝ, झ, झఋ 2 ࠂ Goܳ ܴীࢲ ઁ৻
ݾ ࣻ ੍ӝ न҃ݎ ъച ण ੋҕמ ࢎਊ
ࣻ ੍ӝ MiniMax Monte Carlo Method Monte Carlo Tree Search
MINIMAX
TIC TAC TOE O O O X X O X
O X X O X O - ݢ 3ѐܳ աۆ ֬ ࢎۈ ӝח ѱ - ب ࠂೠ 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
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 धۄҊب פ. ݽٚ ֢٘ח ध ࠗݽੑפ.
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 ࠳ےܳ ઁѢ
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 ࠳ےח ઁѢ
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
None
- (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 ҃ ࣻ
MONTE CARLO METHOD
ਗਯਸ য ࠇद - ਗਯਸ णפ. - ਗਯਸ ా҅ਵ۽ ҳೡ
ࣻ ਸөਃ?
MONTE CARLO METHOD - ےؒೞѱ ਸ ନद. - ਗ উ
іࣻо 314ѐ, ࢎпഋ উ іࣻо 400ѐۄ о೧ ࠇद. - 4(R ^ 2) : π(R ^ 2) = 400 : 314, π = 3.14 - ࠙ ਸ ݆ ନਸ ࣻ۾ ؊ ೠ ਗਯਸ ঌѱ ؾפ.
ޅ ࠊب ݆ ࠁݶ ঋਸө?
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 ୭ҊҊ 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.
Ӓېب ࣻ ৻ী ݽف ےؒ ખ Ӓۧ ঋաਃ?
MONTE CARLO TREE SEARCH
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.
O O O O O O O O O ROOT
- ઁ थܫ ֫ 5ߣ૩ ࣻ ਤ۽ ےؒ Ѿਸ פ. (70%) - աݠ ٜࣻب ےؒ Ѿਸ оՔפ. (30%) - ֬ ࣻо ӝ ٸޙੑפ.
O O O O O O O O O ROOT
- 5ߣ૩ ࣻী ೧ যו ب ےؒ Ѿਸ ೮ਵݶ, 5ߣ૩ ࣻ धب ܻী ನೣפ. - ઁ ےؒೞѱ ف݅ ف ࣻח ҊೞҊ פ. - ର ৌ۰ ח ֢٘о ݆ই פ. X O X O X O O X X O O X O X O X
None
- झח 1996֙ 2ਘ 10ੌ MiniMax ۽ ࠂ. (गಌ ஹೊఠ
٩ ࠶ܖ) - ୭Ӕীח Monte Carlo Tree Search۽ ോಪਵ۽ب Ӓے٘݃झ ఠܳ ӣ. - ೞ݅ ߄ق ࠛоמ೮. - ഛܫਵ۽ ೞӝূ ڜࣻо ցޖ ݆ . ࣻ ੍ӝ݅ਵ۽ ࠙ೠоਃ?
ח Ҕ݅ ٮઉࠁݶ উغਃ?
(ࢎۈ ࢤпী) ח Ҕ݅ ٮઉࠁݶ উغਃ?
न҃ݎ न҃ݎ CNN
न҃ݎ
ӝ नഐܳ ߉ই оҕ೧ ӝ नഐܳ ࠁղח Ѫ? ۠
X1 X2 Y W1 W2 Y = W1X1 + W2X2
ࣻ ݆ ۠ Ѿ೧ (֎ਕ) מਸ ٜ݅যմ Ѫ ۢ, ࣻ рױೠ о ো(ੋҕ ۠)ਸ ֎ਕܳ ٜ݅ݶ যڌѱ ؼөਃ?
- ӝ 0~9ܳ Ҵ NISTо ࣻ. - о۽ 28, ࣁ۽
28 ࣄ۽ ҳࢿ. - ೠ Ӗ 784 ࣄ (28 x 28) MNIST
None
- ࡈр࢝ ਵݶ উػח ڷ. - ۆ࢝ ਵݶ
જח ڷ. - 0 оؘח ਵݶ উػ. - 1 оؘח ਸ оמࢿ ֫ . - ࡈъ -1, ی 1۽ ࠁ. Ѩ ࢝ 0 ӝ ࠂೞ.
X1 X784 W1 W784 W1X1 + … + W784X784 =
Y Yо ݶ ӝо 5ۄח ڷ.
- W1X1 + … + W784X784 = Y - X1ࠗఠ
X784ө ֎ਃ? ੑ۱ 784ѐ - W1ࠗఠ W784ө ֎ਃ? оо 784ѐ - ইۄ࠺ই ंо 0ࠗఠ 9өਗ਼ইਃ? оо 7840ѐ (784 x 10) न҃ݎ
X1 X2 Y1 W1 W2 X784 … Y2 Y10 …
784ѐ ੑ۱ 0ੋоਃ? 1ੋоਃ? 9ੋоਃ? ߣ૩ ࣄ W10 W7840 7840ѐ о 10ѐ ۱
X1 X2 Y1 X784 … Y2 Y10 … ੑ۱க ۱க
оח ־о աਃ?
- ӝ҅о ҕࠗೞӝ ٸޙী ӝ҅ ण - ӝ҅о ҕࠗೡ ࣻ
ѱ ޙઁ৬ ਸ ળ࠺פ. - MNISTۄח ӝ ࣁب ৬ э ઓפ. - ف ઙܨ ؘఠܳ ৮ ܻ࠙ೡ ӝӝ৬ ಞਸ ӝ҅о ইմҊ ࠁݶ ؾפ. - ण ؘఠܳ оҊ ࣻ হ ࠙ਸ ߈ࠂ ೞݴ оܳ ઑӘঀ Ҋоݴ ৢ߄ܲ ݽ؛ਸ ٜ݅যцפ. - ਸ Ҋ णਸ दఃח Ѫਸ ب णۄҊ פ. ӝ҅ण
None
- Ӓېਸ ਤ೧ GPUח ࣻ হ Ӓܿ৬ ࡄ ై ؼ
ܳ ҅ೞҊ ژ ҅פ. - GPU ఌਗೡ ҅ מ۱ ݠन۞җ ঐഐ ചತ ী פ. - ߈ݶ GPUח ਬোೠ ౸ױਸ ޅפ. - 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.
NVIDIA ୡഋ GPU (2রо)
GOOGLE ݠन۞ ਊ TPU (TENSOR PROCESSING UNIT)
־о աਃ?
None
None
࠙ܨ ޙઁܳ ۽Ӓې߁ ਵ۽ ೧Ѿೡ ࣻ হաਃ?
None
- ೦࢚ ৻о ਸ ࣻ णפ. (Ҋন৬ 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
- য়ܲଃ ޙઁח ࢶਸ Ӓযࢲ ޙઁܳ ೧Ѿೡ ࣻ হ. -
ࠂೠ ޙઁח க ۨযо ਃ. ੑ۱கҗ ۱கਵ۽ח উغח ޙઁ
X1 X2 X784 … … ੑ۱க Y1 Y2 Y10 …
۱க ץக - ࠂೠ ޙઁח 1ѐ ࢚ ץக(Hidden Layer)ਸ ٟ݅פ. - 2ѐ ࢚ ץகਸ णೞח ҃ Deep LearningۄҊ ೞݴ बக न҃ݎ (Deep Neural Network)ۄח ݺடਸ ࢎਊפ. - Microsoftח 152கਸ о ResNetਸ ٜ݅ റ 1001ѐ கө ٜ݅णפ.
MNIST पઁ ٘ ؘݽ
CNN
Ҋন VS ѐ - बகݎਵ۽ח ࠙ܨ ޙઁܳ ࠂೞӝ য۰ਛणפ.
- CNN (Convolutional Neural Network) AlexNet աয়ݶࢲ ޙઁܳ ೧Ѿ. - 2012֙ 9ਘ 30ੌࠗఠ ஹೊఠо ҙ ਸ ֈযࢲӝ द೮णפ.
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 ъചೠҊ ࢤпೞҊ णפ. - যڃ ఠח ਮҘਸ ъച? - যڃ ఠח ӈա ܳ ؊ ੜ ࠁѱ ъച? э ࢚࢝ՙܻ ғೞҊ ؊פ.
X1 X2 X784 … … ੑ۱க Y1 Y2 Y10 …
۱க ץக ழօ 3ѐ ఠ
CNNਸ ా೧ റࠁܳ ҳೞҊ ހప ܳ۽ ܻ ࢲ۽ ࣁࠗੋ ࣻܳ
੍णפ.
CNN पઁ ٘ ؘݽ
ъചण ъചण
None
0ࠗఠ 8ө ਸ ࢚కۄҊ ࢤпद. ࢚ೞઝ ೞաܳ Ҋܰח Ѫਸ
ঘ࣌ۄҊ ࠁҊਃ.
ъചण - ഒ ۨ (Self Play)ܳ ೞݴ प۱ਸ ט۰х. -
(࢚క, ঘ࣌, थಁ ৈࠗ)ܳ оҊ ण, റ ࢚కী ೧ ࣻ೯೧ঠ ೡ ঘ࣌ਸ ঌ۰ષ. - ঌҊח झझ۽ ؊ ъ೧. (ࢎ णೠ ӝࠁח ӂ ೧Ѿػ 16݅ ӝࠁ. ࠗ࠙ അ ۽о ইש.) - CNN (଼ݎ, оݎ) + ހప ܳ۽ ܻ ࢲ + ъച ण - ঌ ઁ۽ח ӝࠁ णب ೞ ঋҊ ъച ण݅ਵ۽ ֎ਕܳ ҳ୷. X1 X2 X784 … … ࢚క Y1 Y2 Y10 … ঘ࣌ ץக ழօ
MINIMAX۽ ೧ب ؾפ. ъചण + न҃ݎ + ހపܳ۽ ܻ ࢲ۽
ٜ݅णפ. ALPHAZEROی э ߑधਵ۽ਃ.
ALPHA ZERO ۿ ؘݽ
ੋҕמ ࢎਊ ܻ࠭ ಣо ߣӝ ࢤࢿ न҃ݎ ਯ
೯ ର GPT-3 न࠙ૐ ੋध ୁࠈ नਊಣоݽ؛
- “ঘ࣌ ഴܯೞҊ ࠺ ഴܯ೮णפ.” -> ଵ (୶ୌ) - “ցޖ
ਤҊ Ѣܻо ࡞೮णפ.” -> Ѣ (࠺୶) - CNNਵ۽ णदெ ࢜۽ ܻ࠭ী ೧ ୶ୌੋ ࠺୶ੋ ഛੋ. ܻ࠭ ಣо
- RNNਸ ࢎਊ. CNNҗ ର ץக ؘఠܳ द ץகਵ۽ (ӝর۱)
- ҳӖ न҃ݎ ӝ߈ ߣӝܳ ݅ٚ റ ֎ߡ Ҋ ١ ࠗ࠙ न҃ݎਵ۽. - ҳӖ য ܳ Ҋਊೞ ঋחҊ ೣ. ߣӝ RNN Y X RNN I RNN LOVE RNN YOU RNN դ RNN օ RNN ࢎی೧
- GAN (Generative Adversarial Network) - ف ࢎۈ ઓೞ
ঋ ࢎۈ. ࢤࢿ न҃ݎ
CNNਸ ਊ೧ࢲ о ܳ ٜ݅য ղח GENERATOR৬ о ܳ ౸ݺೞח
DISCRIMINATORо ݀ೣ. थܫ 50%о ؼ ٸө ߈ࠂೞݶ о ࢤࢿ.
IMAGE COLORING & IMAGE NOISE REDUCTION
None
ਯ ೯ ର
None
- openAIо ݅ٚ ੋҕמ. - Generative Pre-trained Transformer 3 -
ߣҗ 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
GPT-2о ೠҴয ؘݽо যࢲ ੋਊ. SKTо ݅ٚ KOGPT2
- न࠙ૐ ղਊ ੋध. - о न࠙ૐੋח Ѩૐ೧ঠೣ. - CNNਸ
ਊ. न࠙ૐ ੋध
- ࢚ਗ ച ࢸ҅. ୁࠈ റ पઁ ࢎۈ .
- ୡӝ ୁࠈ RNNਵ۽ ҳഅ - బஎ, ण, QA ਵ۽ . - য়ח 89.7%, য়ߛח 34.1% ࢚ ୁࠈਵ۽ ৮ܐ. - ੌ߈೯ 10% بо ୁࠈ . ୁࠈ
- ਬ ੑ۱ ղਸ ࠙ࢳ೧ࢲ زਵ۽ Ә. Әా ز ݽਵӝ
- ৈ۞ Ҋё ؘఠܳ ਊ೧ ݠन۞ਸ ਊ. - नਊಣоܳ ా೧
नਊ. नਊಣоݽഋ
хࢎפ.