Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
AGI福岡 第3回
Search
Sponsored
·
SiteGround - Reliable hosting with speed, security, and support you can count on.
→
yuky_az
November 14, 2024
Technology
150
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
AGI福岡 第3回
AGI時代に向けて、今からできることを共に考え、実行していこう!
yuky_az
November 14, 2024
More Decks by yuky_az
See All by yuky_az
AGI福岡 第8回
yukinaga
0
160
AGI福岡 第6回
yukinaga
0
190
AGI福岡 第5回
yukinaga
0
240
AGI福岡 第2回
yukinaga
0
170
生成AIの現状と展望: AIと共生する未来への道程
yukinaga
3
1.2k
BERTによる自然言語処理を学ぼう!【 Live!人工知能 #26】 #Live人工知能
yukinaga
0
470
iOSアプリは「感情」を宿すのか? AIとアプリの未来について
yukinaga
2
1.1k
iOSアプリに「意識」は宿るのか? ディープラーニングの先にある人工知能(AI)
yukinaga
6
6.3k
ヒトとAIの共生、そしてシンギュラリティ
yukinaga
0
1.1k
Other Decks in Technology
See All in Technology
分析の民主化を支えるメタデータ整備の実践
kakehashi
PRO
1
660
子育てエンジニアの可処分時間が減っても成長を諦めない戦い方
sudoakiy
1
1.4k
SDDの運用にめげずに向き合った話
sansantech
PRO
1
130
予測不能な実行エージェントを安全にサクッと試せるサンドボックス環境の話
sansantech
PRO
0
130
【データ横丁主催】AI Agentがコンテキストを使って仕事をした後、何が残るのか― 組織の経験を次の判断に引き継ぐ「Agent Memory」
shisyu_gaku
2
330
並行性の問題を防げ!実践トランザクション入門
occhi
0
280
Execution in the Kingdom of Agents: Reflections on Abstraction and Complexity
bcantrill
0
870
検証フェーズはAIの回答を判断するための学習機会
toru_kubota
2
510
The seven pitfalls of AI (revised version)
ufried
0
280
Futexes the good, the bad, the ugly
ennael
PRO
0
130
AI時代のAPI開発を加速する品質ガードレール / API Quality Guardrails in the AI Era
yokawasa
0
160
Google Cloud Next Tokyo 26登壇時のスクリプト
recruitengineers
PRO
0
200
Featured
See All Featured
Redefining SEO in the New Era of Traffic Generation
szymonslowik
1
460
Put a Button on it: Removing Barriers to Going Fast.
kastner
60
4.6k
Applied NLP in the Age of Generative AI
inesmontani
PRO
4
2.5k
Crafting Experiences
bethany
1
370
The MySQL Ecosystem @ GitHub 2015
samlambert
251
13k
Designing Experiences People Love
moore
143
24k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
234
18k
Tips & Tricks on How to Get Your First Job In Tech
honzajavorek
1
790
Product Roadmaps are Hard
iamctodd
55
13k
What the history of the web can teach us about the future of AI
inesmontani
PRO
1
720
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
35
3.8k
Writing Fast Ruby
sferik
630
63k
Transcript
"(*Ԭୈճ !ΤϯδχΞΧϑΣ
Φʔϓχϯά
ʮ൚༻ਓೳʢ"(*ʣʯͱʮಛԽܕਓೳʯ ൚༻ਓೳʢ˺ڧ͍"*ʣ ˠώτͷೳͷΑ͏ͳ൚༻ੑΛ࣋ͭ"* FHυϥ͑ΜɺమΞτϜͳͲ ಛԽܕਓೳʢ˺ऑ͍"*ʣ ˠݶఆతͳղܾਪ FHνΣεকعͷ"*ɺը૾ೝࣝͳͲ ࠷৽ͷ--.ͲͪΒʁ
"(*ͷొؒۙʁ 4*56"5*0/"-"8"3&/&44 ΑΓ IUUQTTJUVBUJPOBMBXBSFOFTTBJGSPNBHJ UPTVQFSJOUFMMJHFODF
ϊʔϕϧཧֶ © Johan Jarnestad/The Royal Swedish Academy of Sciences John
Hop fi eld and Geoffrey Hinton. Ill. Niklas Elmehed © Nobel Prize Outreach ͷϊʔϕϧཧֶɺ δϣϯɾ+ɾϗοϓϑΟʔϧυʢถࠃʣͱδΣϑϦʔɾ&ɾώϯτϯʢΧφμʣ͕ड
ϊʔϕϧཧֶ • डཧ༝ → ػցֶशͱਓχϡʔϥϧωοτϫʔΫͷج൫Λߏங͠ɺσʔλύλʔϯͷ هԱɾ࠶ݱΛՄೳʹͨ͠ޭɻ • ओͳۀ → ϗοϓϑΟʔϧυɿεϐϯܥͷཧֶͷݪཧΛԠ༻ͨ͠ʮϗοϓϑΟʔϧυ
ωοτϫʔΫʯΛ։ൃ͠ɺσʔλύλʔϯͷอଘͱ෮ݩΛ࣮ݱɻ → ώϯτϯɿʮϘϧπϚϯϚγϯʯΛ։ൃ͠ɺσʔλ͔ΒಛΛֶश͢Δٕज़ Λཱ֬ɻAIͱػցֶशͷඈ༂తͳਐలʹߩݙɻ • Өڹ → AIͱཧֶͷ༥߹ʹΑΔൃలΛଅਐ͠ɺԠ༻ͰͷՄೳੑΛେ͖֦͘େɻ
"(*Ԭͷҙٛ w "(*࣌ʹ͚ͯɺࠓ͔ΒͰ͖Δ͜ͱΛڞʹߟ͑ɺ ࣮ߦ͍ͯ͜͠͏ʂ
ӡӦऀհ w զ࠺ʢ4"*-BCגࣜձࣾʣ w ۙ౻ݑࣇʢגࣜձࣾελσΟετʣ w େ৴ߊʢ/0#%"5"גࣜձࣾʣ w ੨྄ʢςΠϧΠϯυגࣜձࣾʣ
දऀհ զ࠺ :VLJOBHB"[VNB !ZVLZ@B[ 4"*-BCגࣜձࣾදऔక w "*ؔ࿈ͷڭҭɺݚڀ w ཧֶത࢜ʢཧֶʣ w
๏େֶσβΠϯֶ෦݉ߨࢣ w 6EFNZͰສਓΛࢦಋ w ༗໊اۀͰ"*ݚमΛ୲ w ஶॻʹʮ͡ΊͯͷσΟʔϓϥʔχϯάʯͳͲ w झຯϒϥδϦΞϯॊज़"*Ξʔτ
ϋογϡλά BHJGVLVPLB
"(*Ԭ%JTDPSEίϛϡχςΟ
:PV5VCFϥΠϒ ʮ&OHJOFFSDBGF"(*ԬʯͰ:PV5VCFݕࡧ
ࠓճͷίʔυ IUUQTHJUIVCDPNZVLJOBHBIPQ fi FME@CPMU[NBOO
ϗοϓϑΟʔϧυωοτϫʔΫͱʁ
δϣϯɾϗοϓϑΟʔϧυࢯͷհ ݚڀऀͱͯ͠ͷܦྺ • ੜ·ΕɺΞϝϦΧͷཧֶऀɾੜཧֶऀ • ϓϦϯετϯେֶͰཧֶΛઐ߈ʢ1I%ʣ • ΧϦϑΥϧχΞՊେֶɺϓϦϯετϯେֶͳͲͰڭ佃 • ཧֶ͔ΒਆܦՊֶݚڀྖҬΛ֦େ
• ෳࡶܥͷݚڀͰଟେͳޭ ओͳۀͱධՁ • ɿϗοϓϑΟʔϧυωοτϫʔΫΛఏҊ • ཧֶͷ֓೦Λ༻͍ͯχϡʔϥϧωοτΛཧԽ • ݱͷਂֶशͷཧతج൫Λங͍ͨઌۦऀͷҰਓ Ill. Niklas Elmehed © Nobel Prize Outreach
ϗοϓϑΟʔϧυωοτϫʔΫͷݩจ ถࠃՊֶΞΧσϛʔلཁʹܝࡌ IUUQTXXXQOBTPSHEPJQOBT
جຊΞʔΩςΫνϟ ωοτϫʔΫߏ • ͯ͢ͷχϡʔϩϯ͕૬ޓʹ݁߹ • ֤χϡʔϩϯํʹଓɺͨͩࣗ͠ݾ݁߹ଘࡏ͠ͳ͍ • ݁߹ͷॏΈରশతʢ ʣ •
χϡʔϩϯͷঢ়ଶࢄʢ ʣ·ͨೋʢ ʣ ಈ࡞ͷಛ • ΤωϧΪʔؔ ʹجͮ͘ঢ়ଶߋ৽ ◦ ɿχϡʔϩϯ ͷᮢʢόΠΞεʣ • ඇಉظతͳঢ়ଶߋ৽ʢҰʹͭͷχϡʔϩϯʣ • ہॴతͳใͷΈͰߋ৽அ • ΤωϧΪʔ࠷খԽʹΑΔ҆ఆঢ়ଶͷऩଋ • ཧֶͷεϐϯάϥεϞσϧͱྨࣅͨ͠ৼΔ͍ wij = wji si ∈ − 1, + 1 si ∈ 0,1 E = − 1 2 ∑ i,j wij si sj − ∑ i θi si θi i
ಈ࡞ݪཧ ঢ়ଶߋ৽ͷϝΧχζϜ • ֤χϡʔϩϯͷೖྗɿ • ߋ৽نଇɿ ࢄʢۃੑʣͷ߹ɿ ೋʢ୯ۃੑʣͷ߹ɿ
• ඇಉظతͳߋ৽ʢϥϯμϜʹબΜͩͭͷχϡʔϩϯΛߋ৽ʣ • ͯ͢ͷχϡʔϩϯ͕҆ఆ͢Δ·Ͱ܁Γฦ͠ hi = ∑ j wij sj + θi si = { +1 JG hi ≥ 0 −1 JG hi < 0 si = { 1 JG hi ≥ 0 0 JG hi < 0
ಈ࡞ݪཧ ΤωϧΪʔͱऩଋ • ΤωϧΪʔؔঢ়ଶߋ৽ʹΑΓ୯ௐݮগ • γεςϜඞͣہॴ࠷খʹऩଋ • ࠷ऴঢ়ଶॳظঢ়ଶʹґଘ • ΤωϧΪʔܗͷΠϝʔδɿ
◦ࢁʢෆ҆ఆঢ়ଶʣ͔Β୩ʢ҆ఆঢ়ଶʣ ◦ෳͷ҆ఆঢ়ଶʢہॴ࠷খʣ͕ଘࡏ ◦هԱύλʔϯ҆ఆঢ়ଶͷҰͭʹରԠ
ֶशͱهԱ ֶशʢॏΈͷܾఆʣ • ݁߹ՙॏͷֶशଇɿ ʢ ʣ ◦ ɿχϡʔϩϯ ◦ ɿهԱͤ͞Δύλʔϯ
◦ ɿύλʔϯ ʹ͓͚Δχϡʔϩϯ ͷঢ়ଶ • ҰͷܭࢉͰॏΈΛܾఆʢඇ෮తʣ • ੜֶతͳϔϒଇʹجͮ͘ هԱ༰ྔͱಛ • ཧతͳهԱ༰ྔɿ ◦ ݸͷχϡʔϩϯͰ ݸͷύλʔϯΛهԱՄೳ • ϊΠζΛؚΉೖྗ͔Βͷى͕Մೳ • ύλʔϯͷҰ෦͔ΒશମΛ࠶ߏ wij = 1 N ∑p μ=1 sμ i sμ j i ≠ j N p sμ i μ i pmax ≈ 0.14N N 0.14N
Ԡ༻ྫͱಛ දతͳԠ༻ྫ • ࿈هԱγεςϜ ◦ύλʔϯ෮ݩɾϊΠζআڈ ◦ෆશͳೖྗ͔Βͷى • Έ߹Θͤ࠷దԽ ◦८ճηʔϧεϚϯ ◦εέδϡʔϦϯά
རͱݶք • རɿฒྻॲཧ͕ՄೳɺϋʔυΣΞ࣮͕༰қ • ݶքɿهԱ༰ྔͷ੍ݶʢ ʣɺہॴղʹؕΔՄೳੑ 0.14N
ൃలͱӨڹ తͳൃల • ϘϧπϚϯϚγϯͷਐԽ ◦֬తͳৼΔ͍ͷಋೖɺӅΕͷՃ • ࿈ଓϞσϧͷ։ൃ ◦ΞφϩάχϡʔϩϯͷಋೖɺΑΓ๛͔ͳදݱೳྗͷ֫ಘ ݱͷӨڹ •
ཧֶͷݟͷ׆༻ ◦ΤωϧΪʔ࠷খԽʹΑΔֶश ◦֬తΞϓϩʔνͷجૅ
ϗοϓϑΟʔϧυωοτϫʔΫͷίʔυ IUUQTHJUIVCDPNZVLJOBHBIPQ fi FME@CPMU[NBOOUSFFNBJOTFDUJPO@
ϘϧπϚϯϚγϯͱʁ
δΣϑϦʔɾώϯτϯࢯͷհ Ill. Niklas Elmehed © Nobel Prize Outreach ܦྺͱݚڀ •
ΠΪϦεੜ·ΕɺʮσΟʔϓϥʔχϯάͷʯ • τϩϯτେֶڭतɺ(PPHMF3FTFBSDIॴଐ • ೝ৺ཧֶ͔Βਓೳݚڀ • ਂֶशͷཧతجૅΛཱ֬ ओͳۀ • ޡࠩٯ๏ͷ࠶ൃݟʢʣ • %FFQ#FMJFG/FUXPSLͷఏҊʢʣ • νϡʔϦϯάडʢʣ • "*҆શੑͷܯʢʣ
ϘϧπϚϯϚγϯͷݩจ τϩϯτେֶͷΣϒαΠτʹܝࡌ IUUQTXXXDTUPSPOUPFEVdGSJU[BCTQTDPHTDJCNQEG
എܠͱੜ ։ൃͷഎܠ • ɺϗοϓϑΟʔϧυωοτϫʔΫͷݶք • ౷ܭྗֶʢϘϧπϚϯʣͷԠ༻Λண • )JOUPO4FKOPXTLJʹΑΔڞಉݚڀʢʣ • ֬తͳৼΔ͍ͷಋೖ
ओཁͳֵ৽ • ֬తͳχϡʔϩϯͷಋೖ • ԹύϥϝʔλʹΑΔ੍ޚ • ӅΕϢχοτͷಋೖ • ΤωϧΪʔ࠷খԽͱ֬తֶशͷ౷߹
جຊΞʔΩςΫνϟ ωοτϫʔΫߏ • ՄࢹͱӅΕͷߏ • ͯ͢ͷϢχοτ͕ؒํʹ݁߹ • ֬తͳঢ়ଶભҠɿ ◦
ɿঢ়ଶมԽʹ͏ΤωϧΪʔมԽ ◦ ɿԹύϥϝʔλ ΤωϧΪʔؔ • ◦ ɿϢχοτؒͷ݁߹ॏΈ ◦ ɿϢχοτͷঢ়ଶʢ·ͨʣ ◦ ɿόΠΞε߲ • Թ ʹΑͬͯঢ়ଶΛ੍ޚ p(si = 1) = 1 1 + e−ΔEi/T ΔEi T E = − ∑ i<j wij si sj − ∑ i θi si wij si θi T
ಈ࡞ϝΧχζϜ ঢ়ଶભҠͷΈ • ϘϧπϚϯʹै͏֬తͳঢ়ଶߋ৽ ◦ ◦ ɿؔ ɿશϢχοτͷঢ়ଶϕΫτϧ •
ΪϒεαϯϓϦϯάʹΑΔঢ়ଶભҠ ◦ϥϯμϜʹબΜͩϢχοτΛ֬తʹߋ৽ ◦ฏߧঢ়ଶʹ౸ୡ͢Δ·Ͱ܁Γฦ͠ Թ੍ޚͱ࠷దԽ • γϛϡϨʔςουΞχʔϦϯάͷར༻ ◦ߴԹ͔Β։࢝͠ɺঃʑʹԹΛԼ͛Δ ◦ہॴղճආͱେҬత࠷దԽ • ԹʹΑΔ୳ࡧɾऩଋͷ੍ޚ ◦ߴԹɿϥϯμϜͳ୳ࡧԹɿہॴతͳ࠷దԽ P(s) = 1 Z e−E(s)/T Z s
ֶशΞϧΰϦζϜ ֶशͷجຊࣜ • ॏΈߋ৽ଇɿ ◦ ɿֶश ◦ ɿσʔλͷظʢਖ਼૬ʣ ◦
ɿϞσϧͷظʢෛ૬ʣ ্࣮ͷ • ίϯτϥεςΟϒμΠόʔδΣϯεʢ$%ʣ๏ͷಋೖ ◦શͳऩଋΛͨͣʹֶश ◦গεςοϓͷαϯϓϦϯάͰ༻ • Թεέδϡʔϧͷઃఆ ◦ֶशॳظߴԹͰ୳ࡧతʹ ◦ֶशޙظԹͰऩଋΛଅਐ Δwij = η(⟨si sj ⟩data − ⟨si sj ⟩model ) η ⟨ ⋅ ⟩data ⟨ ⋅ ⟩model
ൃలͱԠ༻ Ϟσϧͷൃల • ੍ݶϘϧπϚϯϚγϯʢ3#.ʣͷߟҊ ◦ؒͷΈͷ݁߹ʹΑΔޮԽ ◦%FFQ#FMJFG/FUXPSLͷجૅʹ • %FFQ-FBSOJOHֵ໋ͷߩݙ ◦ਂֶशͷࣄલֶशख๏ͱͯ͠׆༻ ◦ੜϞσϧͷجຊ֓೦Λཱ֬
ϘϧπϚϯϚγϯͷίʔυ IUUQTHJUIVCDPNZVLJOBHBIPQ fi FME@CPMU[NBOOUSFFNBJOTFDUJPO@
Ԭͷ"(* Γ্͍͖͛ͯ·͠ΐ͏ʂ