Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
注意機構を用いた言語創発ゲーム
Search
Ryokan RI
August 19, 2023
Research
0
280
注意機構を用いた言語創発ゲーム
2023 JSAI オーガナイズドセッション
言語とコミュニケーションの創発 ~記号創発システムから共創的言語進化まで~
Ryokan RI
August 19, 2023
Tweet
Share
More Decks by Ryokan RI
See All by Ryokan RI
[論文紹介] Intuitive Fine-Tuning
ryou0634
0
180
[論文紹介] Language is primarily a tool for communication rather than thought
ryou0634
4
1.1k
マルチリンガルな言語モデル入門:これまでとこれから
ryou0634
4
4.6k
What Are You Token About? Dense Retrieval as Distributions Over the Vocabulary
ryou0634
0
710
人工言語を使った事前訓練:言語間転移が可能なエンコーダの持っている知識とは何か?
ryou0634
0
990
MIROSTAT で意外さを コントロールした文章生成
ryou0634
1
1.1k
Other Decks in Research
See All in Research
データサイエンティストの業務変化
datascientistsociety
PRO
0
280
離散凸解析に基づく予測付き離散最適化手法 (IBIS '25)
taihei_oki
PRO
1
710
生成的情報検索時代におけるAI利用と認知バイアス
trycycle
PRO
0
350
Earth AI: Unlocking Geospatial Insights with Foundation Models and Cross-Modal Reasoning
satai
3
580
When Learned Data Structures Meet Computer Vision
matsui_528
1
3.7k
ドメイン知識がない領域での自然言語処理の始め方
hargon24
1
260
2026-01-30-MandSL-textbook-jp-cos-lod
yegusa
1
530
音声感情認識技術の進展と展望
nagase
0
500
SREはサイバネティクスの夢をみるか? / Do SREs Dream of Cybernetics?
yuukit
3
420
LLM-jp-3 and beyond: Training Large Language Models
odashi
1
780
20251023_くまもと21の会例会_「車1割削減、渋滞半減、公共交通2倍」をめざして.pdf
trafficbrain
0
190
Collective Predictive Coding and World Models in LLMs: A System 0/1/2/3 Perspective on Hierarchical Physical AI (IEEE SII 2026 Plenary Talk)
tanichu
1
280
Featured
See All Featured
Rebuilding a faster, lazier Slack
samanthasiow
85
9.4k
Site-Speed That Sticks
csswizardry
13
1.1k
The SEO Collaboration Effect
kristinabergwall1
0
380
Claude Code どこまでも/ Claude Code Everywhere
nwiizo
63
53k
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
9
1.2k
Public Speaking Without Barfing On Your Shoes - THAT 2023
reverentgeek
1
330
Applied NLP in the Age of Generative AI
inesmontani
PRO
4
2.1k
How to make the Groovebox
asonas
2
2k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
133
19k
Un-Boring Meetings
codingconduct
0
220
The Impact of AI in SEO - AI Overviews June 2024 Edition
aleyda
5
760
Visualization
eitanlees
150
17k
Transcript
ҙػߏΛ༻͍ͨ ݴޠൃήʔϜ ཥ ྇פɹ্ా ྄ɹJason Naradowsky ౦ژେֶ
ࢦࣔήʔϜͱʁ - ίϛϡχέʔγϣϯ͔Βࢦࣔͷଆ໘͚ͩΛऔΓग़ͨ͠ήʔϜɻ - ൃίϛϡχέʔγϣϯͷݚڀͰΑ͘ΘΕΔɻ ࢦࣔήʔϜ (referential game)
ࢦࣔήʔϜ (referential game) Speaker 3 14 2 25 ඪΦϒδΣΫτ Listener
ΦϒδΣΫτީิू߹
ࢦࣔήʔϜ (referential game) Speaker 3 14 2 25 Listener ඪΦϒδΣΫτ
ΦϒδΣΫτީิू߹ ᶃ Speaker ͕ ඪΦϒδΣΫτΛݟͯ ϝοηʔδΛੜ
ࢦࣔήʔϜ (referential game) Speaker 3 14 2 25 Listener ඪΦϒδΣΫτ
ΦϒδΣΫτީิू߹ ᶄ Listener ͕ϝοηʔδΛݟͯ ΦϒδΣΫτީิू߹͔Β ඪΦϒδΣΫτΛ༧ଌ͢Δ
ࢦࣔήʔϜ (referential game) Speaker 3 14 2 25 ඪΦϒδΣΫτ Listener
ΦϒδΣΫτީิू߹ ᶅ Listener ͷ༧ଌͱ ඪΦϒδΣΫτ͕Ұக͍ͯ͠Ε ใु͕༩͑ΒΕΔɻ
ࢦࣔήʔϜͷΤʔδΣϯτʹҙػߏΛ࣮͢Δ ຊݚڀͰͬͨ͜ͱ Speaker Listener
ͳͥҙػߏ͔ʁ - ൃݴޠͷߏੑ্͕͕Δͣ - ൃݴޠͷղऍੑ্͕͕Δͣ ຊݚڀͰͬͨ͜ͱ
ͳͥҙػߏ͔ʁ - ൃݴޠͷߏੑ্͕͕Δͣ - ൃݴޠͷղऍੑ্͕͕Δͣ ຊݚڀͰͬͨ͜ͱ
ߏੑͱɿશମ͕෦ͷ૯͔ΒͳΔ͜ͱ ҙػߏൃݴޠͷߏੑΛ্͛Δʁ ߏੑ ߴ🔺 ߏੑ 🔻 5 3 2 3
2 1 6 4
ҙػߏൃݴޠͷߏੑΛ্͛Δʁ 5 Speaker ҙػߏ୯ޠͱΦϒδΣΫτͷ෦ͷ݁ͼ͖ͭ Λֶश͢Δͷʹཱͪͦ͏
ͳͥҙػߏ͔ʁ - ൃݴޠͷߏੑ্͕͕Δͣ - ൃݴޠͷղऍੑ্͕͕Δͣ ຊݚڀͰͬͨ͜ͱ
ҙػߏൃݴޠͷղऍੑΛ্͛Δʁ 5 Speaker ͜ͷ୯ޠ͜ͷҙຯͳͷͩͱΘ͔Δ
࣮ݧઃఆɿࢦࣔήʔϜ Speaker 3 14 2 25 ඪΦϒδΣΫτ Listener ΦϒδΣΫτީิू߹ Fashion
MNIST Λ༻͍ͨࢦࣔήʔϜ ʢεχʔΧʔͱϓϧΦʔόʔʣ
࣮ݧઃఆɿࢦࣔήʔϜ Speaker 3 14 2 25 ඪΦϒδΣΫτ Listener ΦϒδΣΫτީิू߹ Fashion
MNIST Λ༻͍ͨࢦࣔήʔϜ ʢεχʔΧʔͱϓϧΦʔόʔʣ
FashionMNIST - 10 छྨͷΞΠςϜͷը૾ΛؚΉσʔληοτ ࣮ݧઃఆɿࢦࣔήʔϜ
FashionMNIST - 10 छྨͷΞΠςϜͷը૾ΛؚΉσʔληοτ ࣮ݧઃఆɿࢦࣔήʔϜ ̎छྨͷΞΠςϜΛબΜͰ ̍ͭͷΦϒδΣΫτΛ࡞Δ
࣮ݧઃఆɿࢦࣔήʔϜ FashionMNIST Λ༻͍ͨࢦࣔήʔϜ - FashionMNIST த͔Βը૾Λαϯϓϧͯ͠ଟ༷ͳΦϒδΣΫ τΛ࡞ɻ - ಉ͡छྨͷΞΠςϜؚ͕·Ε͍ͯΕಉ͡ΦϒδΣΫτͰ͋Δ ͱΈͳ͢ɻ
- ΦϒδΣΫτͷछྨ - ܇࿅ηοτʹ30छྨɺධՁηοτʹ15छྨ༻ 10 C2 = 45
࣮ݧઃఆɿࢦࣔήʔϜ ਓؒʹ͍ۙݴޠΛֶश͢ΔͨΊʹ ಉ͡छྨͷҟͳΔΦϒδΣΫτΛ Speaker ͱ Listener ʹݟͤΔ ඪΦϒδΣΫτ ΦϒδΣΫτީิू߹
ҙػߏͳ͠ͷ Speaker ҙػߏ͋Γͷ Speaker ࣮ݧઃఆɿϞσϧ ҙػߏͳ͠ͷ Listner ҙػߏ͋Γͷ Listener ×
ͦΕͧΕͷΈ߹ΘͤͰ࣮ݧ ͞ΒʹΤʔδΣϯτΛ࣮͢ΔωοτϫʔΫͱͯ͠ LSTM ͱ Transformer ͦΕͧΕͰ࣮ݧ
࣮ݧઃఆɿϞσϧ ϕʔεϥΠϯɾΤʔδΣϯτ otgt o1 %FDPEFS &ODPEFS m1 m2 m3 ϝοηʔδΦϒδΣΫτ
߹கείΞ m0 m1 m2 m1 m2 m3 Speaker Listener - ҙػߏͳ͠ - ΦϒδΣΫτ୯ҰͷϕΫτϧͰදݱ͞ΕΔ
࣮ݧઃఆɿϞσϧ ҙػߏ͋ΓΤʔδΣϯτ otgt o1 %FDPEFS &ODPEFS m1 m2 m3 "UUFOUJPO
"UUFOUJPO m0 m1 m2 m1 m2 m3 Speaker Listener ϝοηʔδΦϒδΣΫτ ߹கείΞ - ҙػߏ͋Γ - ΦϒδΣΫτෳͷϕΫτϧͰදݱ͞ΕΔ
ҙػߏ͋Γ/ͳ͠ΤʔδΣϯτͷൺֱ - ҙػߏ͋Γͩͱߏੑͷई (TopSim) ্͕͢Δ ҙػߏ͋ΓΤʔδΣϯτͷݴޠͷௐࠪ ࣮ݧͨ͜͠ͱ
ҙػߏ͋Γ/ͳ͠ΤʔδΣϯτͷൺֱ ܇࿅ηοτͷਖ਼ղ ධՁηοτͷਖ਼ղ TopSimʢߏੑͷईʣ - ҙػߏ͕͋Δํ͕ਖ਼ղ͕ߴ͘ͳΔʹ͋Δ - ҙػߏ͕͋Δํ͕ɺTopSim ͕ߴ͘ͳΔʹ͋Δ
ҙػߏ͋ΓΤʔδΣϯτͷ ҙػߏͷՄࢹԽ Speaker Listener Speaker Listener Speaker Listener ޭͨ͠ύλʔϯ ޭͨ͠ύλʔϯ
ࣦഊͨ͠ύλʔϯ
ҙػߏ͋ΓΤʔδΣϯτͷ ൃͨ͠ݴޠͷγϯϘϧͱΦϒδΣΫτͷରԠ Speaker Listener
ҙػߏ͋ΓΤʔδΣϯτͷ ൃͨ͠ݴޠͷγϯϘϧͱΦϒδΣΫτͷରԠ Speaker Listener ্ணܥͷΦϒδΣΫτ ۺܥͷΦϒδΣΫτ
ҙػߏ͋ΓSpeaker ͱ Listener ͷ ҙൣғͷͣΕͱਖ਼ղͷؔ ͣΕ͕গͳ͍΄Ͳਖ਼ղ͍͢͠
ͬͨ͜ͱ - ࢦࣔήʔϜͷΤʔδΣϯτʹҙػߏΛ͚ͭΔ - FashionMNIST ͷը૾ΛͬͨࢦࣔήʔϜͰ࣮ݧ ͔ͬͨ͜ͱ - ҙػߏ͕͋Δͱߏੑ͕ߴ͍ݴޠ͕ൃ͍͢͠
- ҙػߏΛ࣋ͬͨΤʔδΣϯτղऍੑͷ͋ΔݴޠΛൃͤ͞ΒΕΔ ·ͱΊ
࣮͏ͪΐͬͱࣗવͰෳࡶͳը૾ΛͬͨࢦࣔήʔϜͰ ͏·͍ͬͯ͘΄͔͕ͬͨ͠ɺֶश͕͔ͬͨ͠ - ෳࡶͳը૾ͩͱղऍෆೳͳҙͷύλʔϯ͔͠؍͞Εͣ ➡︎ ҙػߏ͕ਓؒͷΑ͏ͳબੑΛ͍࣋ͬͯͳ͍͔Βʁ ਓؒͷ࣋ͭೝػೳͱݴޠͷੑ࣭ͷؔ࿈ͳͲ͕ௐΒΕΔͱ ָͦ͠͏ʁ ࠓޙͷൃల