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
[Human-AI Decision Making勉強会] 正確に予測できるAIは人間の意思決...
Search
Kei Moriyama
January 16, 2024
0
360
[Human-AI Decision Making勉強会] 正確に予測できるAIは人間の意思決定を助けるか?
Kei Moriyama
January 16, 2024
Tweet
Share
More Decks by Kei Moriyama
See All by Kei Moriyama
[読み会]Not All Tokens Are Equal: Human-centric Visual Analysis via Token Clustering Transformer
keimoriyama
0
51
Featured
See All Featured
Sam Torres - BigQuery for SEOs
techseoconnect
PRO
0
190
Lessons Learnt from Crawling 1000+ Websites
charlesmeaden
PRO
1
1.1k
Site-Speed That Sticks
csswizardry
13
1.1k
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
46
2.7k
職位にかかわらず全員がリーダーシップを発揮するチーム作り / Building a team where everyone can demonstrate leadership regardless of position
madoxten
58
50k
Building Experiences: Design Systems, User Experience, and Full Site Editing
marktimemedia
0
410
Joys of Absence: A Defence of Solitary Play
codingconduct
1
290
We Are The Robots
honzajavorek
0
170
Between Models and Reality
mayunak
1
200
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
47
7.9k
Data-driven link building: lessons from a $708K investment (BrightonSEO talk)
szymonslowik
1
920
AI Search: Implications for SEO and How to Move Forward - #ShenzhenSEOConference
aleyda
1
1.1k
Transcript
ਖ਼֬ʹ༧ଌͰ͖ΔAIਓؒͷҙࢥܾఆ Λॿ͚Δ͔? 2024 / 01 / 1 7 Human-AI Decision
making Learning Human-Compatible Representations for Case-Based Decision Support Han Liu, Yizhou Tian, Chacha Chen, Shi Feng, Yuxin Chen, Chenhao Tan ICLR 2 023 https://iclr.cc/virtual/ 20 23 /poster/ 118 14 ౦ژେֶ അݚڀࣨ M1 कࢁ ܛ
ਖ਼֬ͳ༧ଌΛ͢ΔϞσϧҙࢥܾఆͷʹཱͭͷ͔?
ࠓճѻ͏༰ 香 生 AI 文 AI 人 目 人 AI
AI ? AI 十 ? AI 目
ࠓͷจ Learning Human-Compatible Representations for Case-Based Decision Support ICLR 20
23 人 手
߳ઌੜͷൃදͰ৮ΕΒΕ͍ͯ·ͨ͠ https://speakerdeck.com/rinabouk/ humanai-dm-jc- 202 311 15 ?slide= 38
എܠɿAIʹΑΔλεΫͷࣗಈԽ AI
എܠɿAIͷ༧ଌͷઆ໌ʹΘΕΔͷ AI 人 見 Giang Nguyen, Daeyoung Kim, and Anh
Nguyen. The e ff ectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems, 34 : 2642 2 – 2 6 43 6 , 2 02 1 .
ิɿAIͷຒΊࠐΈදݱ AI
എܠɿAIͷຒΊࠐΈදݱͷಛੑ AI 人
എܠɿAIͷຒΊࠐΈදݱͷಛੑ
తɿਓؒͷೝʹԊͬͨຒΊࠐΈදݱͷ֫ಘ
ఏҊख๏ͷલʹɿڑֶशʹ͍ͭͯ
ڑֶशʹ༻͍Δଛࣦؔɿ ه߸ͷఆٛ
ڑֶशʹ༻͍Δଛࣦؔ ℒtriplet (x, xp , xn ) = max {0,d(x,
xp ) − d(x, xn )} 小 d(x, xp ) ≤ d(x, xn ) ℒtriplet (x, xp , xn ) = 0 d(x, xp ) > d(x, xn ) ℒtriplet (x, xp , xn ) > 0 方 小
ఏҊख๏ɿมͷఆٛ
ఏҊख๏ɿଛࣦؔ λ − ∑ (x,y)∼D log(pθ (y|x)) + (1 −
λ) ∑ (xr,x+,x−∼T) max(d(xr, x+) − d(xr, x−) + 1,0) ⾒ 人 ⾒ x y D = {(x, y)i }N i=0 T = {(xr, x+, x−)i }M i=0 λ
࣮ݧ 人工 用 人 比
࣮ݧɿൺֱख๏ MLE 目 ( ) TML 行 目 ( )
HC 手 ( ) RIRO 示 λ = 0 λ = 1 λ = 0.5 λ − ∑ (x,y)∼D log(pθ (y|x)) + (1 − λ) ∑ (xr,x+,x−∼T) max(d(xr, x+) − d(xr, x−) + 1,0) 用 目
࣮ݧɿਓσʔλʹ͓͚Δ࣮ݧ 人工 手 Vespula vs Weevil 虫 用 用
࣮ݧɿਓσʔλʹ͓͚Δ࣮ݧ AI 用 行
࣮ݧ݁Ռɿਓσʔλʹ͓͚Δ࣮ݧʹ͓͚ΔείΞ 用
࣮ݧ݁ՌɿͦΕͧΕͷ࣮ݧʹ͓͚ΔείΞ R 1 R 2 HC 方 人工
࣮ݧ݁ՌɿͦΕͧΕͷ࣮ݧʹ͓͚ΔείΞ 見 ⾒ 見 ⾒ Neutral Persuasive ⾒ 手 人工
示
࣮ݧɿਓؒͷҙࢥܾఆΛରͱ࣮ͨ͠ݧ
࣮ݧઃఆ 2 行
࣮ݧ݁Ռ 手 (HC) 手
࣮ݧ݁Ռ (RIRO) MLE 人
·ͱΊ 文 人 手 人 人 示
ٙ 人工 文 Appendix