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
[読み会] Learning Representations by Humans, for H...
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
mei28
April 19, 2022
49
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
[読み会] Learning Representations by Humans, for Humans
読み会資料
Learning Representations by Humans, for Humans(ICML2021)
mei28
April 19, 2022
More Decks by mei28
See All by mei28
[FAccT'26] Do User-Aligned Explanations Steer Human Decisions? Context-Dependent Influence and Ethical Implications
mei28
0
31
[Human-AI Decision Making勉強会] 説明を部分的に見せることで人に考えさせ、AIへの不適切な依存を減らす
mei28
0
130
[読み会] CHI2025論文紹介
mei28
1
90
[読み会] “Are You Really Sure?” Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making
mei28
0
180
[JSAI'24] 人間の判断根拠は文脈によって異なるのか?〜信頼されるXAIに向けた人間の判断根拠理解〜
mei28
2
890
[CHI'24] Fair Machine Guidance to Enhance Fair Decision Making in Biased People
mei28
0
130
[DEIM2024] 卓球の得点予測における重要要素の分析
mei28
0
72
[Human-AI Decision Making勉強会] 意思決定 with AIは個人vsグループで変わるの?
mei28
0
270
[読み会] Words are All You Need? Language as an Approximation for Human Similality Judgements
mei28
0
97
Featured
See All Featured
世界の人気アプリ100個を分析して見えたペイウォール設計の心得
akihiro_kokubo
PRO
71
40k
Being A Developer After 40
akosma
91
590k
Ruling the World: When Life Gets Gamed
codingconduct
0
250
Jamie Indigo - Trashchat’s Guide to Black Boxes: Technical SEO Tactics for LLMs
techseoconnect
PRO
0
170
The Illustrated Children's Guide to Kubernetes
chrisshort
51
52k
The Limits of Empathy - UXLibs8
cassininazir
1
360
VelocityConf: Rendering Performance Case Studies
addyosmani
333
25k
Speed Design
sergeychernyshev
33
1.8k
Amusing Abliteration
ianozsvald
1
210
The Language of Interfaces
destraynor
162
27k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
32
2.9k
Gemini Prompt Engineering: Practical Techniques for Tangible AI Outcomes
mfonobong
2
440
Transcript
ൃදऀɿ༶໌ ݄!ಡΈձ -FBSOJOH3FQSFTFOUBUJPOT CZ)VNBOT GPS)VNBOT
จใͱબΜͩཧ༝ બཧ༝ • ਓؒʹରͯ͠ใٕज़Ͱ࡞༻ʢհೖʣʹ͍ͭͯΔͨΊ • )VNBOJOUIFMPPQܥͷจ
ߩݙɿਓؒΛհͯ͠ਓؒʹ༗ӹͳදݱΛ֫ಘՄೳʹ എܠɿػցֶश͕ൃల͖ͯͯ͠ɼҙࢥܾఆʹΘΕ͖͍ͯΔ ɿ҆શੑɼެฏੑͳͲΛߟྀ͢ΔࡍʹӏವΈʹͣ͠Β͍ ఏҊɿਓ͕ؒཧղՄೳͳํ๏ͰใΛఏࣔ͠ɼ ਓؒͷ࠷ऴܾఆΛࢧԉ͢ΔϑϨʔϜϫʔΫ.P.ΛఏҊ ݁ՌɿͭͷλεΫΛ࣮ࢪɽ ਓؒͷҙࢥܾఆʹର্ͯͤ͠͞Δ͜ͱΛࣔͨ͠
͍··Ͱͷҙࢥܾఆࢧԉɿग़ྗΛӏವΈ͠ͳ͍ ػցͷग़ྗ Λਓ͕ؒड͚औΓɼߦಈ ΛͱΔ ̂ y a fθ ̂
y x a
͍··Ͱͷख๏ɿΑ͍ग़ྗΛֶशʢ܇࿅࣌ʣ ڭࢣ͋ΓֶशͰྑ͍ग़ྗʹͳΔΑ͏ʹֶश fθ ̂ y x y L(y, ̂
y)
ݱ࣮ɿਪ݁ՌͱߦಈҰக͠ͳ͍ ਓ͕ؒؒʹೖΔͷͰ ʹͳΔ͜ͱ͕͋Δ ҙࢥܾఆͷঢ়گʹΑͬͯػցͱਓؒͷҙࢥܾఆ͕Ұக͠ͳ͍ ̂ y ≠ a fθ
̂ y x a ≠ ̂ y
ཧɿਓؒͷҙࢥܾఆΛ࠷దԽ ਓؒͷߦಈΛؚΊͯ࠷దԽ͢Δ͜ͱ͕ཧ ͕ͨͩͷग़ྗͰਓؒʹͱͬͯཧղ͍͠ ̂ y y fθ ̂ y
x a = h(x, ̂ y) L(y, a)
ՄࢹԽ ఏҊख๏ɿਓؒͰཧղͰ͖Δํ๏Ͱࢧԉ͍ͨ͠ ਓ͕͍ؒΔͨΊɼޯΛͰ͖ͳ͍ y දݱֶश x ϕ a =
h(z, ̂ y) L(y, a) z
ਓؒͷཧϞσϧΛઃఆͯ͠ɼޯΛՄೳʹʂ y x ϕ a = h(z, x) L(y,
a) h z
දݱΛͬͨదͳՄࢹԽ͍͠ දݱ ͔Βਓ͕ؒཧղͰ͖ΔΑ͏ʹ͢ΔͨΊʹɼ ՄࢹԽͷૢ࡞͕ඞཁ • άϥϑɼը૾ɼϋΠϥΠτͱ͔ʜ ҙࢥܾఆʹରͯ͠ɼྑ͍հೖ͕ߦ͑ΔΑ͏ͳՄࢹԽͦΕ͚ͩͰݚڀʹ ͳΔ͘Β͍͍͠
ຊݚڀͰɼྑͦ͞͏ͳՄࢹԽΛબΜͰར༻͢Δ z
࣮ݧɿͭͷλεΫͰ༗ޮੑΛࣔ͢ λεΫɿߴ࣍ݩσʔλΛ࣍ݩʹѹॖͨ͠ͱ͖ʹ ༗༻ͳදݱΛ֫ಘͰ͖Δ͔ʁ λεΫɿ࣮ࡍͷλεΫͰ ɹɹɹɹ֫ಘͨ͠දݱʹΑΔࢧԉ༗ޮ͔ʁ λεΫɿػց͕Γಘͳ͍Ճใ
ɹɹɹɹදݱͱͯ֫͠ಘͰ͖Δ͔ʁ
λεΫɿྑ͍දݱΛ֫ಘͰ͖Δ͔ʁ ࣍ݩѹॖΛͯ͠σʔλͷՄࢹԽ͕ՄೳʹͳΔ • طଘͷ࣍ݩѹॖ౷ܭతͳ࠷దԽ͔͠ߟ͍͑ͯͳ͍ ߴ࣍ݩͳσʔλΛਓతʹ࡞ΓɼྨʹऔΓΉ • ަࣹӨͨ࣌͠ʹɼz9zͱz0zͷܗʹฒͿΑ͏ʹ࡞
λεΫɿදݱϞσϧͱཧϞσϧ දݱϞσϧ • Yͷઢܕࣸ૾ߦྻ ཧϞσϧ • ҰͰYͷΈࠐΈωοτϫʔΫ • ਓؒͷࢹ֮Λ͓͓·͔ʹ࠶ݱ͢Δ
λεΫɿ݁Ռ "DDVSBDZˠ ͷ࣮ݧࢀՃऀ͕ਫ਼Λୡ
λεΫɿϦΞϧͳσʔλͰࢧԉͰ͖Δ͔ ϩʔϯ৹ࠪΛࡐʹͨ͠λεΫ • ࡁͨ͠ɼ ೲɼ ঝೝɼ ڋ൱ • ଛࣦؔɿ
.5VSLͰࢀՃऀΛूΊܭճΛूΊͨ ඪɿදݱʹΑΔΞυόΠεͰ ҙࢥܾఆͷࢧԉ͕Մೳ͔Ͳ͏͔ y = 1 y = 0 a = 1 a = 0 l(y, a) = 1y≠a
λεΫɿإදʹΑͬͯදݱΛՄࢹԽ 'BDJBMBWBUBSΛͬͯɼإදͰՄࢹԽΛߦ͏ • $IFSOP ff GBDFTإύʔπʹͦΕͧΕ͕ରԠͯ͠มԽ
λεΫɿදݱϞσϧͱཧϞσϧ දݱϞσϧ • શ݁߹ͷϢχοτΛ ཧϞσϧ • શ݁߹ͷϢχοτΛ
λεΫɿతͳΞυόΠεͱಉ /PBEWJDFΑΓ্ ༧ଌΞυόΠεͱಉ
λεΫɿਓ͔ؒ͠Γಘͳ͍ใ֫ಘͰ͖Δ͔ ਓ͔ؒ͠Γಘͳ͍ใʢࣝɼৗࣝͱ͔ʣΛදݱͱͯ֫͠ಘ Ͱ͖Δ͔ΛΈ͍ͨ ҩྍஅͷλεΫΛઃܭ • Ճใ Λઃఆ͢Δˠਓ͔ؒ͠ݟΕͳ͍ • ͜ͷՃใग़ྗʹӨڹΛٴ΅͢ઃఆ
σʔληοτࣗମਓతʹੜ s
λεΫɿදݱϞσϧͱཧϞσϧ දݱϞσϧ • ॏΈͱಛྔͷઢܗ ཧϞσϧ • ॏΈͱಛྔͷઢܗ Ճใͷઢܗ
λεΫɿ • ७ਮʹ܇࿅Λߦ͏߹ .P. • දݱҰॹʹ܇࿅͢Δ .P.Ҏ֎ͰɼՃใʹ Αͬͯաֶश͕ى͖͍͢
h(.BDIJOF)
·ͱΊͱײ ·ͱΊ ਓ͔ؒΒਓؒͷͨΊʹҙࢥܾఆΛࢧԉ͢Δ ϑϨʔϜϫʔΫΛఏҊ ਓؒͷ࠷ऴܾఆΛࢧԉͰ͖ΔΑ͏ʹɼදݱΛֶश͠ՄࢹԽ