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]Teaching Categories to Human Learners with...
Search
mei28
May 18, 2021
0
22
[読み会]Teaching Categories to Human Learners with Visual Explanations
読み会資料
Teaching Categories to Human Learners with Visual Explanations (CVPR 2018)
mei28
May 18, 2021
Tweet
Share
More Decks by mei28
See All by mei28
[読み会] “Are You Really Sure?” Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making
mei28
0
50
[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
1
400
[CHI'24] Fair Machine Guidance to Enhance Fair Decision Making in Biased People
mei28
0
37
[DEIM2024] 卓球の得点予測における重要要素の分析
mei28
0
27
[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
180
[読み会] Words are All You Need? Language as an Approximation for Human Similality Judgements
mei28
0
28
[参加報告] AAAI'23
mei28
0
81
[計算機構論] Learning Models of Individual Behavior in Chess
mei28
0
69
[計算機構論] Why do tree-based models still outperform deep learning on tabular data?
mei28
0
44
Featured
See All Featured
Build your cross-platform service in a week with App Engine
jlugia
229
18k
The Pragmatic Product Professional
lauravandoore
31
6.3k
No one is an island. Learnings from fostering a developers community.
thoeni
19
3k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
31
2.7k
Rebuilding a faster, lazier Slack
samanthasiow
79
8.7k
A Modern Web Designer's Workflow
chriscoyier
693
190k
How to Think Like a Performance Engineer
csswizardry
20
1.1k
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
159
15k
Performance Is Good for Brains [We Love Speed 2024]
tammyeverts
6
410
Optimising Largest Contentful Paint
csswizardry
33
2.9k
Testing 201, or: Great Expectations
jmmastey
38
7.1k
How to Ace a Technical Interview
jacobian
276
23k
Transcript
Teaching Categories to Human Learners with Visual Explanations ಡΈձ@2021/05/18 ༶໌
•ը૾ྨͰػցڭࣔΛߦ͏ͱ͖ʹɼͲ͜ݟΔ͖͔ͷઆ໌Λ ͯ͠ਓؒͷύϑΥʔϚϯεΛ͋͛ͨΑʂ ͻͱ͜ͱͰ͍͏ͱ ػցڭࣔ × ը૾
•ஶऀ: •Oisin Mac Aodha, Shihan Su, Yuxin Chen, Pietro
Person, Yisong Yue •California Institute of Technology •ग़య: CVPR 2018 •ͳΜͰಡΜ͔ͩ?: ࠷৽ͷػցڭࣔΛΩϟονΞοϓ͍͔ͨ͠Β จใ
•ܭࢉػ͕ิॿ͢Δڭҭ → ݸਓ͝ͱʹಛԽͰ͖ΔΑ͏ʹͳ͍ͬͯ Δɽ •ֶݴޠڭҭͰɼશࣗಈͰߦ͑ΔΑ͏ʹͳΓͭͭ͋Δ •͔͠͠ɼઐతͳ(ҩֶͱ͔)ͰະͩͰ͖ͳ͍ •υϝΠϯࣝΛڭ͑Δͷ͕͍͠ Πϯτϩ
•ΫϥυιʔγϯάͷϫʔΧΛڭҭ͢Δ͜ͱ͕ඞཁ •ઐՈΛ͑Δͷʹίετ͕͔͔Γ੍͔ͭݶ͕͋Δ •ڭҭͰ͖ͨΒߴ࣭ͳσʔληοτ͕࡞ΕΔ •΄͔ͷυϝΠϯʹରͯ͠ਓؒͷ൚Խྗ͕ద༻Ͱ͖Δ͔? Πϯτϩ
•୯७ʹਖ਼ղϥϕϧͱαϯϓϧΛฦ͢ •͚Ͳ͜ΕͰຊʹ͍͍ͷʁ܇࿅Ͱ͖ͯΔͷʁʁ •આ໌Λ༩ֶͯ͠शޮՌΛߴΊΔ Πϯτϩ ͜Ε·ͰͷػցڭࣔͲ͏ͳͷʁ
ఏҊ: Interpretable Visual Teaching ༻ޠͷఆٛΛ͍ͯ͘͠Α : ೖྗը૾ X :
ਖ਼ղϥϕϧ Y :Ծઆू߹ அج४ͷू߹ H
•Ծઆ: ֶशࡁΈϞσϧͦͷͷɽೖྗۭ͔ؒΒग़ྗू߹ͷؔ •Ծઆू߹: Ծઆ͕ू·͍ͬͯΔͷ. MLΞϧΰϦζϜͰ࡞ΒΕΔ Մೳੑͷ͋ΔϞσϧͷू·Γ •Ծઆू߹ͷதʹ͋ΔਅͷԾઆ ʹ͚͍ۙͮͯ͘ͷ͕త h⋆
ఏҊख๏ ͪΐͬͱৄ͘͠
Title Text
• ͳը૾ू߹ ʹରֶͯ͠शऀͷԾઆ มԽ͢Δ Ծઆ ͷࣄޙ: ਪ࣌: T ⊂
X T h h P(h ∣ T) ∝ P(h) ∏ xt ∈ T yt ≠ ̂ yh t P (y t ∣ h, x t) P (y t ∣ h, x t) = 1 1 + exp ( −αh (xt) yt) ఏҊख๏ STRICTΞϧΰϦζϜ: Կ͠ͳͱ͖ͷ ճʹର͢Δ ֬৴
•ߋ৽ࣜ࣍ͷΑ͏ʹม͑Δ •৽͘͠2ͭͷݮਰ߲ΛՃ͢Δ ఏҊख๏ EXPLAINΞϧΰϦζϜ: ϑΟʔυόοΫΛߟ͑Δͱ͖ P(h ∣ T) ∝
P(h) ∏ xt ∈ T yt ≠ ̂ yh t P (y t ∣ h, x t)∏ x t ∈T ( E (e t) D (x t))
•આ໌ͷ࣭ը૾ͷ͠͞ͱಉ͡Α͏ʹଌΕͳ͍ɽ •ը૾ͷқఆڥքͱͷڑͰܭࢉͰ͖Δ •ࣗಈੜ͢Δํ๏͋ͱͰग़ͯ͘ΔΑ ఏҊख๏ EXPLAINΞϧΰϦζϜ: Modeling Explanations E (e
t) = 1 1 + exp ( −β diff (et)) ը૾ ʹର͢Δ༩͑ΒΕͨ આ໌ ͷ͠͞ x t e t
•αϯϓϧtͷઆ໌ Λ࡞Γ͍ͨ •Ϋϥυιʔγϯάͱ͔ઐՈͱ͔ʹͬͯΒ͏ͱ͔͋Δ͚ͲࣗಈͰ࡞ ΕΔͱΑ͘ͳ͍ʁ •CNNͷClass Activation MappingʹΑͬͯࣗಈͰઆ໌Λ࡞Δ e t
ఏҊख๏ EXPLAINΞϧΰϦζϜ: ࣗಈੜ e(j) = ∑ k wk c fk j (x) + b c
•͖ͬ͞ఆٛͨ͠ը૾ͷઆ໌߹͍͔Βɼը૾ͷқΛఆٛ •ࣗಈͰઆ໌Λੜ͢Δ࣌ͷϞσϧͱͯ͠ResNetϕʔεͷϞσϧ Λར༻ ఏҊख๏ EXPLAINΞϧΰϦζϜ: ը૾ͷઆ໌ੑˠқͷఆٛ diff(e) = −
1 J ∑ j e(j)log(e(j))
•ैདྷᩦཉʹޡࠩ࠷খΛ࠷దԽ → ඞͣ͠༗ӹͰͳ͍ •ೳಈֶशʹώϯτΛಘͯɼΫϥεͷදྫΛఏࣔ • ʹͳΔͱSTRICTͱಉ͡ʹͳΔ β, γ →
∞ ఏҊख๏ EXPLAINΞϧΰϦζϜ: Modeling Representativeness D (x t) = 1 1 + exp ( −γ dist (xt)) ଞͷը૾ͱൺͯ ͲΕ͘Β͍Ε͍ͯΔ͔ dist (x t) = 1 N N ∑ n=1 x t − x n 2 2
•ڭࡐू߹ ͰͳʹΛબ͢Δ͔ → ֶशऀͷޡࠩΛݮΒ͍ͨ͠ •Ծઆ ʹରͯ͠ɼ؍ଌՄೳͳσʔλͱͷޡࠩΛ࣍ͷΑ͏ʹఆٛ T h ఏҊख๏
Teaching Algorithm: ͲͷαϯϓϧΛఏࣔ͢Δ͔ʁ err c (h) = x : ( ̂ yh ≠ y c ∧ y = y c) ∨ ( ̂ yh = y c ∧ y ≠ y c) | | .
•ޡࠩͷظ͕Ұ൪େ͖ܰ͘ݮͰ͖ΔΑ͏ͳू߹Λબ •͜ͷRΛ࠷େʹ͢ΔΑ͏ͳू߹T͕ཉ͍͠ڭࡐू߹ •͔͠͠ɼٻΊΔͷྼϞδϡϥੑ͔Βࠔ •ྑ͍αϯϓϧΛ1ͭͣͭՃ͍ͯ͘͠ ఏҊख๏ ڭࡐू߹ͷબ R(T) = 1
C ∑ c ( [err c (h)] − [err c (h) ∣ T]) = 1 C ∑ c∈ ∑ h∈ℋ (P c (h) − P c (h ∣ T)) err c (h) খ͘͞ͳΔ΄Ͳ خ͍͠ x t = argmax x R(T ∪ {x})
•3ͭͷσʔληοτΛ༻͍ͯ༗ޮੑΛ֬ೝ͍ͯ͘͠ɽ 1. Butterflies (ࣝผ) 2. OCT Eyes (ບஅ) 3.
Chinese Characters (จࣈࣝผ) ࣮ݧ σʔληοτ
•Amazon Mechanical TurkͰඃݧऀ40ਓ •ࢼը૾ϥϯμϜʹఏࣔͯ͠ɼબճͷॱ൪ϥϯμϜʹ •ҐஔʹΑΔόΠΞεΛͳ͍ͯ͘͠Δɽ •ର߅ख๏ •RAND_IM: ϥϯμϜʹը૾ͱਖ਼ղϥϕϧ •RAND_EXP:
ϥϯμϜը૾ͱͦͷઆ໌ •STRICT: ͍͍ײ͡ͷը૾Λબ͢Δ ࣮ݧઃఆ ͪΐͬͱৄࡉʹ
࣮ݧઃఆ ػցڭࣔͷྲྀΕ
࣮ݧ݁Ռ Butterfly ਖ਼ͷώετάϥϜ ͕ӈʹγϑτ͍ͯ͠Δ ͜ͷσʔληοτ͍͠ ࣅͨ3छࠞཚ͕ͪ͠
࣮ݧ આ໌ը૾ͷΠϝʔδ: Butterfly
࣮ݧ݁Ռ OCT Eyes ਖ਼ͷώετάϥϜ ͕ӈʹγϑτ͍ͯ͠Δ ϥϯμϜͰ্ ͯ͠͠·ͬͯΔ
࣮ݧ આ໌ը૾ͷΠϝʔδ: OCT Eyes
࣮ݧ݁Ռ Chinese Character CNNͷઆ໌͕ࣦഊ͠ ͍ͯΔ खಈͷઆ໌͕ ੑೳྑ͍
࣮ݧ આ໌ը૾ͷΠϝʔδ: Chinese Character
•ࢹ֮తઆ໌ੑΛը૾ʹ༩͑ͯͦΕΛͱʹػցڭࣔΛߦͳͬͯ ͍͘ɽ •ैདྷͷਖ਼ղϥϕϧ͚ͩڭ͑Δํ๏ΑΓɼઆ໌͕͋Δํֶ͕शޮ Ռ͕ߴ͘ɼ͞ΒʹޮՌͷߴ͍ڭࡐू߹Λݟ͚ͭΕ͍ͯΔɽ •কདྷతʹΦϯϥΠϯͰΠϯλϥΫςΟϒʹΓ͍ͨΑͶ ·ͱΊ ը૾આ໌͖ػցڭࣔ