[読み会]Teaching Categories to Human Learners with Visual Explanations

Slide 1

Slide 1 text

Teaching Categories to Human Learners with Visual Explanations ಡΈձ@2021/05/18 ༶໌఩

Slide 2

Slide 2 text

•ը૾෼ྨͰػցڭࣔΛߦ͏ͱ͖ʹɼͲ͜ݟΔ΂͖͔ͷઆ໌Λ଍ ͯ͠ਓؒͷύϑΥʔϚϯεΛ͋͛ͨΑʂ ͻͱ͜ͱͰ͍͏ͱ ػցڭࣔ × ը૾

Slide 3

Slide 3 text

•ஶऀ: •Oisin Mac Aodha, Shihan Su, Yuxin Chen, Pietro Person, Yisong Yue •California Institute of Technology •ग़య: CVPR 2018 •ͳΜͰಡΜ͔ͩ?: ࠷৽ͷػցڭࣔΛΩϟονΞοϓ͍͔ͨ͠Β ࿦จ৘ใ

Slide 4

Slide 4 text

•ܭࢉػ͕ิॿ͢Δڭҭ → ݸਓ͝ͱʹಛԽͰ͖ΔΑ͏ʹͳ͍ͬͯ Δɽ •਺ֶ΍ݴޠڭҭͰ͸ɼશࣗಈͰߦ͑ΔΑ͏ʹͳΓͭͭ͋Δ •͔͠͠ɼઐ໳తͳ࿩(ҩֶͱ͔)Ͱ͸ະͩͰ͖ͳ͍ •υϝΠϯ஌ࣝΛڭ͑Δͷ͕೉͍͠ Πϯτϩ

Slide 5

Slide 5 text

•Ϋϥ΢υιʔγϯάͷϫʔΧΛڭҭ͢Δ͜ͱ͕ඞཁ •ઐ໳ՈΛ࢖͑Δͷʹίετ͕͔͔Γ੍͔ͭݶ͕͋Δ •ڭҭͰ͖ͨΒߴ඼࣭ͳσʔληοτ͕࡞ΕΔ •΄͔ͷυϝΠϯʹରͯ͠΋ਓؒͷ൚Խྗ͕ద༻Ͱ͖Δ͔΋? Πϯτϩ

Slide 6

Slide 6 text

•୯७ʹਖ਼ղϥϕϧͱαϯϓϧΛฦ͢ •͚Ͳ͜ΕͰຊ౰ʹ͍͍ͷʁ܇࿅Ͱ͖ͯΔͷʁʁ •આ໌Λ෇༩ֶͯ͠शޮՌΛߴΊΔ Πϯτϩ ͜Ε·Ͱͷػցڭࣔ͸Ͳ͏ͳͷʁ

Slide 7

Slide 7 text

ఏҊ: Interpretable Visual Teaching ༻ޠͷఆٛΛ͍ͯ͘͠Α : ೖྗը૾ X : ਖ਼ղϥϕϧ Y :Ծઆू߹ ൑அج४ͷू߹ H

Slide 8

Slide 8 text

•Ծઆ: ֶशࡁΈϞσϧͦͷ΋ͷɽೖྗۭ͔ؒΒग़ྗू߹΁ͷؔ਺ •Ծઆू߹: Ծઆ͕ू·͍ͬͯΔ΋ͷ. MLΞϧΰϦζϜͰ࡞ΒΕΔ Մೳੑͷ͋ΔϞσϧͷू·Γ •Ծઆू߹ͷதʹ͋ΔਅͷԾઆ ʹ͚͍ۙͮͯ͘ͷ͕໨త h⋆ ఏҊख๏ ͪΐͬͱৄ͘͠

Slide 9

Slide 9 text

Title Text

Slide 10

Slide 10 text

• ͳը૾ू߹ ʹରֶͯ͠शऀͷԾઆ ͸มԽ͢Δ Ծઆ ͷࣄޙ෼෍: ਪ࿦࣌: 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: Կ΋͠ͳͱ͖ͷ ճ౴ʹର͢Δ ֬৴౓

Slide 11

Slide 11 text

•ߋ৽ࣜ͸࣍ͷΑ͏ʹม͑Δ •৽͘͠2ͭͷݮਰ߲Λ௥Ճ͢Δ 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))

Slide 12

Slide 12 text

•આ໌ͷ࣭͸ը૾ͷ೉͠͞ͱಉ͡Α͏ʹଌΕͳ͍ɽ •ը૾ͷ೉қ౓͸൑ఆڥքͱͷڑ཭ͰܭࢉͰ͖Δ •ࣗಈੜ੒͢Δํ๏͸͋ͱͰग़ͯ͘ΔΑ ఏҊख๏ EXPLAINΞϧΰϦζϜ: Modeling Explanations E (e t) = 1 1 + exp ( −β diﬀ (et)) ը૾ ʹର͢Δ༩͑ΒΕͨ આ໌ ͷ೉͠͞ x t e t

Slide 13

Slide 13 text

•αϯϓϧtͷઆ໌ Λ࡞Γ͍ͨ •Ϋϥ΢υιʔγϯάͱ͔ઐ໳Ոͱ͔ʹ΍ͬͯ΋Β͏ͱ͔͋Δ͚ͲࣗಈͰ࡞ ΕΔͱΑ͘ͳ͍ʁ •CNNͷClass Activation MappingʹΑͬͯࣗಈͰઆ໌Λ࡞Δ e t ఏҊख๏ EXPLAINΞϧΰϦζϜ: ࣗಈੜ੒ e(j) = ∑ k wk c fk j (x) + b c

Slide 14

Slide 14 text

•͖ͬ͞ఆٛͨ͠ը૾ͷઆ໌౓߹͍͔Βɼը૾ͷ೉қ౓Λఆٛ •ࣗಈͰઆ໌Λੜ੒͢Δ࣌ͷϞσϧͱͯ͠ResNetϕʔεͷϞσϧ Λར༻ ఏҊख๏ EXPLAINΞϧΰϦζϜ: ը૾ͷઆ໌ੑˠ೉қ౓ͷఆٛ diﬀ(e) = − 1 J ∑ j e(j)log(e(j))

Slide 15

Slide 15 text

•ैདྷ͸ᩦཉʹޡࠩ࠷খΛ࠷దԽ → ඞͣ͠΋༗ӹͰ͸ͳ͍ •ೳಈֶशʹώϯτΛಘͯɼΫϥεͷ୅දྫΛఏࣔ • ʹͳΔͱSTRICTͱಉ͡ʹͳΔ β, γ → ∞ ఏҊख๏ EXPLAINΞϧΰϦζϜ: Modeling Representativeness D (x t) = 1 1 + exp ( −γ dist (xt)) ଞͷը૾ͱൺ΂ͯ ͲΕ͘Β͍཭Ε͍ͯΔ͔ dist (x t) = 1 N N ∑ n=1 x t − x n 2 2

Slide 16

Slide 16 text

•ڭࡐू߹ ͰͳʹΛબ୒͢Δ͔ → ֶ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) | | .

Slide 17

Slide 17 text

•ޡࠩͷظ଴஋͕Ұ൪େ͖ܰ͘ݮͰ͖ΔΑ͏ͳू߹Λબ୒ •͜ͷRΛ࠷େʹ͢ΔΑ͏ͳू߹T͕ཉ͍͠ڭࡐू߹ •͔͠͠ɼ௚઀ٻΊΔͷ͸ྼϞδϡϥੑ͔Βࠔ೉ •ྑ͍αϯϓϧΛ1ͭͣͭ௥Ճ͍ͯ͘͠ 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})

Slide 18

Slide 18 text

•3ͭͷσʔληοτΛ༻͍ͯ༗ޮੑΛ֬ೝ͍ͯ͘͠ɽ 1. Butterﬂies (௏ࣝผ) 2. OCT Eyes (໢ບ਍அ) 3. Chinese Characters (จࣈࣝผ) ࣮ݧ σʔληοτ

Slide 19

Slide 19 text

•Amazon Mechanical TurkͰඃݧऀ40ਓ •໛ࢼը૾͸ϥϯμϜʹఏࣔͯ͠ɼબ୒ճ౴ͷॱ൪΋ϥϯμϜʹ •ҐஔʹΑΔόΠΞεΛͳ͍ͯ͘͠Δɽ •ର߅ख๏ •RAND_IM: ϥϯμϜʹը૾ͱਖ਼ղϥϕϧ •RAND_EXP: ϥϯμϜը૾ͱͦͷઆ໌ •STRICT: ͍͍ײ͡ͷը૾Λબ୒͢Δ ࣮ݧઃఆ ͪΐͬͱৄࡉʹ

Slide 20

Slide 20 text

࣮ݧઃఆ ػցڭࣔͷྲྀΕ

Slide 21

Slide 21 text

࣮ݧ݁Ռ Butterﬂy ਖ਼౴཰ͷώετάϥϜ ͕ӈʹγϑτ͍ͯ͠Δ ͜ͷσʔληοτ͸೉͍͠܏޲ ࣅͨ3छ͸ࠞཚ͕ͪ͠

Slide 22

Slide 22 text

࣮ݧ આ໌ը૾ͷΠϝʔδ: Butterﬂy

Slide 23

Slide 23 text

࣮ݧ݁Ռ OCT Eyes ਖ਼౴཰ͷώετάϥϜ ͕ӈʹγϑτ͍ͯ͠Δ ϥϯμϜͰ΋޲্ ͯ͠͠·ͬͯΔ

Slide 24

Slide 24 text

࣮ݧ આ໌ը૾ͷΠϝʔδ: OCT Eyes

Slide 25

Slide 25 text

࣮ݧ݁Ռ Chinese Character CNNͷઆ໌͕ࣦഊ͠ ͍ͯΔ खಈͷઆ໌͕ ੑೳྑ͍

Slide 26

Slide 26 text

࣮ݧ આ໌ը૾ͷΠϝʔδ: Chinese Character

Slide 27

Slide 27 text

•ࢹ֮తઆ໌ੑΛը૾ʹ༩͑ͯͦΕΛ΋ͱʹػցڭࣔΛߦͳͬͯ ͍͘ɽ •ैདྷͷਖ਼ղϥϕϧ͚ͩڭ͑Δํ๏ΑΓɼઆ໌͕͋Δํֶ͕शޮ Ռ͕ߴ͘ɼ͞ΒʹޮՌͷߴ͍ڭࡐू߹Λݟ͚ͭΕ͍ͯΔɽ •কདྷతʹΦϯϥΠϯͰΠϯλϥΫςΟϒʹ΍Γ͍ͨΑͶ ·ͱΊ ը૾આ໌෇͖ػցڭࣔ