Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
Pathologies of Neural Models Make Interpretatio...
Search
Yasufumi Taniguchi
December 09, 2018
Research
1.8k
1
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Pathologies of Neural Models Make Interpretations Difficult
Yasufumi Taniguchi
December 09, 2018
More Decks by Yasufumi Taniguchi
See All by Yasufumi Taniguchi
AllenNLPを使った開発
yasufumy
0
2.3k
Making Neural QA as Simple as Possible but not Simpler
yasufumy
0
110
Other Decks in Research
See All in Research
VLMの推論を高速化する視覚トークン削減の仕組み
tattaka
2
310
Language and AI
ayaniwa
0
270
Spatial Active Noise Control Based onSound Field Interpolation Incorporating Physical Constraints
skoyamalab
0
190
SAM3を用いたコマ・吹き出しの 領域検出と分割構造からの読み順推定
kzmssk
0
130
【ローカルAIに向き合う展示会vol.2】液体時間定数型モジュールを用いた オリジナルの双方向エンコーダーモデルNexteraBERT 推論速度向上検討並びにダウンストリーム評価
rikkabotan7
0
200
[最先端NLP勉強会2026] Checklists Are Better Than Reward Models For Aligning Language Models
nzw0301
1
310
秋葉原ウォーカブル基礎調査報告書
izumiyama_lab
1
120
20260624 NLP colloquium: 単一のhubテキストがCLIPを壊す:hubnessによる埋め込みの脆弱性特定
de9uch1
2
260
OWASP AISVS - C7
shiell
2
780
某助成金プロジェクト採択に向けて企業研究所のアウトリーチ専任者がやったこと
afroscript
0
180
Research Engineerという仕事 / Research Engineering: Bridging Research and Business
chck
1
340
[BlackHatAsia2026] Hidden Telemetry: Uncovering TraceLogging ETW Providers You're Not Using (Yet)
asuna_jp
1
720
Featured
See All Featured
Heart Work Chapter 1 - Part 1
lfama
PRO
10
36k
Build The Right Thing And Hit Your Dates
maggiecrowley
39
3.4k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
287
14k
What’s in a name? Adding method to the madness
productmarketing
PRO
24
4.2k
Rails Girls Zürich Keynote
gr2m
96
14k
What the history of the web can teach us about the future of AI
inesmontani
PRO
1
690
The Curious Case for Waylosing
cassininazir
1
510
Agile Actions for Facilitating Distributed Teams - ADO2019
mkilby
0
280
Impact Scores and Hybrid Strategies: The future of link building
tamaranovitovic
0
440
WCS-LA-2024
lcolladotor
0
830
Building AI with AI
inesmontani
PRO
1
1.2k
We Analyzed 250 Million AI Search Results: Here's What I Found
joshbly
1
1.9k
Transcript
ൃදऀ ୩ޱହ࢙ ҟৗͳڍಈ
!2 Pathological behavior ࣭จ͕did͚ͩͰ Ϟσϧͷग़ྗಉ͡ ֬ߴ͍
֓ཁ w NLPʹ͓͚ΔχϡʔϥϧϞσϧͷղੳख๏ΛఏҊ w Ϟσϧ͕λεΫΛղ্͘Ͱॏཁͳ୯ޠΛநग़͢Δख๏ w நग़͞Εͨ୯ޠਓʹͱͬͯҙຯෆ໌ w ҰํͰϞσϧநग़୯ޠͰਖ਼͘͠༧ଌ(Pathology) w
ղੳ݁Ռʹجͮ͘ਖ਼ଇԽ߲ΛఏҊ w ਖ਼ଇԽ߲ʹΑͬͯϞσϧͷղऍੑ্ !3
࣍ Ϟσϧղੳͷطଘख๏ ఏҊख๏ ࣮ݧ ·ͱΊ !4
Ϟσϧղੳͷطଘख๏
Ϟσϧղੳͷطଘख๏ !6 Adversarial Example Ϟσϧʹਓͷײʹ͢ΔڍಈΛͤ͞Δαϯϓϧ NLPͷλεΫ ओʹQAλεΫ Ͱύλʔϯ ਓʹͱͬͯҙຯͷͳ͍มߋ͕ɺϞσϧͷग़ྗΛܹมͤ͞Δέʔε
ਓʹͱͬͯ໌Β͔ͳมߋͰɺϞσϧ͕ग़ྗΛม͑ͳ͍έʔε
ग़ྗ͕ܹม͢Δέʔε !7 Jia et al., 2017 ΫΥʔλʔόοΫͷྸʹ͍ͭͯͷ จॻʹΫΥʔλʔόοΫͷഎ൪߸ʹ ؔ͢ΔจΛՃ Ϟσϧޡ
ग़ྗΛม͑ͳ͍έʔε !8 Mudrakarta et al., 2018 ݐͷന͍ϨϯΨ͕ରশ͔ʁ spherical (ٿঢ়ͷ) ݐͷന͍ϨϯΨ͕ٿঢ়͔ʁ
࣭จͷҙຯมԽ Ϟσϧͷ༧ଌෆม
2. ఏҊख๏
*OQVU3FEVDUJPO • ॏཁͰͳ͍୯ޠΛೖྗ͔ΒΓɺϞσϧͷڍಈΛੳ • Ϟσϧ͕ਖ਼͍͠ग़ྗΛ͢ΔͨΊʹඞཁͳ࠷୯ޠ (ॏཁ ୯ޠ) • Adversarial ExampleϞσϧʹͱͬͯͷॏཁ୯ޠʹண
*OQVU3FEVDUJPO !11 x y Ϟσϧͷ༧ଌ f( ⋅ ) χϡʔϥϧϞσϧ ೖྗܥྻ
(จจॻ) xi ೖྗܥྻͷ͋Δཁૉ (୯ޠ) g(xi |x) = f(y|x) − f(y|x−i ) ͋Δ୯ޠ ʹର͢Δ ॏཁΛఆٛ xi g i൪ͷ୯ޠΛফͨ͠ೖྗ
*OQVU3FEVDUJPO !12 g(xcontest |x) = f(y|x) − f(y|x−contest ) What
company won free advertisement due to QuickBooks contest ? What company won free advertisement due to QuickBooks contest ? g͕େ͖͚Εɺcontest͕ॏཁͳ୯ޠͱͳΔ Ϟσϧͷग़ྗʹେ͖͘د༩͍ͯ͠ΔͨΊ
*OQVU3FEVDUJPO !13 g(xi |x) = f(y|x) − f(y|x−i ) ॏཁͷ͍୯ޠΛআ
y͕มԽ͠ͳ͍Α͏ʹɺg͕࠷খͱͳΔ୯ޠiΛআ ͍ͯ͘͠
3. ࣮ݧ
ղੳͷରλεΫ 1. SQuAD w จॻͱ࣭จ͕༩͑ΒΕΔˠ࣭จʹରͯ͠Input Reduction w จॻ͔ΒղΛநग़͢ΔλεΫ 2. SNLI
w จ͕༩͑ΒΕΔˠͭͷจʹରͯ͠Input Reduction w จͷؔΛਪఆ͢ΔλεΫ 3. VQA w ը૾ͱ࣭จ͕༩͑ΒΕΔˠ࣭จʹରͯ͠Input Reduction w ղΛੜ͢ΔλεΫ !15
࣮ݧ༰ Input Reduction w Ϟσϧ͕ਖ਼͍͠ग़ྗΛ͢ΔαϯϓϧΛରʹ࣮ݧ w Input ReductionΛద༻ͨ͠ೖྗ(Reduced)ʹର͢ΔਓखධՁ w ReducedͱϥϯμϜʹ୯ޠΛམͱͨ͠߹(Random)ͷࠩҟͷධՁ
Regularization on Reduced Inputs w Input ReductionʹΑΔϞσϧͷPathological behaviorΛܰݮ͢Δਖ਼ଇԽ߲ ޙड़ ͷಋೖ !16
Reducedʹର͢ΔਓखධՁ !17 Reducedʹରͯ͠ ਓਖ਼͍͠༧ଌΛͰ ͖ͳ͍ w Reducedʹର͢Δਓͷਖ਼ w Ϟσϧͷਖ਼͕ͷαϯϓϧΛ༻
Reducedʹର͢ΔਓखධՁ !18 w ReducedͱRandomͷͲͪΒ͕ࣗવͳจ͔ w vs. Randomfifty-fiftyͱׂ͑ͨ߹ Reducedਓʹͱͬ ͯRandomͱಉ͡
Reducedͷࣄྫ !19 ʮͲ͜Ͱ࿅शͨ͠ ͔ʯΛฉ͔Ε͍ͯ ΔͷΘ͔Δ͕ɺ ʮͲͷνʔϜʯ͔ Θ͔Βͳ͍
Reducedͷฏۉ୯ޠ ͭͷλεΫͱɺ ਖ਼͢Δͷʹඞཁͳ୯ޠฏۉd
Reducedʹର͢ΔϞσϧͷ֬ !21 • Input Reductionͷద༻લޙͰϞσϧͷ ֬ʹมԽ΄ͱΜͲͳ͍ • ϞσϧӶ͍ϐʔΫΛ࣋ͭΑ͏ͳ Λֶश͍ͯ͠Δ͜ͱ͕ݪҼ
ਖ਼ଇԽ߲ͷಋೖ !22 ∑ (x,y)∈(X,Y) log(f(y|x)) + λ∑ ¯ x∈ ¯
X H(f(y| ¯ x)) Reducedʹରͯ͠ਖ਼͍͠yΛ ग़ྗ͠ʹ͘͘͢Δ ௨ৗͷతؔ Reducedαϯϓϧ௨ৗͷతؔΛֶͬͯशͨ͠ ϞσϧΛ༻͍ͯੜ
ਖ਼ଇԽ߲ͷޮՌ !23 • Ϟσϧͷਫ਼͕ඍ૿ • ਖ਼ʹඞཁͳ୯ޠ ͕૿Ճ
ਖ਼ଇԽ߲ͷޮՌ !24 ਓखධՁͷਫ਼্ Input Reductionͨ͠ೖྗ ͷղऍੑ্͕
ਖ਼ଇԽͨ͠Ϟσϧͷࣄྫ !25 Input Reductionͨ͠ೖྗ͕ਓͰ ղऍՄೳʹͳͬͨ
·ͱΊ ఏҊख๏ w NLPͷχϡʔϥϧϞσϧղੳख๏ͱͯ͠ɺInput ReductionΛఏҊ w ༧ଌʹد༩͠ͳ͍୯ޠΛೖྗ͔ΒΓɺϞσϧΛղੳ ࣮ݧ݁Ռ w ఏҊख๏Λద༻ͨ͠ೖྗਓʹͱͬͯҙຯෆ໌
w ҰํͰχϡʔϥϧϞσϧਖ਼͍͠༧ଌΛߦ͏ w ਖ਼ଇԽ߲Λಋೖ͢ΔͱϞσϧͷڍಈվળ !26