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
iclr2020deepsemi-supervisedanomalydetectionyama...
Search
Yamato.OKAMOTO
June 14, 2020
Technology
180
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
iclr2020deepsemi-supervisedanomalydetectionyamatookamoto-200531022507.pdf
Yamato.OKAMOTO
June 14, 2020
More Decks by Yamato.OKAMOTO
See All by Yamato.OKAMOTO
いまAI組織が求める企画開発エンジニアとは?
roadroller
2
1.6k
Slide ICCV2023 Constructing Image Text Pair Dataset from Books
roadroller
0
170
第11回 全日本コンピュータビジョン勉強会 CVPR2022 "A Self-Supervised Descriptor for Image Copy Detection"
roadroller
0
710
第9回 全日本コンピュータビジョン勉強会 発表資料
roadroller
0
690
第七回全日本コンピュータビジョン勉強会 A Multiplexed Network for End-to-End, Multilingual OCR
roadroller
1
1k
部下のマネジメントはAI開発に学べ
roadroller
0
200
Domain Generalization via Model-Agnostic Learning of Semantic Features NeurIPS’19 読み会 in 京都
roadroller
0
330
ICML’2019 読み会in京都 Federated Learningの研究動向
roadroller
0
150
CVPR2019@Long Beach 参加速報(本会議)
roadroller
0
170
Other Decks in Technology
See All in Technology
Driving AI Adoption Using In-House GPUs to Serve Qwen
po3rin
2
460
Microsoft 365 Copilot chat -tekoälypalvelun tietosuojaongelmat
hponka
0
430
AI時代のAPI品質を支えるガードレール / API Guardrails for API quality in the AI era
yokawasa
1
120
現場に行くだけでは足りない——プロダクトエンジニアが業務の流れを捉える観点と、その鍛え方
takumiengineering
0
300
module Synths; end
asonas
1
120
振り返りこそエンジニアの本領
negima
0
310
Genie Code ワークショップ 基礎編 / Genie-Code-Workshop-fundamental
databricksjapan
PRO
0
360
Bet AI Day 2026丨バクラク Autopilot、業務システムの再設計
layerx
PRO
2
1.5k
Bet AI Day 2026丨Production-Ready AI Agents — エンタープライズの実務を任せるための設計と運用
layerx
PRO
3
2.4k
Jetpack Compose で挑む新聞紙面UI ─ 複合ジェスチャー・ポリゴン記事領域・適応的ページ構成という3つの壁/droidkaigi2026
nikkei_engineer_recruiting
0
230
AIに丸投げしないトイル削減 / Eliminating Toil Without Leaving It All to AI
kohbis
2
610
PfEingのアプローチで働こう
rindrics
0
180
Featured
See All Featured
Why Mistakes Are the Best Teachers: Turning Failure into a Pathway for Growth
auna
0
250
HDC tutorial
michielstock
2
850
Beyond borders and beyond the search box: How to win the global "messy middle" with AI-driven SEO
davidcarrasco
3
230
The Power of CSS Pseudo Elements
geoffreycrofte
82
6.5k
Leading Effective Engineering Teams in the AI Era
addyosmani
9
2.5k
DevOps and Value Stream Thinking: Enabling flow, efficiency and business value
helenjbeal
1
380
Marketing to machines
jonoalderson
1
5.7k
Lessons Learnt from Crawling 1000+ Websites
charlesmeaden
PRO
1
1.6k
Bridging the Design Gap: How Collaborative Modelling removes blockers to flow between stakeholders and teams @FastFlow conf
baasie
0
690
Between Models and Reality
mayunak
4
450
Speed Design
sergeychernyshev
33
2.1k
Getting science done with accelerated Python computing platforms
jacobtomlinson
2
470
Transcript
2020/6/14 Yamato OKAMOTO ICLRΦϯϥΠϯಡΈձ Deep Semi-supervised Anomaly Detection
ࣗݾհʢ͘!!ʣ ɹԬຊେʢ͓͔ͱ·ͱʣ • ژେֶඒೱݚڀࣨͰύλʔϯೝࣝΛݚڀͯ͠म࢜՝ఔमྃ • ΦϜϩϯͰ৽نࣄۀΛܦݧޙɺ͍·ࣾձγεςϜࣄۀ෦ͷݚڀॴϦʔμʔ • ເژΛϙετɾγϦίϯόϨʔʹ͢Δ͜ͱɺؔͷίϛϡχςΟΛڧԽ͍ͨ͠ ɹ㱺 ژͷมਓύϫʔΛੈքʹΒ͠Ί͍ͨ
Twitter : RoadRoller_DESU ҆৺҆શͳࣾձͷ࣮ݱʹ͚ͯɺ ࠷ۙ Anomaly Detection ʹڵຯΞϦ
Anomaly Detection ͋Δ͋Δ ఆٛࠔ • ҟৗʹ༷ʑͳόϦΤʔγϣϯ͕͋Δ • ҟৗݕग़͍͚ͨ͠ͲʮWhat is ҟৗʁʯ͕ఆٛͰ͖ͳ͍
ֶशσʔλ͕ೖखࠔ • ҟৗ໓ଟʹൃੜ͠ͳ͍ʢ※ සൟʹൃੜ͢ΔΠϕϯτҟৗ͡Όͳ༷ͯ͘ʣ • ѹతʹҟৗσʔλ͕ෆͯ͠ػցֶश͕ࠔ ैདྷख๏ɿਖ਼ৗΛఆٛ͢Δ • ʮWhat is ҟৗʁʯͷఆٛΛఘΊΔɺҟৗσʔλͷֶशఘΊΔ • ͦͷΘΓʮWhat is ਖ਼ৗʁʯͷఆٛΛֶशͯ͠ɺʮNot ਖ਼ৗʯΛҟৗͱఆ͢Δ
Anomaly Detection ͷैདྷݚڀ Deep One-Class Classification (ICML’18) • ਖ਼ৗσʔλͷΈΛ༻͍ͯɺClassifierͳΓAutoEncoderͳΓΛैདྷ௨Γʹֶश •
͜ͷͱ͖ɺಛྔ͕࣍ݩ෦ۭؒʹऩଋ͢ΔΑ͏LOSSΛՃ͑Δ • ਖ਼ৗσʔλͳΒٿʹ͢ΔͣͳͷͰɺٿ͔Β֎ΕͨσʔλΛҟৗͱఆ͢Δ ୈҰ߲ʹΑͬͯٿʹ͕ԡ͠ࠐ·ΕΔ cɿ ٿͷத৺ʢͨͩ͠≠0ʣ nɿֶश͢Δਖ਼ৗσʔλͷ
Anomaly Detection ͷධՁ؍ ͲΕ͚ͩਖ਼֬ʹҟৗΛݕͰ͖͔ͨʁ • ਖ਼ৗσʔλΛਖ਼ৗͱఆͯ͠ɺҟৗσʔλΛҟৗͱఆ͢Δਫ਼ ԼྲྀλεΫΛअຐ͠ͳ͍͔ʁ • ԼྲྀλεΫ͕͋Δ߹ɺҟৗݕػೳͷՃʹΑͬͯѱӨڹ͕ͳ͍͔Ͳ͏͔ •
ྫ͑ɺ10ΫϥεͷࣈࣝผثʹɺਤܗͳͲࣈҎ֎͕ೖྗ͞Εͨͱ͖ҟৗͱఆ͢Δػ ೳΛ͚Ճ͍͑ͨͤͰɺैདྷͷ10Ϋϥεࣝผੑೳ͕Լ͢ΔͱࠔΔ ad-hoc͔post-hoc͔ʁ • ҟৗݕ͢ΔͨΊʹϞσϧߏֶशํ๏·Ͱม͑Δඞཁ͕͋Δ͔ʁ • ·ͨɺLOSSΛޙ͔Β͚͚̍ͭͩͯ͠Ճֶश͢Δ͚ͩͰOK͔ʁ • ͲͪΒ͕ྑ͍ѱ͍ͳͲҰ֓ʹݴ͑ͳ͍͕ɺpost-hocͷํ͕ѻ͍͍͢ɻ
հจͷ֓ཁ ʮSemi-supervisedʹֶश͠Α͏ʂʯ Anomaly Detection ͷݚڀUnsupervised͕ओྲྀͷΑ͏ͩ Ͱɺֶश༻ͷҟৗσʔλ͕ೖखࠔͩͱͯ͠ɺ ӡ༻Λଓ͚ͯͨΒҟৗσʔλʹ͍ͣΕग़ձ͏ͣ ͳΒɺͦΕΒগྔͷҟৗσʔλΛͬͯɺ Semi-supervisedʹֶशͨ͠ํ͕ྑ͍ͷͰʁ ※Semi-supervisedͷAnomaly
Detectionݚڀඇৗʹগͳ͍
ఏҊख๏ ʮLOSSʹ߲Λ̍ͭՃ͠·ͨ͠ʯ Deep One-Class Classification (ICML’18) ͷLOSSʹSemi-supervisedͷ߲Λ̍ͭՃ • ࣮ಉ͡ஶऀͰͨ͠ɻࣗͷݚڀΛࣗͰΞοϓσʔτͨ͠ܗʹͳΔɻ ͠ҟৗσʔλʹग़ձͬͨΒɺ
ٿͷ֎ଆʹߦ͘Α͏ֶश͢Δ mɿsemi-supervisedʹֶश͢Δσʔλ yj ɿਖ਼ৗorҟৗͷϥϕϧ
࣮ݧ݁Ռ ॎ࣠ɿҟৗσʔλͷݕग़ੑೳ ʢHigher is Betterʣ Unsupervised Semi-supervised ԣ࣠ɿSemi-supervisedͰڭࢣ͖ͷҟৗσʔλΛֶशׂͨ͠߹ ఏҊख๏ MNISTɺFashion-MNISTɺCIFAR-10ͷσʔληοτͰධՁ
• ̍Ϋϥεͱਖ਼ৗͱఆٛͯ͠ɺAutoEncoderʴఏҊख๏ͰಛྔදݱΛֶश • Γͷ̕ΫϥεΛೖྗͨ͠ͱ͖ɺҟৗͱఆͰ͖Δ͔Ͳ͏͔ධՁ ੑೳվળΛ֬ೝ
·ͱΊͱߟ ਂֶशʹΑΔ Semi-supervised ͳ Anomaly Detection ख๏ΛఏҊ • ॳΊͯͰͳ͍ͱࢥ͏͕ɺਂֶशʹΑΔAnomaly DetectionͰsemi-supervised͍͠
• ͔ͨ͠ʹࣾձ࣮Λߟ͑Δͱɺ͜ͷઃఆద • ख๏γϯϓϧͰɺpost-hocͳͷͰѻ͍͍͢ • ࠓճԼྲྀλεΫ͕AE͕ͩͬͨɺClassificationͩͱͲ͏ͳΔ͔ʁ • Anomaly DetectionͷධՁσʔληοτͬͯଞʹͳ͍ͷ͔ͳɺɺɺɺ ʢ͍ͭ·ͰMNISTʹΑΔධՁ͕ଓ͘ͷͩΖ͏͔ʣ