iclr2020deepsemi-supervisedanomalydetectionyamatookamoto-200531022507.pdf

Transcript

1. 2020/6/14
 Yamato OKAMOTO ICLRΦϯϥΠϯಡΈձ Deep Semi-supervised Anomaly Detection

2. ࣗݾ঺հʢ୹͘!!ʣ ɹԬຊେ࿨ʢ͓͔΋ͱ΍·ͱʣ • ژ౎େֶඒೱݚڀࣨͰύλʔϯೝࣝΛݚڀͯ͠म࢜՝ఔमྃ • ΦϜϩϯͰ৽نࣄۀ૑଄Λܦݧޙɺ͍·͸ࣾձγεςϜࣄۀ෦ͷݚڀॴϦʔμʔ • ເ͸ژ౎ΛϙετɾγϦίϯόϨʔʹ͢Δ͜ͱɺؔ੢ͷίϛϡχςΟΛڧԽ͍ͨ͠
 ɹ㱺 ژ౎ͷมਓύϫʔΛੈքʹ஌Β͠Ί͍ͨ

   Twitter : RoadRoller_DESU ҆৺҆શͳࣾձͷ࣮ݱʹ޲͚ͯɺ ࠷ۙ͸ Anomaly Detection ʹڵຯΞϦ

3. Anomaly Detection ͋Δ͋Δ ఆٛࠔ೉ • ҟৗʹ͸༷ʑͳόϦΤʔγϣϯ͕͋Δ • ҟৗݕग़͍͚ͨ͠ͲʮWhat is ҟৗʁʯ͕ఆٛͰ͖ͳ͍

   ֶशσʔλ͕ೖखࠔ೉ • ҟৗ͸໓ଟʹൃੜ͠ͳ͍ʢ※ සൟʹൃੜ͢ΔΠϕϯτ͸ҟৗ͡Όͳͯ͘࢓༷ʣ • ѹ౗తʹҟৗσʔλ͕ෆ଍ͯ͠ػցֶश͕ࠔ೉ ैདྷख๏ɿਖ਼ৗΛఆٛ͢Δ • ʮWhat is ҟৗʁʯͷఆٛΛఘΊΔɺҟৗσʔλͷֶश΋ఘΊΔ • ͦͷ୅ΘΓʮWhat is ਖ਼ৗʁʯͷఆٛΛֶशͯ͠ɺʮNot ਖ਼ৗʯΛҟৗͱ൑ఆ͢Δ

4. Anomaly Detection ͷैདྷݚڀ Deep One-Class Classification (ICML’18) • ਖ਼ৗσʔλͷΈΛ༻͍ͯɺClassifierͳΓAutoEncoderͳΓΛैདྷ௨Γʹֶश •

   ͜ͷͱ͖ɺಛ௃ྔ෼෍͕௿࣍ݩ෦෼ۭؒʹऩଋ͢ΔΑ͏LOSSΛՃ͑Δ • ਖ਼ৗσʔλͳΒ௒ٿ಺ʹ෼෍͢Δ͸ͣͳͷͰɺ௒ٿ͔Β֎ΕͨσʔλΛҟৗͱ൑ఆ͢Δ ୈҰ߲ʹΑͬͯ௒ٿ಺ʹ෼෍͕ԡ͠ࠐ·ΕΔ cɿ ௒ٿͷத৺ʢͨͩ͠≠0ʣ nɿֶश͢Δਖ਼ৗσʔλͷ਺

5. Anomaly Detection ͷධՁ؍఺ ͲΕ͚ͩਖ਼֬ʹҟৗΛݕ஌Ͱ͖͔ͨʁ • ਖ਼ৗσʔλΛਖ਼ৗͱ൑ఆͯ͠ɺҟৗσʔλΛҟৗͱ൑ఆ͢Δਫ਼౓ ԼྲྀλεΫΛअຐ͠ͳ͍͔ʁ • ԼྲྀλεΫ͕͋Δ৔߹ɺҟৗݕ஌ػೳͷ௥ՃʹΑͬͯѱӨڹ͕ͳ͍͔Ͳ͏͔ •

   ྫ͑͹ɺ10Ϋϥεͷ਺ࣈࣝผثʹɺਤܗͳͲ਺ࣈҎ֎͕ೖྗ͞Εͨͱ͖ҟৗͱ൑ఆ͢Δػ ೳΛ෇͚Ճ͍͑ͨͤͰɺैདྷͷ10Ϋϥεࣝผੑೳ͕௿Լ͢ΔͱࠔΔ ad-hoc͔post-hoc͔ʁ • ҟৗݕ஌͢ΔͨΊʹϞσϧߏ଄΍ֶशํ๏·Ͱม͑Δඞཁ͕͋Δ͔ʁ • ·ͨ͸ɺLOSSΛޙ͔Β͚̍ͭͩ෇͚଍ͯ͠௥Ճֶश͢Δ͚ͩͰOK͔ʁ • ͲͪΒ͕ྑ͍ѱ͍ͳͲҰ֓ʹ͸ݴ͑ͳ͍͕ɺpost-hocͷํ͕ѻ͍΍͍͢ɻ

6. ঺հ࿦จͷ֓ཁ ʮSemi-supervisedʹֶश͠Α͏ʂʯ Anomaly Detection ͷݚڀ͸Unsupervised͕ओྲྀͷΑ͏ͩ Ͱ΋ɺֶश༻ͷҟৗσʔλ͕ೖखࠔ೉ͩͱͯ͠΋ɺ ӡ༻Λଓ͚ͯͨΒҟৗσʔλʹ͍ͣΕग़ձ͏͸ͣ ͳΒ͹ɺͦΕΒগྔͷҟৗσʔλΛ࢖ͬͯɺ Semi-supervisedʹֶशͨ͠ํ͕ྑ͍ͷͰ͸ʁ ※Semi-supervisedͷAnomaly

   Detectionݚڀ͸ඇৗʹগͳ͍

7. ఏҊख๏ ʮLOSSʹ߲Λ̍ͭ௥Ճ͠·ͨ͠ʯ Deep One-Class Classification (ICML’18) ͷLOSSʹSemi-supervisedͷ߲Λ̍ͭ௥Ճ • ࣮͸ಉ͡ஶऀͰͨ͠ɻࣗ෼ͷݚڀΛࣗ෼ͰΞοϓσʔτͨ͠ܗʹͳΔɻ ΋͠ҟৗσʔλʹग़ձͬͨΒɺ

   ௒ٿͷ֎ଆʹߦ͘Α͏ֶश͢Δ mɿsemi-supervisedʹֶश͢Δσʔλ਺ yj ɿਖ਼ৗorҟৗͷϥϕϧ

8. ࣮ݧ݁Ռ ॎ࣠ɿҟৗσʔλͷݕग़ੑೳ ʢHigher is Betterʣ Unsupervised Semi-supervised ԣ࣠ɿSemi-supervisedͰڭࢣ෇͖ͷҟৗσʔλΛֶशׂͨ͠߹ ఏҊख๏ MNISTɺFashion-MNISTɺCIFAR-10ͷσʔληοτͰධՁ

   • ̍Ϋϥεͱਖ਼ৗͱఆٛͯ͠ɺAutoEncoderʴఏҊख๏Ͱಛ௃ྔදݱΛֶश • ࢒Γͷ̕ΫϥεΛೖྗͨ͠ͱ͖ɺҟৗͱ൑ఆͰ͖Δ͔Ͳ͏͔ධՁ ὎ ੑೳվળΛ֬ೝ

9. ·ͱΊͱߟ࡯ ਂ૚ֶशʹΑΔ Semi-supervised ͳ Anomaly Detection ख๏ΛఏҊ • ॳΊͯͰ͸ͳ͍ͱࢥ͏͕ɺਂ૚ֶशʹΑΔAnomaly DetectionͰsemi-supervised͸௝͍͠

   • ͔ͨ͠ʹࣾձ࣮૷Λߟ͑Δͱɺ͜ͷ໰୊ઃఆ͸ద੾ • ख๏΋γϯϓϧͰɺpost-hocͳͷͰѻ͍΍͍͢ • ࠓճ͸ԼྲྀλεΫ͕AE͕ͩͬͨɺClassificationͩͱͲ͏ͳΔ͔ʁ • Anomaly DetectionͷධՁσʔληοτͬͯଞʹͳ͍ͷ͔ͳɺɺɺɺ
 ʢ͍ͭ·ͰMNISTʹΑΔධՁ͕ଓ͘ͷͩΖ͏͔ʣ