Upgrade to Pro — share decks privately, control downloads, hide ads and more …

iclr2020deepsemi-supervisedanomalydetectionyamatookamoto-200531022507.pdf

 iclr2020deepsemi-supervisedanomalydetectionyamatookamoto-200531022507.pdf

Yamato.OKAMOTO

June 14, 2020
Tweet

More Decks by Yamato.OKAMOTO

Other Decks in Technology

Transcript

  1. 2020/6/14

    Yamato OKAMOTO
    ICLRΦϯϥΠϯಡΈձ
    Deep Semi-supervised Anomaly Detection

    View full-size slide

  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

    View full-size slide

  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

    View full-size slide

  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਺

    View full-size slide

  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

    View full-size slide

  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

    View full-size slide

  7. ఏҊख๏ ʮLOSSʹ߲Λ̍ͭ௥Ճ͠·ͨ͠ʯ
    Deep One-Class Classification (ICML’18) ͷLOSSʹSemi-supervisedͷ߲Λ̍ͭ௥Ճ
    • ࣮͸ಉ͡ஶऀͰͨ͠ɻࣗ෼ͷݚڀΛࣗ෼ͰΞοϓσʔτͨ͠ܗʹͳΔɻ
    ΋͠ҟৗσʔλʹग़ձͬͨΒɺ
    ௒ٿͷ֎ଆʹߦ͘Α͏ֶश͢Δ
    mɿsemi-supervisedʹֶश͢Δσʔλ਺
    yj
    ɿਖ਼ৗorҟৗͷϥϕϧ

    View full-size slide

  8. ࣮ݧ݁Ռ
    ॎ࣠ɿҟৗσʔλͷݕग़ੑೳ
    ʢHigher is Betterʣ
    Unsupervised Semi-supervised
    ԣ࣠ɿSemi-supervisedͰڭࢣ෇͖ͷҟৗσʔλΛֶशׂͨ͠߹
    ఏҊख๏
    MNISTɺFashion-MNISTɺCIFAR-10ͷσʔληοτͰධՁ
    • ̍Ϋϥεͱਖ਼ৗͱఆٛͯ͠ɺAutoEncoderʴఏҊख๏Ͱಛ௃ྔදݱΛֶश
    • ࢒Γͷ̕ΫϥεΛೖྗͨ͠ͱ͖ɺҟৗͱ൑ఆͰ͖Δ͔Ͳ͏͔ධՁ ὎ ੑೳվળΛ֬ೝ

    View full-size slide

  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

    View full-size slide