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

Human-in-the-Loop 機械学習 / Human-in-the-Loop Machine Learning

Human-in-the-Loop 機械学習 / Human-in-the-Loop Machine Learning

Yukino Baba
PRO

February 24, 2021
Tweet

More Decks by Yukino Baba

Other Decks in Research

Transcript

  1. ஜ೾େֶγεςϜ৘ใܥഅ৔ઇ೫
    CBCB!DTUTVLVCBBDKQ
    !ZVLJOP
    )VNBOJOUIF-PPQ

    ػցֶश

    View Slide

  2. /35
    )VNBOJOUIF-PPQػցֶशɿਓ͕ؒࢀՃ͢Δػցֶशϓϩηε
    2
    1BSU

    ܈ऺʹΑΔ

    ܇࿅σʔλ࡞੒
    1BSU

    ܈ऺ͔Βͷֶश
    1BSU

    ػցֶशͷ༷ʑͳ৔໘Ͱͷ

    ܈ऺ׆༻
    Ϟσϧ
    adelie
    gentoo
    chinstrap
    ܇࿅σʔλ
    Ϟσϧ
    ໰߹ͤ
    ڭࢣϥϕϧ

    ͷఏڙ
    ໰߹ͤ
    ༷ʑͳ஌ݟ

    ͷఏڙ
    2ΑΓྑ͍ϞσϧΛޮ཰తʹֶश͢ΔͨΊʹਓؒΛͲ͏׆༻͢Δ͔ʁ
    ڭࢣϥϕϧ

    ͷఏڙ
    ෆಛఆଟ਺ͷਓʢ܈ऺʣΛ૝ఆ

    View Slide

  3. /35
    Ϋϥ΢υιʔγϯάɿෆಛఆଟ਺ͷਓʹগֹͰ୯७࡞ۀΛґཔ͢Δ࢓૊Έ
    3
    Ϩγʔτͷॻ͖ى͜͠
    ʢʣ
    ֆըͷײ৘ϥϕϧ෇༩
    ʢʣ
    ྫɿ"NB[PO.FDIBOJDBM5VSL
    ґཔҰཡ ࡞ۀը໘
    https://www.mturk.com

    View Slide

  4. 1BSU

    ܈ऺʹΑΔ܇࿅σʔλ࡞੒

    View Slide

  5. /35
    ˔ ܈ऺʹϥϕϧΛ໰͍߹Θͤ܇࿅σʔλΛ࡞੒
    ˔ ෆಛఆଟ਺ͷਓ͕ࢀՃ͢ΔͨΊڭࢣϥϕϧͷ඼࣭͕՝୊
    ৴པੑ͕௿͍ࢀՃऀ΋͍Δ
    ϥϕϧ෇͚ཁ͕݅े෼ʹ఻ΘΒͳ͍͜ͱ͕͋Δ
    େྔͷϥϕϧ෇͚݁ՌΛґཔऀ͕શͯݕূ͢Δͷ͸ࠔ೉
    ඼࣭อূ
    ڭࢣϥϕϧͷ඼࣭͕՝୊
    5
    Choose the correct category
    Adelie
    Chinstrap
    Gentoo
    “This was the best book I ever read!!! Thank you so much! :)”
    What emotion does this text convey?
    Anger Disgust Fear Happiness
    Sadness Surprise

    View Slide

  6. /35
    ࢀՃऀબൈ΍ฒྻԽɾ௚ྻԽͰ඼࣭อূ
    6
    ࢀՃऀબൈ ฒྻԽ ௚ྻԽ
    ࣄલςετͰબൈ
    ճ౴
    ໰୊ Chinstrap
    ໰୊ Adelie
    ໰୊ Adelie
    ਖ਼ղط஌ͷ໰୊Λࠞͥͯબൈ
    ճ౴
    ໰୊ Gentoo
    ໰୊ Adelie
    ໰୊ Gentoo
    ·ͨ͸
    ໰୊ͷਖ਼ղ͸"EFMJF

    ˠޡ౴ͨ͠ɹɹͷճ౴͸શͯআ֎
    ඼࣭อূ
    Adelie
    Adelie Adelie Gentoo
    ಉ͡໰୊΁ͷෳ਺ਓͷճ౴Λ౷߹
    Adelie
    Chinstrap
    Gentoo
    ଟ਺ܾ౳Ͱ

    ౷߹
    ͋Δਓͷճ౴Λଞऀ͕ݕূɾमਖ਼
    ճ౴
    ໰୊ Gentoo
    ໰୊ Adelie
    ໰୊ Gentoo
    ճ౴
    ໰୊ Gentoo
    ໰୊ Adelie
    ໰୊ Gentoo



    ݕূ
    ࣮૷͕༰қͳฒྻԽ͕

    ޿͘༻͍ΒΕ͍ͯΔ

    View Slide

  7. /35
    ฒྻԽʹΑΔ඼࣭อূ
    ଟ਺ܾͰ͸֤ࣗͷ৴པੑΛߟྀͰ͖ͳ͍
    7
    NO YES YES YES
    YES YES NO YES
    YES YES YES NO
    YES
    YES
    YES
    ໰୊
    ଟ਺ܾ͸֤ࣗͷճ౴ͷॏΈ͕౳͍͠ͱΈͳ͕͢

    ճ౴ऀ͝ͱʹ৴པੑ͸ҟͳΔ͸ͣ
    ଟ਺ܾͰ

    ༧ଌͨ͠ਖ਼ղ
    “Is a bird in

    the picture?"



    View Slide

  8. /35
    ฒྻԽʹΑΔ඼࣭อূ
    ֤ࣗͷ৴པੑ͕ෆ໌ͳͨΊॏΈ෇͖ଟ਺ܾ͸࢖͑ͳ͍
    8
    ໰୊
    NO YES YES YES
    YES YES NO YES
    YES YES YES NO
    ɹɹɹ YES
    ɹɹɹ YES
    ɹɹɹ NO
    ॏΈ


    ॏΈ


    ॏΈ


    ॏΈ


    :&4ථ

    /0ථ
    :&4ථ

    /0ථ
    :&4ථ

    /0ථ
    ॏΈ෇͖ଟ਺ܾͰ

    ༧ଌͨ͠ਖ਼ղ
    ࣮ࡍʹ͸ॏΈ͸Θ͔Βͳ͍
    “Is a bird in

    the picture?"



    View Slide

  9. /35
    ˔ ճ౴͔Β֤ࣗͷ৴པੑͳͲΛਪఆ͠ਖ਼ղͷ༧ଌʹ༻͍Δ
    ˔ ੜెͷճ౴͚͔ͩΒࢼݧͷਖ਼ղΛ༧ଌ͢ΔΑ͏ͳ΋ͷ
    ਅ࣮ൃݟ
    ਅ࣮ൃݟ 5SVUIEJTDPWFSZ
    ɿෳ਺ਓͷճ౴͔Βͷਖ਼ղ༧ଌ໰୊
    9
    ਖ਼ղ
    ໰୊
    NO YES YES YES
    YES YES NO YES
    YES YES YES NO
    ?
    ?
    ?
    “Is a bird in

    the picture?"

    View Slide

  10. /35
    ਅ࣮ൃݟ
    %BXJE4LFOFճ౴ऀͷ৴པੑΛࠞಉߦྻͰදݱ
    10
    A. P. Dawid and A. M. Skene: Maximum likelihood estimation of observer error-rates using the EM algorithm. Journal of the Royal Statistical Society,
    Series C (Applied Statistics), 1979.
    ɿճ౴ऀ ͕ਖ਼ղ͕YESͷ໰୊ʹYESͱ౴͑Δ֬཰
    αj j
    ɿճ౴ऀ ͕ਖ਼ղ͕NOͷ໰୊ʹNOͱ౴͑Δ֬཰
    βj j
    ճ౴ऀͷ৴པੑύϥϝʔλʢࠞಉߦྻʣ
    ti
    ਖ਼ղ
    YES
    ti
    =
    NO
    ti
    =
    yij
    βj
    ճ౴
    ճ౴Ϟσϧʢ໰୊ ʹର͢Δճ౴ऀ ͷճ౴ʣ
    i j
    αj
    ճ౴
    YES NO


    ղ
    YES
    NO
    αj
    βj
    1 − αj
    1 − βj
    ࠞಉߦྻ
    Pr[yij
    ∣ ti
    = 1] = αyij
    j
    (1 − αj
    )(1−yij
    )
    Pr[yij
    ∣ ti
    = 0] = β(1−yij
    )
    j
    (1 − βj
    )yij

    View Slide

  11. /35
    ਅ࣮ൃݟ
    %BXJE4LFOFճ౴ऀͷ৴པੑΛࠞಉߦྻͰදݱ
    11
    qi
    = Pr[ti
    = 1 ∣ {yij
    }] ∝ p∏
    j
    αyij
    j
    (1 − αj
    )1−yij
    αj
    =

    i
    qi
    yij

    i
    qi
    , βj
    =

    i
    (1 − qi
    )yij

    i
    (1 − qi
    )
    , p =

    i
    qi
    N
    ਖ਼ղͷࣄޙ֬཰ ˔
    ΋ Λ࢖ͬͯܭࢉ͠ Λࢉग़
    ˔
    Pr[ti
    = 0 ∣ {yij
    }] βj
    qi
    p = Pr[ti
    = 1]
    ਖ਼ղ͕YESͷ໰୊Ͱͷਖ਼౴཰ͷΑ͏ͳ΋ͷ
    &TUFQճ౴ऀ৴པੑΛݻఆͯ͠ਖ਼ղΛਪఆ
    .TUFQਖ਼ղΛݻఆͯ͠ճ౴ऀ৴པੑΛਪఆ
    ˞ ͸໰୊਺
    N
    &.ΞϧΰϦζϜΛ࢖͍ճ౴ऀ৴པੑͱਖ਼ղΛަޓʹਪఆ

    View Slide

  12. /35
    ਅ࣮ൃݟ
    %BXJE4LFOFʢͷվྑ൛ʣ͸"NB[PO.FDIBOJDBM5VSLͰར༻Մೳ
    12
    Amazon SageMaker GroundTruth

    "NB[PO.FDIBOJDBM5VSLΛར༻ͨ͠܇࿅σʔλ࡞੒Λࢧԉ
    https://aws.amazon.com/sagemaker/groundtruth/

    https://aws.amazon.com/jp/blogs/news/use-the-wisdom-of-crowds-with-amazon-sagemaker-ground-truth-to-annotate-data-more-accurately/
    ฒྻλεΫΛࣗಈൃߦɺ

    %BXJE4LFOFͰ

    ਖ਼ղΛ༧ଌͯ͠ग़ྗ
    ฒྻ਺Λࢦఆ
    ୯ՁΛࢦఆ
    λεΫͷछྨΛࢦఆ

    View Slide

  13. /35
    ਅ࣮ൃݟ
    $PNNVOJUZ#$$ࠞಉߦྻΛෳ਺ਓͰڞ௨Խ
    13
    M. Venanzi, J. Guiver, G. Kazai, P. Kohli, and M. Shokouhi: Community-based bayesian aggregation models for crowdsourcing. WWW 2014.
    Truth Predicted
    Truth Predicted
    Truth Predicted
    0 1 234
    4
    3
    2 3
    1
    0
    0 1 234
    4
    3
    2 3
    1
    0
    0 1 234
    4
    3
    2 3
    1
    0
    0.8
    0.4
    0
    0.8
    0.4
    0
    0.8
    0.4
    0
    Decisive (5%)
    Conservative (4%)
    Calibrated (91%)
    ճ౴ऀ͕൑அ܏޲ʹج͍ͮͯ

    ίϛϡχςΟΛܗ੒͍ͯ͠Δͱ͢Δ
    ίϛϡχςΟͷࠞಉߦྻ͔Β

    ֤ࣗͷࠞಉߦྻ͕ੜ੒͞ΕΔ
    ࠞಉߦྻ
    ࠞಉߦྻͷਪఆྫɿίϛϡχςΟ͝ͱʹҟͳΔ൑அ܏޲Λଊ͍͑ͯΔ
    0: Negative


    1: Neutral


    2: Positive


    3: Not-related


    4: Unknown
    ࠞಉߦྻ

    View Slide

  14. /35
    ˔ ྫɿ܈ऺʹΑΔ࠶൜ϦεΫ༧ଌ͸ಛఆͷਓछʹରͯ͠ෆެฏ
    ˔ ճ౴ऀͷόΠΞεΛਪఆɾআڈ͠ͳ͕Βެฏͳਖ਼ղΛਪఆ
    ਅ࣮ൃݟ
    'BJS5%ෆެฏͳճ౴͔Βͷެฏͳਅ࣮ൃݟ
    14
    Y. Li, H. Sun, W. H. Wang: Towards fair truth discovery from biased crowdsourced answers. KDD 2020.
    yij
    = ti
    +
    𝒩
    (0,σ2
    j
    ) + bA=a
    j
    “The defendant is a [RACE] [SEX] aged [AGE]. They have been charged with: [CRIME CHARGE]. This crime is
    classi
    fi
    ed as a [CRIMINAL DEGREE]. They have been convicted of [NON-JUVENILE PRIOR COUNT] prior crimes.
    They have [JUVENILE- FELONY COUNT] juvenile felony charges and [JUVENILE-MISDEMEANOR COUNT]
    juvenile misdemeanor charges on their record.”
    Do you think this person commit another crime within 2 years?
    J. Dressel and H. Farid: The accuracy, fairness, and limits of predicting recidivism. Science advances, 2018.
    ճ౴Ϟσϧ
    อޢάϧʔϓʹର͢ΔόΠΞε
    ΞϧΰϦζϜ

    ਪఆͨ͠ਖ਼ղͷόΠΞεΛࢉग़ɺ

    ٯ޲͖ͷόΠΞεΛ࣋ͭճ౴ऀΛҰਓબ୒

    બ୒ͨ͠ճ౴ऀҎ֎ͷόΠΞεΛআڈͯ͠ਖ਼ղΛਪఆ

    ਪఆͨ͠ਖ਼ղΛ༻͍ͯ Λߋ৽
    bA=a
    j
    , σj

    View Slide

  15. /35
    අ༻ͷޮ཰Խ
    Ұఆਓ਺͕ಉ͡ճ౴Λͨ࣌͠఺Ͱ໰߹ͤΛதࢭ
    15
    L. von Ahn, B. Maurer, C. McMillen, D. Abraham, and M. Blum: reCAPTCHA: Human-based character recognition via web security measures. Science,
    2008.
    morning morninq morning
    ̎ਓ͕ಉ͡౴͑Λฦͨ͠ͷͰ

    morning

    Λೝࣝ݁Ռͱͯ͠࠾༻ͯ͠ऴྃ
    Type the word Type the word
    SF$"15$)"ҹ࡮෺ͷจࣈΛਓؒʹೝࣝͤ͞ΔϓϩδΣΫτ

    View Slide

  16. /35
    ˔ ݱࡏ·Ͱͷ:&4/0ͷճ౴਺ʹԠͯ͡໰߹ͤͷ௥ՃΛܾఆ
    ˔ ޡ൑ఆͷ֬཰͕ҰఆҎ্ͷ৔߹ʹ໰߹ͤΛ௥Ճ͢Δ
    අ༻ͷޮ཰Խ
    ෆ࣮֬ੑʹԠͯ͡৽ͨͳ໰߹ͤΛ௥Ճ
    16
    V. S. Sheng, F. Provost, and P. G. Ipeirotis: Get another label? Improving data quality and data mining using multiple, noisy labelers. KDD 2008.
    NO YES YES YES NO :&4ͷਓ਺
    /0ͷਓ਺
    n = 3
    ¯
    n = 2
    ผͷਓʹ໰͍߹ΘͤΔ͔ʁ
    ਖ਼ղ͕YESͰ͋Δࣄޙ֬཰
    ͕
    ʹै͏ͱ͢Δ
    q Beta(n + 1,¯
    n + 1)
    I0.5
    (n, ¯
    n)
    NOͱ൑ఆ YESͱ൑ఆ
    ਖ਼ղ͕YESͳΒޡ൑ఆͷ֬཰͸I0.5
    (n, ¯
    n)
    ਖ਼ղ͕NO ͳΒޡ൑ఆͷ֬཰͸
    1 − I0.5
    (n, ¯
    n)
    ὎ ͷͱ͖໰߹ͤ௥Ճ
    min{I0.5
    (n, ¯
    n),1 − I0.5
    (n, ¯
    n)} > ϵ

    View Slide

  17. /35
    ଟ਺ܾͷ౴͑Λਖ਼ղͱݟͳͯ͠

    ֤ࣗͷਖ਼౴֬཰ͱͦͷ৴པ۠ؒΛਪఆ
    ৴པ۠ؒͷ্ݶ͕ߴ͍ճ౴ऀɹɹɹΛ࠾༻
    ˔ ৴པੑͷߴ͍ճ౴ऀʹ໰͍߹ΘͤΔ͜ͱͰ༧ࢉΛޮ཰ར༻͍ͨ͠
    ˔ ୭͕৴པͰ͖Δ͔͸ࣄલʹ͸Θ͔Βͳ͍ͷͰ

    ଟ࿹όϯσΟοτ໰୊ʹ͓͚Δ6$#ઓུΛར༻ͯ͠

    ৴པͰ͖Δճ౴ऀ΁ͷ໰߹ͤʢ׆༻ʣͱಉ࣌ʹ৴པͰ͖Δճ౴ऀΛ୳ࡧ
    අ༻ͷޮ཰Խ
    *&5ISFTI৴པͰ͖Δճ౴ऀΛ୳ࡧɾ׆༻
    17
    P. Donmez, J. G. Carbonell, and J. Schneider: E
    ff
    i
    ciently learning the accuracy of labeling sources for selective sampling. KDD 2009.
    ਖ਼౴֬཰
    ਖ਼౴֬཰͕ߴ͍ˠ׆༻
    ୳ࡧ͕ෆे෼ɹˠ୳ࡧ
    {

    View Slide

  18. 1BSU

    ܈ऺ͔Βͷֶश

    View Slide

  19. /35
    ܈ऺ͔Βͷֶश
    ܈ऺͷճ౴͔ΒػցֶशϞσϧΛ௚઀ֶश
    19
    ௨ৗͷֶश
    ܈ऺ͕࡞੒ͨ͠

    ܇࿅σʔλ͔Βͷֶश
    ܈ऺ͔Βͷֶश
    adelie chinstrap chinstrap chinstrap
    gentoo adelie gentoo adelie
    chinstrap gentoo gentoo gentoo
    adelie
    gentoo
    chinstrap
    ਖ਼ղ
    ֶश
    adelie
    gentoo
    chinstrap
    ֶश
    ਖ਼ղΛਪఆ
    adelie chinstrap chinstrap chinstrap
    gentoo adelie gentoo adelie
    chinstrap gentoo gentoo gentoo
    ֶश
    Ϟσϧ

    View Slide

  20. /35
    ܈ऺ͔Βͷֶश
    ճ౴Ϟσϧ͔Βਪఆͨ͠ਖ਼ղΛར༻ͯ͠෼ྨϞσϧΛֶश
    20
    V. C. Raykar, S. Yu, L. H. Zhao, G. H. Valadez, C. Florin, L. Bogoni, and L. Moy: Learning from crowds. Journal of Machine Learning Research, 2010.
    ෼ྨϞσϧ
    fw
    (xi
    ) = Pr [yi
    = 1 ∣ xi] =
    1
    1 + exp(−w⊤xi
    )
    ෼ྨϞσϧͷֶशखॱʢऩଋ͢Δ·Ͱ܁Γฦ͢ʣ
    &TUFQ

    ճ౴ऀࠞಉߦྻΛݻఆͯ͠ਖ਼ղΛਪఆʢ%BXJE4LFOFͱಉ༷ʣ
    .TUFQ

    ਪఆͨ͠ਖ਼ղΛ༻͍ͯճ౴ऀࠞಉߦྻͱ෼ྨϞσϧΛߋ৽
    ௨ৗͷ܇࿅σʔλ {(xi
    , yi
    )} {(xi
    , yi1
    , yi2
    , …)}
    ܈ऺ͔Βͷֶशͷ܇࿅σʔλ
    ճ౴Ϟσϧ
    Pr[yij
    ∣ yi
    = 1] = αyij
    j
    (1 − αj
    )(1−yij
    )
    Pr[yij
    ∣ yi
    = 0] = β(1−yij
    )
    j
    (1 − βj
    )yij

    View Slide

  21. /35
    ܈ऺ͔Βͷֶश
    "HH/FUճ౴Ϟσϧ͔Βਪఆͨ͠ਖ਼ղΛར༻ͯ͠ਂ૚ֶशϞσϧΛֶश
    21
    S. Albarqouni, C. Baur, F. Achilles, V. Belagiannis, S. Demirci, and N. Navab: AggNet: Deep learning from crowds for mitosis detection in breast cancer
    histology images. IEEE Transactions on Medical Imaging, 2016.
    ਂ૚ֶशʹΑΔ෼ྨϞσϧ ճ౴Ϟσϧ
    &TUFQͰਪఆͨ͠ਖ਼ղЖΛ
    ༻͍ͯ෼ྨϞσϧΛߋ৽
    ALBARQOUNI et al.: AGGNET: DEEP LEARNING FROM CROWDS FOR MITOSIS DET
    Fig. 2. AggNet architecture: The same CNN architecture is used for different
    w
    e
    t
    b
    a
    g
    f
    m

    View Slide

  22. /35
    ܈ऺ͔Βͷֶश
    ڞ௨ɾݸผͷࠞಉߦྻΛ༻ҙͯ͠ਪఆΛޮ཰Խ
    22
    Z. Chu, J. Ma, and H. Wang: Learning from crowds by modeling common confusions. AAAI 2021.
    model parameter estimation. We define cross-entropy loss
    on the observed annotations and use error back-propagation
    to update the classifier’s output and the network parameters
    simultaneously.
    e
    1:R
    xi
    Classifier
    Aux.Net
    Wa
    1:R
    i
    1:R
    i
    (1 1:R
    i
    )
    + =
    f
    i
    Wgf
    i
    W1:Rf
    i
    h1:R
    i
    input
    parallel noise
    adaptation layers
    backbone model predicted anno. dist.
    where (W
    terms fo
    moid fun
    fer the p
    the mag
    or small
    we norm
    before c
    Based
    modelin
    the netw
    ing the
    feature v
    ing the c
    and pred
    ڞ௨ࠞಉߦྻ
    ݸผࠞಉߦྻ ࣮ࡍͷճ౴ͱ

    ൺֱ͠ଛࣦΛܭࢉ
    ෼ྨϞσϧ
    ೖྗ
    ճ౴ऀಛ௃ ڞ௨ͷࠞಉߦྻΛ࢖͏֬཰
    ิॿϞσϧ

    View Slide

  23. /35
    ೳಈֶश
    ֶशʹ༗༻ͳαϯϓϧΛબͼϥϕϧ෇͚ճ਺ΛݮΒ͢
    23
    ֶशʹ༗༻ͳαϯϓϧΛબΜͰϥϕϧ෇͚

    ˠ৽͍͠ϥϕϧͰ෼ྨϞσϧΛߋ৽
    adelie
    ༗༻ͳ

    αϯϓϧΛબ୒
    ϞσϧΛߋ৽
    ϥϕϧ෇͚
    ೳಈֶश "DUJWFMFBSOJOH

    ༗༻ੑͷࢦඪ
    ༧ଌ෼෍Λ༻͍ͯෆ࣮֬ੑΛଌΔ

    ʢΤϯτϩϐʔ౳Λར༻ʣ adelie
    chinstrap
    gentoo
    0 0.25 0.5 0.75 1
    adelie
    chinstrap
    gentoo
    0 0.25 0.5 0.75 1
    ༧ଌ෼෍
    ༧ଌ͕࣮֬ ༧ଌ͕ෆ࣮֬

    View Slide

  24. /35
    ܈ऺ͔Βͷೳಈֶश
    ༗༻ͳαϯϓϧʹ৴པੑͷߴ͍ճ౴ऀΛׂΓ౰ͯΔ
    24
    Y. Yan, R. Rosales, G. Fung, and J. G. Dy: Active learning from crowds. ICML 2011.
    fw
    (x) = Pr [y = 1 ∣ x] =
    1
    1 + exp(−w⊤x)
    ෼ྨϞσϧ
    ηwj
    (x) =
    1
    1 + exp(−w⊤
    j
    x)
    ճ౴Ϟσϧʢਖ਼౴֬཰ʣ
    ೳಈֶशͷखॱ

    ෼ྨϞσϧʹΑΔ༧ଌ͕࠷΋ෆ࣮֬ͳαϯϓϧ
    ΛબͿ

    ਖ਼౴֬཰
    ͕࠷େͷճ౴ऀΛબͼճ౴Λऔಘ

    ಘΒΕͨճ౴Λ༻͍ͯ෼ྨϞσϧͱճ౴Ϟσϧ
    Λߋ৽
    x*
    ηwj
    (x*)
    w, {wj
    }

    View Slide

  25. /35
    ܈ऺ͔Βͷೳಈֶश
    ਂ૚ֶशʹ༗༻ͳαϯϓϧʹ৴པੑͷߴ͍ճ౴ऀΛׂΓ౰ͯΔ
    25
    J. Yang, T. Drake, A. Damianou, and Y. Maarek: Leveraging crowdsourcing data for deep active Learning an application: Learning intents in Alexa. WWW
    2018.
    ϞϯςΧϧϩυϩοϓΞ΢τʹΑΔෆ࣮֬ੑධՁ
    ˔ ਂ૚ֶशωοτϫʔΫͷҰ෦ΛϥϯμϜʹ

    ܽམͤͨ͞΋ͷΛෳ਺༻ҙ
    ˔ ༧ଌ෼෍ͷฏۉΛ༻͍ͯෆ࣮֬ੑΛධՁ
    ճ౴Ϟσϧʢਖ਼౴֬཰ʣ
    ηwj
    (x) =
    1
    1 + exp(−w⊤
    j
    Fx)
    , wj
    ∈ Rk, F ∈ Rk×d, x ∈ Rd
    Λ௿࣍ݩʹຒΊࠐΉ
    x
    ˞ೳಈֶशͷखॱ͸લεϥΠυͷख๏ͱಉ༷

    View Slide

  26. 1BSU

    ػցֶशͷ༷ʑͳ৔໘Ͱͷ

    ܈ऺ׆༻

    View Slide

  27. /35
    ಛ௃நग़
    ܭࢉػͰ͸ଊ͑ΒΕͳ͍ಛ௃Λਓ͕ؒநग़
    27
    S. Branson, C. Wah, F. Schro
    ff
    , B. Babenko, P. Welinder, P. Perona, S. Belongie: Visual recognition with humans in the loop. ECCV 2010.
    Visual Recognition with Humans in the Loop 3
    mputer vision is helpful Computer vision is not helpful
    mputer vision is helpful Computer vision is not helpful
    The bird is a
    Black‐footed
    Albatross
    Is the belly
    white? yes
    Are the eyes
    white? yes
    Th bi d i
    Is the beak cone‐shaped? yes
    Is the upper‐tail brown? yes
    Is the breast solid colored? no
    Is the breast striped? yes
    I h h hi ?
    The bird is a
    Parakeet Auklet
    Is the throat white? yes
    The bird is a Henslow’s
    Sparrow
    2. Examples of the visual 20 questions game on the 200 class Bird dataset.
    an responses (shown in red) to questions posed by the computer (shown in blue)
    sed to drive up recognition accuracy. In the left image, computer vision algorithms
    uess the bird species correctly without any user interaction. In the middle image,
    uter vision reduces the number of questions to 2. In the right image, computer
    n provides little help.
    2ෲ͕ന͍ʁ
    2໨͕ന͍ʁ
    2ͪ͘͹͕͠ίʔϯܕʁ
    YES
    YES
    NO
    ಛ௃̍ ಛ௃ ಛ௃
    1 1 0
    ػցֶशͰ͸ಛ௃ઃܭ͕ॏཁ

    ˠࣄલʹ࣭໰Λ༻ҙ͓͖ͯ͠܈ऺʹճ౴ͤͯ͞ಛ௃Λநग़

    View Slide

  28. /35
    ಛ௃ઃܭ
    "EB'MPDLಛ௃நग़͚ͩͰ͸ͳ͘ಛ௃ઃܭ΋ਓ͕࣮ؒࢪ
    28
    R. Takahama, Y. Baba, N. Shimizu, S. Fujita, and H. Kashima: AdaFlock: Adaptive feature discovery for human-in-the-loop predictive modeling. AAAI 2018.
    B
    ਖ਼ྫɾෛྫΛ

    ਓؒʹఏࣔ
    C
    ਖ਼ྫɾෛྫΛ

    ۠ผ͢Δ࣭໰จΛ

    ਓ͕ؒੜ੒

    ʢಛ௃ઃܭʣ
    D
    ਓ͕࣭ؒ໰ʹճ౴

    ʢಛ௃நग़ʣ
    E
    ෼ྨثΛߋ৽͠

    ಛ௃ઃܭʹ࢖͏

    αϯϓϧΛબ୒
    ྫɿϞωͱγεϨʔͷֆͷ෼ྨ

    View Slide

  29. /35
    Ϟσϧݕࠪ
    ෆద੾ͳಛ௃Λਓؒʹআڈͤͯ͞ϞσϧΛվળ
    29
    M. T. Ribeiro, S. Singh, and C. Guestrin: "Why should I trust you?": Explaining the predictions of any classi
    fi
    er. KDD 2016.


    https://drive.google.com/
    fi
    le/d/0ByblrZgHugfYZ0ZCSWNPWFNONEU/view
    ྫɿफڭؔ࿈ϝʔϧͷ൑ఆ
    Ϟσϧͷ

    ൑அࠜڌΛఏࣔ
    )PTU 1PTUJOH //51౳ͷ

    ແؔ܎ͳ୯ޠΛॏࢹ͢Δͷ͸ෆద੾

    View Slide

  30. /35
    ղऍੑ
    ਓؒʹΑΔ൑அ࣌ؒΛߟྀͯ͠ղऍੑͷߴ͍ϞσϧΛൃݟ
    30
    I. Lage, A. Ross, S. J. Gershman, B. Kim, and F. Doshi-Velez: Human-in-the-loop interpretability prior. NeurIPS 2018.

    ਫ਼౓ͷߴ͍ϞσϧΛީิͱͯ͠ྻڍ

    ਓؒʹධՁͤ͞ΔϞσϧ
    Λબ୒
    M

    ਓ͕ؒϞσϧ
    ͷղऍੑ
    ΛධՁ
    M p(M)

    ࠷ྑͷϞσϧΛܾఆ
    p(M) ≈
    1
    N ∑
    x
    𝖧
    𝖨𝖲
    (x, M)
    𝖧𝖨𝖲
    (x, M) = max{0,
    𝗆𝖺
    𝗑 𝖱𝖳

    𝗆𝖾𝖺 𝗇𝖱𝖳
    (x, M)}
    ฏۉ൑அ࣌ؒ
    (a) An example of our interface with a tree trained on
    Ϟσϧͷઆ໌ʹج͍ͮͯਓؒʹ༧ଌΛͤ͞Δ

    ˠॴཁ͕࣌ؒ୹͍΄Ͳղऍੑͷߴ͍Ϟσϧ

    View Slide

  31. /35
    ఢରతੜ੒ωοτϫʔΫʢ("/ʣ
    )VNBO("/ਓؒΛ("/ͷࣝผثʹͯ͠ਓؒͷײੑΛऔΓࠐΉ
    31
    K. Fujii, Y. Saito, S. Takamichi, Y. Baba, and H. Saruwatari: HumanGAN: Generative adversarial network with human-based discriminator and its evaluation
    in speech perception modeling. ICASSP 2020.
    Pertur-
    bation
    Crowd-
    workers
    Backpropagation using
    approximated
    Worker’s answer to “to what degree
    are two samples different?”
    [times]
    2
    Fig. 2. Generator training procedure of proposed HumanGAN.
    Crowdworkers state a perceptual difference (i.e., difference of pos-
    terior probabilities) of two perturbed samples. Answer and perturba-
    tion are used for backpropagation to train generator.
    University of Tokyo, Japan.
    ty of Tsukuba, Japan.
    minator
    Natu-
    ral
    -based
    ata Distr. of human perception
    of basic GAN and proposed HumanGAN.
    rator by fooling DNN-based discriminator
    scriminator), and generator finally represents
    n. In comparison, HumanGAN trains gener-
    ࣝผثͱͯ͠ਓؒΛ༻͍Δ
    ਓؒΛ᱐͢ੜ੒ثΛֶश͢Δ͜ͱͰ
    ਓؒͷײੑΛऔΓࠐΉ

    View Slide

  32. /35
    ਓؒͱܭࢉػͷڠௐϞσϧ
    ਓؒ΁ͷ໰͍߹ΘͤΛؚΊͨϞσϧΛֶश
    32
    B. Wilder, E. Horvitz, and E. Kamar: Learning to complement humans. IJCAI 2020.
    ਓؒͱܭࢉػͷڠௐϞσϧ
    ೖྗ
    x
    ࣭໰ثq(x) ਓؒʹ໰͍߹Θͤͳ͍
    ਓؒʹ໰͍߹ΘͤΔ
    h
    ෼ྨث
    ̂
    y = f(x)
    ̂
    y = f(x, h)
    ڠௐϞσϧͷֶश
    ܇࿅σʔλ Λ༻͍ͯ ͕࠷େʹͳΔΑ͏ʹ

    ෼ྨثͱ࣭໰ثΛֶश
    {(x, y, h)} q(x){u(y, f(x, h)) − c} + (1 − q(x))u(y, f(x))
    ༧ଌͷਖ਼͠͞ͷޮ༻ؔ਺
    Λग़ྗ
    Λग़ྗ
    අ༻

    View Slide

  33. ·ͱΊ

    View Slide

  34. /35
    )VNBOJOUIF-PPQػցֶशɿਓ͕ؒࢀՃ͢Δػցֶशϓϩηε
    34
    1BSU

    ܈ऺʹΑΔ

    ܇࿅σʔλ࡞੒
    1BSU

    ܈ऺ͔Βͷֶश
    1BSU

    ػցֶशͷ༷ʑͳ৔໘Ͱͷ

    ܈ऺ׆༻
    Ϟσϧ
    adelie
    gentoo
    chinstrap
    ܇࿅σʔλ
    Ϟσϧ
    ໰߹ͤ
    ڭࢣϥϕϧ

    ͷఏڙ
    ໰߹ͤ
    ༷ʑͳ஌ݟ

    ͷఏڙ
    ڭࢣϥϕϧ

    ͷఏڙ
    ˔ ਅ࣮ൃݟ
    ˔ අ༻ͷޮ཰Խ
    ˔ ܈ऺ͔Βͷਂ૚ֶश
    ˔ ೳಈֶश
    ˔ ಛ௃நग़ɾઃܭ
    ˔ ղऍੑ޲্
    ˔ ਓؒͱܭࢉػͷڠௐϞσϧ
    ˔ FUD

    View Slide

  35. /35
    35
    ࣛౡٱ࢚ খࢁ૱ അ৔ઇ೫

    ώϡʔϚϯίϯϐϡςʔγϣϯͱΫϥ΢ιʔγϯά

    ߨஊࣾ
    Robert (Munro) Monarch.

    Human-in-the-Loop Machine Learning:

    Active learning and annotation for human-centered AI.

    Manning Publications, 2021.
    ࢀߟจݙ

    View Slide