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

人工知能の基本問題:これまでとこれから

itakigawa
September 27, 2023
49

 人工知能の基本問題:これまでとこれから

人工知能学会 第110回人工知能基本問題研究会(SIG-FPAI)
2019年9月24日(火)・25日(水)
カナモトホール(札幌市民ホール)・第 2 会議室

総説(Open Access)
瀧川一学, 人工知能基本問題研究会(FPAI). 人工知能, 2019 年 34 巻 5 号 p. 603-611. https://doi.org/10.11517/jjsai.34.5_603

itakigawa

September 27, 2023
Tweet

Transcript

  1. 110 (SIG-FPAI)

    2019 9 24-25
    ( )
    (AIP)

    iPS
    (WPI-ICReDD)
    [email protected]
    Permalink : http://id.nii.ac.jp/1004/00010344/

    View full-size slide

  2. Vol.34 No.5 (2019/9)
    !2
    ಛूɿʮݚڀձ঺հʯ
    w ਓ޻஌ೳجຊ໰୊ݚڀձʢ'1"*ʣ
    w ஌ࣝϕʔεγεςϜݚڀձʢ,#4ʣ
    w ݴޠɾԻ੠ཧղͱର࿩ॲཧݚڀձʢ4-6%ʣ
    w ઌਐతֶशՊֶͱ޻ֶݚڀձʢ"-45ʣ

    w ࢢຽڞ૑஌ݚڀձʢ$$*ʣ
    w ܭଌΠϯϑΥϚςΟΫεݚڀձʢ.&*ʣ
    w ਎ମ஌ݚڀձʢ4,-ʣ
    w "*νϟϨϯδݚڀձʢ$IBMMFOHFʣ
    w ൚༻ਓ޻஌ೳݚڀձʢ"(*ʣ
    w ҩ༻ਓ޻஌ೳݚڀձʢ"*.&%ʣ

    View full-size slide

  3. = ...
    !3
    SIG-FPAI
    kashi_pong ( )
    1.
    2.
    3.
    4. The Hard Thing about Hard Things
    5. :
    6. :
    7. :
    8. :
    9. :

    FAI/FPAI

    FAI/FPAI
    ( 100 )

    View full-size slide

  4. !4
    AI ( ) https://jsai.ixsq.nii.ac.jp/


    (Permlink )

    View full-size slide

  5. ( )
    !5
    10
    (1995 2004)
    7
    (2005 2011)
    7
    (2012 2018)
    ( )

    " L1 "
    ( )

    ( )

    ( )

    + JST :
    ( )
    (2019 )
    AIP iPS 

    ( )

    View full-size slide

  6. (SIG-FPAI)
    !6
    https://sig-fpai.org

    View full-size slide

  7. AI ( )
    !8
    1984 

    ( UUCP JUNET )
    1992 (ISP)
    1993 NCSA Mosaic
    1994 IP
    1995 NTT Windows 95
    1996 Yahoo! JAPAN
    1999 ADSL 2
    2000 Google amazon.co.jp
    2001 Wikipedia ADSL Yahoo! BB ADSL
    2003
    2004 SNS (mixi Ameba GREE )
    2005 YouTube iTunes Music Store
    2008 iPhone Facebook Twitter
    2009 Google Android
    SIG
    FAI
    SIG
    FPAI
    1987
    2004
    Web ...

    ( )
    *+$"*
    5PLZP
    /BHPZB
    /BHPZB

    View full-size slide

  8. 30 : Vol.5, No.1 (1990 1 )
    !9
    Permalink : http://id.nii.ac.jp/1004/00002672/

    View full-size slide

  9. Vol.34 No.5 (2019/9)
    !10
    ಛूɿʮݚڀձ঺հʯ
    w ਓ޻஌ೳجຊ໰୊ݚڀձʢ'1"*ʣ
    w ஌ࣝϕʔεγεςϜݚڀձʢ,#4ʣ
    w ݴޠɾԻ੠ཧղͱର࿩ॲཧݚڀձʢ4-6%ʣ
    w ઌਐతֶशՊֶͱ޻ֶݚڀձʢ"-45ʣ

    w ࢢຽڞ૑஌ݚڀձʢ$$*ʣ
    w ܭଌΠϯϑΥϚςΟΫεݚڀձʢ.&*ʣ
    w ਎ମ஌ݚڀձʢ4,-ʣ
    w "*νϟϨϯδݚڀձʢ$IBMMFOHFʣ
    w ൚༻ਓ޻஌ೳݚڀձʢ"(*ʣ
    w ҩ༻ਓ޻஌ೳݚڀձʢ"*.&%ʣ

    View full-size slide

  10. !12
    Vol.34 No.5 (2019 9 )
    Permalink : http://id.nii.ac.jp/1004/00010296/
    ΦʔϓϯΞΫηεͰ
    ୭Ͱ΋ಡΊ·͢ʂ
    https://sig-fpai.org

    View full-size slide

  11. ...
    !13
    SIG-FPAI
    kashi_pong ( )
    1.
    2.
    3.
    4. The Hard Thing about Hard Things
    5. :
    6. :
    7. :
    8. :
    9. :

    FAI/FPAI

    FAI/FPAI
    ( 100 )

    View full-size slide

  12. 2019/02/25 14:44
    !15
    :
    2019 9 2018
    2016 30
    2 3

    2 28 xxxx
    :

    View full-size slide

  13. 2019/03/07 9:38
    !16
    2
    9 2019 


    View full-size slide

  14. Vol.11 No.3 (1996 5 )
    !17
    (1994-1995 FAI4 )
    (1992-1993 FAI3 )
    Permalink : http://id.nii.ac.jp/1004/00004008/

    View full-size slide

  15. !18
    (1990-1991 FAI2 )
    (1987-1989 FAI )
    Vol.11 No.3 (1996 5 )
    Permalink : http://id.nii.ac.jp/1004/00004008/

    View full-size slide

  16. !19
    Vol.7 No.1 (1992 1 )
    Permalink : http://id.nii.ac.jp/1004/00003060/

    View full-size slide

  17. ͍·ͷʮػցֶशʯͱ

    ͍ͩͿ༷ࢠ͕ҧ͏ ࿦ཧ

    w ʮϓϩάϥϛϯάݴޠݚڀʯ
    w ࿦ཧϓϩάϥϛϯά
    w ੍໿ϓϩάϥϛϯά
    w ؼೲ࿦ཧϓϩάϥϛϯά *-1

    w આ໌ʹجֶͮ͘श
    w ड़ޠ࿦ཧͷֶश
    w ໋୊࿦ཧͷֶश
    w ྨਪ
    w ؼೲਪ࿦
    w จ๏ਪ࿦ ݴޠͷؼೲֶश

    w ܭࢉ࿦తֶश
    w (PMEͷۃݶಉఆ

    w 1MPULJOͷ࠷খ൚Խ

    w "OHMVJOͷؼೲਪ࿦

    w "OHMVJOͷ࣭໰ֶश

    w 4IBQJSPͷϞσϧਪ࿦

    w 7BMJBOUͷ1"$ֶश

    '"*ʹ͓͚Δʮػցֶशʯ
    ܾఆ໦
    1SPHPM

    View full-size slide

  18. !21
    2019
    !?
    Vol.9 No.6 (1994 11 )
    Permalink : http://id.nii.ac.jp/1004/00003683/

    View full-size slide

  19. !22
    Vol.3 No.6 (1988 11 )
    Permalink : http://id.nii.ac.jp/1004/00002442/

    View full-size slide

  20. !23
    4 


    ...
    ஌ࣝ޻ֶΤΩεύʔτγεςϜͷ
    ʮ஌ࣝ֫ಘͷϘτϧωοΫʯ
    ͔Β஌ࣝൃݟɾσʔλϚΠχϯά΁
    Vol.11 No.6 (1996 11 )
    Permalink : http://id.nii.ac.jp/1004/00004121/

    View full-size slide

  21. !24
    ՊݚඅɾಛఆྖҬݚڀ "

    ʮڊେֶज़ࣾձ৘ใ͔Βͷ

    ஌ࣝൃݟʹؔ͢Δجૅݚڀʯ
    Vol.15 No.4 (2000 7 )
    Permalink : http://id.nii.ac.jp/1004/00004986/

    View full-size slide

  22. KDD: Knowledge Discovery in Databases
    !25
    ࠃࡍձٞ,%%͸΋ͱ΋ͱ͸"""*ͷ෼ՊձXPSLTIPQͱͯ͠೥୅ʹ࢝·Δ

    ࠷ޙͷ%Ͱ͋ΔʮJO%BUBCBTFTʯ͸ɺେن໛σʔλϕʔεࣗମ͕ίϞσΟςΟԽͨ͠

    ͜Ζʹ͠ΕͬͱʮBOE%BUB.JOJOHʯʹมߋʁͳͥ,%%.͡Όͳ͍ͷʁͱ͍͏ΞϨ

    View full-size slide

  23. !26
    Vol.18 No.5 (2003 9 )
    Permalink : http://id.nii.ac.jp/1004/00004121/

    View full-size slide

  24. Bernoulli-RIKEN Symposium on Neural Networks and Learning
    !27
    October 25 - 27, 2000
    Ohkouchi Hall, RIKEN (The Institute of Physical and Chemical Research), Japan
    1. Graphical Models and Statistical Methods
    • Steffen L. Lauritzen (Aalborg University)
    • Thomas S. Richardson (University of Warwick)
    • Lawrence Saul (AT&T Labs)
    • Martin Tanner (Northwestern University)
    2. Combining Learners
    • Leo Breiman (University of California, Berkeley)
    • Jerome H. Friedman (Stanford University)
    • Peter Bartlett (Australian National University)
    • Yoram Singer (The Hebrew University)
    3. Information Geometry and Statistical Physics
    • Shinto Eguchi (The Institute of Statistical Mathematics)
    • Shun-ichi Amari (RIKEN Brain Science Institute)
    • Manfred Opper (Aston University)
    • Magnus Rattray (University of Manchester)
    4. VC Dimension and SVM
    • Vladimir Vapnik (AT&T Labs)
    • Michael Kearns (AT&T Labs)
    • Gabor Lugosi (Pompeu Fabra University)
    • Bernhard Schoelkopf (Microsoft Research Ltd.)
    ͪ͜ΒͷྲྀΕ

    /*14ք۾
    ͸
    *#*4ݚڀձ΁
    ೥લޙ͸Ұੈ෩ᴆ
    ͨ͠χϡʔϥϧωοτ͕
    47.౳ͷΧʔωϧ๏ɺ
    3BOEPN'PSFTU΍
    #PPTUJOHͳͲͷΞϯα
    ϯϒϧ๏ɺ֊૚తͳϕΠ
    ζϞσϧͳͲʹҠ͍ͬͯ
    ͘λΠϛϯά

    View full-size slide

  25. !28

    1987 ( ) ...
    Empirical ( )



    empirical ...

    View full-size slide

  26. data driven
    !29



    GAFA
    empiricism theory-driven 

    ( )

    View full-size slide

  27. !30


    α 


    ...

    View full-size slide

  28. Theory-driven vs Data-driven
    !31
    David Hand
    All models are wrong, but some are useful
    (George Box)
    Theory-driven models can be wrong
    But data-driven models cannot be wrong
    or right
    Data-driven are not trying to describe an underlying reality.
    so they could be poor or useless, but not wrong
    But are merely intended to be useful
    http://videolectures.net/kdd2018_hand_data_science/
    ߨԋϏσΦ͸ԼهͰެ։ ൃදͰ͸͍͢͝αϓϥΠζ͕ώϯτɿࢦࣔ๮

    View full-size slide

  29. Here is a title text
    32
    With enough data, the numbers
    speak for themselves.
    Chris Anderson (2008)
    cf.

    View full-size slide

  30. Q. (or )
    !33
    ...
    ...

    (vs )


    View full-size slide

  31. !34
    or
    etc.
    (UAI; Uncertainty in AI)
    /
    ( )
    (or )
    vs
    αʔΧϜεΫϦϓγϣϯ

    View full-size slide

  32. vs :
    !35
    Chomsky-Norvig debate
    Rahimi-LuCun debate
    Leo Breiman "Two Cultures"
    [Biology] Genome Sciences / X-omics / Systems Biology 

    [Neuroscience] Connectomics/Whole-Brain Simulation

    "low input, high throughput, no output science." (Sydney Brenner)
    Gallistel and King 'Memory and the Computational Brain'
    David Marr 'Vision'
    David Hand
    George Box
    Henri Poincaré
    ) ( )
    Description or Explanation? Simulation or Emulation?
    Reverse-engineering a highly complex system whose inner workings are largely a mystery.
    Associationism (Correlation vs Causation), Deterministic Unpredictability (Chaos)
    ) JSAI 2019 (Preferred Networks)


    View full-size slide

  33. debate
    !36
    Vol.3 No.6 (1988 11 )
    Vol.2 No.4 (1987 12 )

    View full-size slide

  34. Chomsky Norvig debate
    !37
    Peter Norvig (Google)
    Noam Chomsky (MIT)
    • (Chomsky) σʔλ͔Βͷ౷ܭత༧ଌʹجͮ͘ݱࡏͷAIͰ͸Պֶ͕ఏڙ͢΂
    ͖આ໌΍ಎ࡯΍ҰൠݪཧʹࢸΒͳ͍ͷͰ͸ͱ൷൑
    • (Chomsky) EngineeringͷՁ஋͸෼͔Δ͠ཧղͰ͖ΔΜ͚ͩͲɺ΋ͱ΋ͱͷ
    scientific question͔Β཭Ε͍ͯͳ͍͔͍ʁʁ (ࢠڙͷݴޠ֫ಘetcʣ
    • (Norvig) Science͸ٕज़΍Factऩूͱڞʹ૑ΒΕΔͷͰ͸ɻFactͷղऍͱ͸
    ֬཰తͳ΋ͷͰ͸ͳ͍͔ɻ(ݕࡧΤϯδϯɾԻ੠ೝࣝɾػց຋༁ɾQAͷ੒ޭ)
    Factͷऩू+౷ܭత༧ଌ(֬཰తͳػցֶश༧ଌ) vs ୈҰݪཧͷཱ֬?

    (ߦಈओٛݴޠֶ vs ੜ੒จ๏΍ϛχϚϦετϓϩάϥϜͷ࠶೩ͱ΋?).

    View full-size slide

  35. )
    !38
    RNN(LSTM/GRU) CNN ...
    Google BERT
    OpenAI GPT-2
    GLUE SOTA
    SQuAD SOTA
    CMU XLNet
    Microsoft MT-DNN
    2018/10/18
    2019/01/31
    2019/02/14
    2019/06/19
    Google
    RNN CNN
    Attention
    Transformer !?

    (Attentive )

    View full-size slide

  36. ) Attention Context
    !39
    Attention ( ) ...

    (mixture modeling...)
    m
    X
    i=1
    ↵ifi(x)
    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
    m
    X
    i=1
    ↵i(x)fi(x)
    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


    context
    (query,key,value)

    context /subtype ...

    ( )
    "Attention Weights"
    m
    X
    i=1
    ↵(x, xi)yi
    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
    ↵(x, xi) =
    k(x, xi)
    P
    j
    k(x, xj)
    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
    query key
    value
    ( )

    ?
    ࢀߟ) Attention Models in Graphs: A Survey (2018, arXiv:1807.07984)

    View full-size slide

  37. ) ICML2019 Attention Tutorial
    !40
    ଞͷྫ
    • DeepSets
    • Squeeze Excitation Networks (SENet)
    • Graph Attention Networks (GAT)

    :

    X
    x
    (x)
    !
    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
    1PPMJOH
    "UUFOUJPO1PPMJOH

    X
    x
    ↵(x, w) (x)
    !
    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

    View full-size slide

  38. Context-Thick...? Attention...?
    !41
    Vol.10 No.3 (1995 5 )
    Permalink : http://id.nii.ac.jp/1004/00003785/

    View full-size slide

  39. Neural Abstract Machines & Program Induction (NAMPI)
    !42
    • Differentiable Neural Computers / Neural Turing Machines (Graves+ 2014)
    • Memory Networks (Weston+ 2014)
    • Pointer Networks (Vinyals+ 2015)
    • Neural Stacks (Grefenstette+ 2015, Joulin+ 2015)
    • Hierarchical Attentive Memory (Andrychowicz+ 2016)
    • Neural Program Interpreters (Reed+ 2016)
    • Neural Programmer (Neelakantan+ 2016)
    • DeepCoder (Balog+ 2016)
    context (attention ) ...
    IUUQTVDMNSHJUIVCJPOBNQJ
    09:00-09:30 Dawn Song: Deep Learning for Program synthesis: Lessons & Open Challenges
    09:30-10:00 Armando Solar-Lezama: Program synthesis and ML join forces
    10:30-11:00 Sumit Gulwani: Programming by Examples: Logical Reasoning meets Machine Learning
    11:00-11:30 Brenden Lake: Program induction for building more human-like machine learning algorithms
    11:30-12:00 Satinder Singh: Program Induction and Language: Two Vignettes
    12:00-12:30 Oriol Vinyals: Generating Visual Programs with Agents
    14:00-14:30 Rishabh Singh: Neural Meta Program Synthesis
    14:30-15:00 Veselin Raychev: Interpretable Probabilistic Models for Code
    15:00-15:30 Richard Evans: Differentiable Inductive Logic Programming
    ICML2018 Workshop on NAMPI

    View full-size slide

  40. Differentiable
    Inductive Logic Programming
    Richard Evans, Ed Grefenstette
    @LittleBimble, @egrefen
    https://jair.org/index.php/jair/article/view/11172

    View full-size slide

  41. Rahimi LeCun debate
    !44
    Yann LeCun

    (Facebook/NYU)
    Ali Rahimi 

    (Google/MIT)
    • (Ali Rahimi) ࠓͷػցֶश͸ʮAlchemyʯʹͳͬͯ͠·ͬͯɺػೳ͢Δ͚Ͳ
    த਎ͷ࢓૊Έ͸ෳࡶͰΑ͘Θ͔Βͳ͍ɻࣾձͷج൫(electricityʹྫ͑ͯ)ʹͭ
    ͔͏ʹ͸·ͩෆ҆ͳٕज़ͳͷͰAlchemy͔ΒElectricityΛऔΓ໭ͦ͏ɻ
    • (LeCun) ٕज़͸͍ͭ΋ཧ࿦తͳཧղΑΓઌߦ͖ͯͨ͠ɺ·ͨγϯϓϧͳఆཧ
    ͱҰൠԽ͸͢͹Β͍͕ͦ͠Ε͕ͳ͍ର৅΋͋Γ͑Δ͸ͣ (ྲྀମΛྫʹ)ɻ

    View full-size slide

  42. ) The Imposteriors NIPS2017
    !45

    View full-size slide

  43. Breiman "Two Cultures" in Statistical Modeling
    !46

    View full-size slide

  44. vs :
    !47
    Chomsky-Norvig debate
    Rahimi-LuCun debate
    Leo Breiman "Two Cultures"
    [Biology] Genome Sciences / X-omics / Systems Biology 

    [Neuroscience] Connectomics/Whole-Brain Simulation

    "low input, high throughput, no output science." (Sydney Brenner)
    Gallistel and King 'Memory and the Computational Brain'
    David Marr 'Vision'
    David Hand
    George Box
    Henri Poincaré
    ) ( )
    Description or Explanation? Simulation or Emulation?
    Reverse-engineering a highly complex system whose inner workings are largely a mystery.
    Associationism (Correlation vs Causation), Deterministic Unpredictability (Chaos)
    ) JSAI 2019 (Preferred Networks)


    View full-size slide

  45. !48
    Permalink : http://id.nii.ac.jp/1004/00007983/

    View full-size slide

  46. !51

    = (ill-posed)
    = regular
    (under tting)
    (over tting)
    etc.
    Inductive Bias

    View full-size slide

  47. Theory-driven vs Data-driven
    !52
    Theory-driven
    Data-driven
    " "
    Blackbox
    ( AI)
    (NP )
    ( )
    ( )
    Data-Driven ( ) 

    Data
    ML etc
    etc

    View full-size slide

  48. ...
    !53
    SIG-FPAI
    kashi_pong ( )
    1.
    2.
    3.
    4. The Hard Thing about Hard Things
    5. :
    6. :
    7. :
    8. :
    9. :

    FAI/FPAI

    FAI/FPAI
    ( 100 )

    View full-size slide

  49. https://itakigawa.github.io/news.html
    https://www.slideshare.net/itakigawa/presentations
    (Permlink )

    View full-size slide