WiTS Virtual - Digital Discrimination: Cognitive Bias in Machine Learning

Digital Discrimination Resources

My team’s work:
http://ibm.biz/codait-trusted-ai
https://developer.ibm.com/blogs/ibm-and-lfai-move-forward-on-trustworthy-and-responsible-ai/
https://www.ibm.com/blogs/policy/facial-recognition-sunset-racial-justice-reforms/

Criminal Recidivism Scoring
http://www.equivant.com/solutions/inmate-classification
https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

Gender Shades/Algorithmic Justice League
http://gendershades.org/
http://www.aies-conference.com/wp-content/uploads/2019/01/AIES-19_paper_223.pdf
https://www.youtube.com/watch?v=Af2VmR-iGkY
https://www.ajlunited.org/fight
https://www.npr.org/2020/06/24/882683463/the-computer-got-it-wrong-how-facial-recognition-led-to-a-false-arrest-in-michig

Ethics in Computer Science
https://www.nytimes.com/2018/02/12/business/computer-science-ethics-courses.html
https://twitter.com/Neurosarda/status/1084198368526680064

Pinterest Skin Tone Search
https://www.engadget.com/2019/01/24/pinterest-skin-tone-search-diversity/

Industry/Government Definitions of Trustworthy AI
https://ec.europa.eu/digital-single-market/en/news/ethics-guidelines-trustworthy-ai
https://wiki.lfai.foundation/display/DL/Trusted+AI+Committee

Transcript

1. Digital Discrimination:
   Cognitive Bias in Machine
   Learning
   Maureen McElaney, Program Manager
   IBM Quantum
   July 28, 2020
   

   View Slide

2. About Me...
   2
   

   View Slide

3. Digital Discrimination:
   Cognitive Bias in Machine
   Learning
   Tweet at me! @Mo_Mack
   

   View Slide

4. 4
   Agenda
   ● Examples of Bias in Machine
   Learning.
   ● Solutions to combat unwanted bias.
   ● Tools to combat unwanted bias.
   ● Resources and how to get involved.
   

   View Slide

5. A cognitive bias is a systematic pattern of
   deviation from norm or rationality in
   judgment.
   People make decisions given their limited
   resources.
   Wilke A. and Mata R. (2012) “Cognitive Bias”, Clarkson University
   @Mo_Mack
   

   View Slide

6. Examples of bias in machine
   learning.
   @Mo_Mack
   

   View Slide

7. @Mo_Mack
   NorthPointe’s
   COMPAS
   Algorithm
   Image Credit: #WOCinTech
   

   View Slide

8. Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
   May 2016 - Northpointe’s COMPAS Algorithm
   http://www.equivant.com/solutions/inmate-
   classification
   

   View Slide

9. May 2016 - Northpointe’s COMPAS Algorithm
   http://www.equivant.com/solutions/inmate-
   classification
   Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
   

   View Slide

10. May 2016 - Northpointe’s COMPAS Algorithm
    http://www.equivant.com/solutions/inmate-
    classification
    Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    

    View Slide

11. May 2016 - Northpointe’s COMPAS Algorithm
    http://www.equivant.com/solutions/inmate-
    classification
    Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    

    View Slide

12. Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    Black Defendant’s Risk Scores
    

    View Slide

13. Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    White Defendant’s Risk Scores
    

    View Slide

14. @Mo_Mack
    BLACK VS. WHITE
    DEFENDANTS
    ○ Falsely labeled black defendants as likely
    of future crime at twice the rate as white
    defendants.
    ○ White defendants mislabeled as low risk
    more than black defendants
    ○ Pegged Black defendants 77% more likely
    to be at risk of committing future violent
    crime
    

    View Slide

15. View Slide

16. @Mo_Mack
    Gender
    Shades Project
    February 2018
    Image Credit: #WOCinTech
    

    View Slide

17. http://gendershades.org/
    

    View Slide

18. “If we fail to make
    ethical and inclusive
    artificial intelligence
    we risk losing gains
    made in civil rights
    and gender equity
    under the guise of
    machine neutrality.”
    - Joy Boulamwini
    @jovialjoy
    

    View Slide

19. http://www.aies-conference.com/wp-content/uploads/2019/01/AIES-19_paper_223.pdf
    

    View Slide

20. @Mo_Mack
    https://www.youtube.com/watch?v=Af2VmR-iGkY
    

    View Slide

21. @Mo_Mack
    

    View Slide

22. View Slide

23. 23
    Agenda
    ● Examples of Bias in Machine
    Learning.
    ● Solutions to combat unwanted
    bias.
    ● Tools to combat unwanted bias.
    ● Resources and how to get involved.
    

    View Slide

24. Solutions?
    What can we do
    to combat bias
    in AI?
    @Mo_Mack
    

    View Slide

25. @Mo_Mack
    EDUCATION IS
    KEY
    Image Credit: #WOCinTech
    

    View Slide

26. https://www.nytimes.com/2018/02/12/business/computer-science-
    ethics-courses.html
    

    View Slide

27. Questions
    posed to
    students
    in these
    courses...
    Is the
    technology
    fair?
    How do you
    make sure
    that the
    data is not
    biased?
    Should
    machines
    be judging
    humans?
    @Mo_Mack
    

    View Slide

28. https://twitter.com/Neurosarda/status/1084198368526680064
    

    View Slide

29. FIX THE
    PIPELINE?
    @Mo_Mack Image Credit: #WOCinTech
    

    View Slide

30. “Cognitive bias in
    machine learning is
    human bias on
    steroids.”
    30
    - Rediet Abebe
    @red_abebe
    

    View Slide

31. https://twitter.com/MatthewBParksSr/status/1133435312921874432
    

    View Slide

32. January 2019 - New Search Feature on...
    https://www.pinterest.com
    Source:
    https://www.engadget.com/2019/01/24/pinterest-skin-tone-search-diversity/
    

    View Slide

33. “By combining the
    latest in machine
    learning and inclusive
    product development,
    we're able to directly
    respond to Pinner
    feedback and build a
    more useful product.”
    33
    - Candice Morgan
    @Candice_MMorgan
    

    View Slide

34. @Mo_Mack
    National and
    Industry
    Standards
    Image Credit: #WOCinTech
    

    View Slide

35. EU Ethics Guidelines for Trustworthy
    Artificial Intelligence
    According to the Guidelines, trustworthy AI should be:
    (1) lawful - respecting all applicable laws and
    regulations
    (2) ethical - respecting ethical principles and values
    (3) robust - both from a technical perspective while
    taking into account its social environment
    Source:
    https://ec.europa.eu/digital-single-market/en/news/ethics-guidelines-trustworthy-ai
    

    View Slide

36. #1 -
    Human
    agency and
    oversight.
    #2 -
    Technical
    robustness
    and safety.
    #3 -
    Privacy and
    data
    governance.
    #4 -
    Transparency
    .
    @Mo_Mack
    #5 -
    Diversity,
    non-
    discrimination
    and fairness.
    #6 -
    Societal and
    environmental
    well-being.
    #7 -
    Accountability
    

    View Slide

37. https://wiki.lfai.foundation/display/DL/Trusted+AI+Committee
    

    View Slide

38. 38
    Agenda
    ● Examples of Bias in Machine
    Learning.
    ● Solutions to combat unwanted bias.
    ● Tools to combat unwanted bias.
    ● Resources and how to get involved.
    

    View Slide

39. @Mo_Mack
    TOOLS TO
    COMBAT BIAS
    Image Credit: #WOCinTech
    

    View Slide

40. Tool #1:
    AI Fairness
    360 Toolkit
    Open Source Library
    @Mo_Mack
    

    View Slide

41. http://aif360.mybluemix.net/
    

    View Slide

42. http://aif360.mybluemix.net/
    

    View Slide

43. Machine Learning
    Pipeline
    In-
    Processing
    Pre-
    Processing
    Post-
    Processing
    Modifying the
    training data.
    Modifying the
    learning
    algorithm.
    Modifying the
    predictions (or
    outcomes.)
    

    View Slide

44. http://aif360.mybluemix.net/
    Demos
    

    View Slide

45. https://github.com/IBM/AIF360
    AI Fairness 360 Toolkit Public Repo
    

    View Slide

46. Tool #2:
    AI Explainability
    360 Toolkit
    Open Source Library
    

    View Slide

47. AIX360 toolkit is an open-source library to help explain AI and machine learning models and
    their predictions. This includes three classes of algorithms: local post-hoc, global post-hoc,
    and directly interpretable explainers for models that use image, text, and structured/tabular
    data. The AI Explainability360 Python package includes a comprehensive set of explainers,
    both at global and local level.
    Toolbox
    Local post-hoc
    Global post-hoc
    Directly interpretable
    AI Explainability
    360
    ↳ (AIX360)
    https://github.com/IBM/AIX360
    http://aix360.mybluemix.net
    THINK 2020 / © 2020 IBM Corporation
    

    View Slide

48. Tackling different ways to explain
    Selected 2018 explainability innovations from IBM Research
    GLOBAL, POST-HOC
    Improving Simple Models with
    Confidence Profiles
    NEURIPS 2018
    LOCAL, POST-HOC
    Explanations Based on the
    Missing: Towards Contrastive
    Explanations with Pertinent
    Negatives
    NEURIPS 2018
    GLOBAL, DIRECTLY INTERPRETABLE
    Boolean Decision Rules via
    Column Generation
    NIPS 2018
    Variational Inference of
    Disentangled Latent Concepts
    from Unlabeled Observations
    ICLR 2018
    INTERACTIVE MODEL VISUALIZATION
    Seq2Seq-Vis: A Visual
    Debugging Tool for
    Sequence-to-Sequence Models
    IEEE VAST 2018
    LOCAL, DIRECTLY
    INTERPRETABLE
    TED: Teaching AI to
    Explain its Decisions
    AIES 2019
    THINK 2020 / © 2020 IBM Corporation
    

    View Slide

49. Three dimensions of explainability
    One explanation does not fit all: There are many ways to explain things
    directly interpretable
    The oldest AI formats, such as decision
    rule sets, decision trees, and decision
    tables are simple enough for people to
    understand. Supervised learning of
    these models is directly interpretable.
    vs. post hoc interpretation
    Start with a black box model and probe
    into it with a companion model to
    create interpretations. The black box
    model continues to provide the actual
    prediction while interpretation improve
    human interactions.
    global (model-level)
    Show the entire predictive model to the
    user to help them understand it (e.g. a
    small decision tree, whether obtained
    directly or in a post hoc manner).
    vs. local (instance-level)
    Only show the explanations associated
    with individual predictions (i.e. what
    was it about the features of this
    particular person that made her loan
    denied).
    static
    The interpretation is simply presented
    to the user.
    vs. interactive (visual analytics)
    The user can interact with
    interpretation.
    

    View Slide

50. Data explanation
    • ProtoDash (Gurumoorthy et al.,
    2019)
    • Disentangled Inferred Prior VAE
    (Kumar et al., 2018)
    Supported explainability algorithms
    Local post-hoc explanation
    • ProtoDash (Gurumoorthy et al.,
    2019)
    • Contrastive Explanations Method
    (Dhurandhar et al., 2018)
    • Contrastive Explanations Method
    with Monotonic Attribute Functions
    (Luss et al., 2019)
    • LIME (Ribeiro et al. 2016, Github)
    • SHAP (Lundberg, et al. 2017, Github)
    Local direct explanation
    • Teaching AI to Explain its Decisions
    (Hind et al., 2019)
    Global direct explanation
    • Boolean Decision Rules via Column
    Generation (Light Edition) (Dash et
    al., 2018)
    • Generalized Linear Rule Models (Wei
    et al., 2019)
    Global post-hoc explanation
    • ProfWeight (Dhurandhar et al., 2018)
    Supported explainability
    metrics
    • Faithfulness (Alvarez-Melis and
    Jaakkola, 2018)
    • Monotonicity (Luss et al., 2019)
    

    View Slide

51. Explain how AI arrived at a prediction
    • Uses contrastive techniques to explain model behavior in the vicinity of the target data
    point.
    • Identifies feature weighting of most and least important features
    • Displays factors that influence a prediction in simple terms.
    • Explanation in terms of the top-K features which played a key role in the prediction. E.g.,
    The loan was rejected because: (1) Credit score=average, (2) Loan Amount>$2M and (3)
    Area=Downtown.
    Explainability
    THINK 2020 / © 2020 IBM Corporation
    

    View Slide

52. Prediction: Partially Granted
    1 2 3 4 5 6 7 8 9
    Input Data Point
    PP PN
    Partially Approved Approved
    Most Frequent Least Frequent
    Number of Married Years
    Contrastive Explanation:
    • PP: If Number of married years was 7 and salary was in the range $190-210K, then
    outcome would have changed to Loan=Approved
    • PN: Even if Number of married years = 3 and salary was in the range $110-130K,
    outcome would have been Loan=Partially Granted
    [90,
    110]
    [70, 90]
    [110,13
    0]
    Most Frequent Least Frequent
    Input Data Point
    PP
    Salary
    PN
    Partially Approved Approved
    [>210]
    [130,15
    0]
    [150,17
    0]
    [170,19
    0]
    [190,21
    0]
    THINK 2020 / © 2020 IBM Corporation
    

    View Slide

53. https://github.com/IBM/AIX360
    AI Explainability 360 Toolkit
    Public Repo
    

    View Slide

54. IBM donates "Trusted AI" projects to
    Linux Foundation AI
    Source: https://www.zdnet.com/article/ibm-donates-trusted-ai-projects-to-linux-foundation-ai/
    

    View Slide

55. Tool #3:
    Model Asset
    eXchange
    Open Source Pre-Trained
    Deep Learning Models
    

    View Slide

56. Step 1: Find a model
    ...that does what you need
    ...that is free to use
    ...that is performant enough
    

    View Slide

57. Step 2: Get the code
    Is there a good implementation available?
    ...that does what you need
    ...that is free to use
    ...that is performant enough
    

    View Slide

58. Step 3: Verify the
    model
    ○ Does it do what you need?
    ○ Is it free to use (license)?
    ○ Is it performant enough?
    ○ Accuracy?
    

    View Slide

59. Step 4: Train the model
    

    View Slide

60. Step 4: Train the model
    

    View Slide

61. Step 5: Deploy your
    model
    ○ Adjust inference code (or write from
    scratch)
    ○ Package inference code, model code, and
    pre-trained weights together
    ○ Deploy your package
    

    View Slide

62. Step 6: Consume your
    model
    

    View Slide

63. Model Asset
    Exchange
    The Model Asset Exchange (MAX) is a one
    stop shop for developers/data scientists to
    find and use free and open source deep
    learning models
    ibm.biz/model-exchange
    

    View Slide

64. ibm.biz/model-exchange
    

    View Slide

65. http://ibm.biz/model-exchange
    Model Asset eXchange (MAX)
    

    View Slide

66. Tool #4:
    Data Asset
    eXchange
    Open Source Data Sets
    

    View Slide

67. http://ibm.biz/data-exchange
    Data Asset eXchange (DAX)
    

    View Slide

68. http://ibm.biz/codait-trusted-ai
    IBM CODAIT Trusted AI Work
    

    View Slide

69. 69
    Agenda
    ● Examples of Bias in Machine
    Learning.
    ● Solutions to combat unwanted bias.
    ● Tools to combat unwanted bias.
    ● Resources and how to get involved.
    

    View Slide

70. https://www.ajlunited.org/fight
    

    View Slide

71. https://www.patreon.com/poetofcode
    

    View Slide

72. Photo by rawpixel on Unsplash
    No matter what it is our
    responsibility to build
    systems that are fair.
    

    View Slide

73. Thank you!
    Resources from this talk:
    https://bit.ly/WiTS-Bias
    Follow me: @Mo_Mack
    

    View Slide