Scribd Logo
Upload

Welcome to Scribd!

  • 23 views

    Question 1 : Fdic 1

    Uploaded by

    ShidehShamsAmiri

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Question 1 : Fdic 1

    Uploaded by

    ShidehShamsAmiri
    23 views7 pages

    Original Title

    FDIC

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    23 views7 pages

    Question 1 : Fdic 1

    Uploaded by

    ShidehShamsAmiri

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 7

    FDIC

    Oversight of Third Parties


    Financial institutions may opt to use AI developed by third parties, rather than
    develop the approach internally. Existing agency guidance (as noted in the
    Appendix) describes information and risks that may be relevant to financial
    institutions when selecting third-party approaches (including ones using AI and
    sets out principles for the validation of such third-party approaches.

    Dynamic Updating
    A particular characteristic of some AI is the ability for it to learn or evolve over
    time, especially as it captures new training data. Over time, this could result in
    drift (i.e., the AI approach could change) as it learns from the new data. This can
    present challenges for validating, monitoring, tracking, and documenting the AI
    approach, including for persons conducting an independent review. It may be
    important to understand whether an AI approach that was independently
    reviewed initially has significantly evolved over time (e.g., using an influx of new
    data). Dynamic updating can also affect how results are tracked over time. For
    example, initial performance thresholds chosen to monitor the approach could
    become less meaningful if the AI approach has significantly changed to focus on
    different target outcomes. Similar risks can arise with AI approaches that are not
    updated as their context evolves, since they are more closely tuned to their
    training data. For example, AI approaches that are validated in one circumstance
    may not perform well in another, and an independent review conducted in a
    previous context may no longer be accurate in new circumstances.

    Question 1
    How do financial institutions identify and manage risks relating to AI
    explainability? What barriers or challenges for explainability exist for developing,

    FDIC 1
    adopting, and managing AI?

    💡 There are three main approaches for developing explainable machine


    learning models.

    1. Using explainable models: Choosing explainable machine learning


    models such as Logistic Regression, Generalized Linear Models,
    Generalized Additive Models, Decision Trees. The disadvantage of using
    interpretable models is that it will limit the developers to use specific
    kinds of machine learning models and decrease in predictive
    performance compared to other machine learning models.

    2. Model-Agnostic Interpretable methods: Separates the explanation


    from the choice machine learning model. The most common model-
    agnostic methods are Partial Dependence Plots, Permutation Feature
    Importance, Local Surrogate Methods LIME, SHapely additive
    explanation.

    Currently, there is a clear trade-off between the performance of


    machine learning models and the ability to produce explainable and
    interpretable predictions. The black-box models, including deep learning
    and model stacking, and model ensembles can produce accurate
    predictions while interpretable models such as logistic regression fail in
    capturing feature interactions and non-linear relationships.

    Question 2:

    How do financial institutions use post-hoc methods to assist in evaluating


    conceptual soundness? How common are these methods? Are there limitations of
    these methods (whether to explain an AI approach’s overall operation or to explain
    a specific prediction or categorization)? If so, please provide details on such
    limitations.

    FDIC 2
    💡 Post-hoc interpretability methods can be used in model validation. The
    explanation generated by these methods could help to examine whether
    a machine learning model has employed the true evidences instead of
    biases in training data. For examples, explanations can be used to
    identify whether models have utilized features that are proxy for gender
    or ethnic biases to make decision.

    Some limitations includes high computational cost of generating


    explanations, instability of explanation in sampling based model-
    agnostic methods such as LIME, ability to intentionally generate
    misleading explanations.

    Question 3

    For which uses of AI is lack of explainability more of a challenge? Please describe


    those challenges in detail. How do financial institutions account for and manage
    the varied challenges and risks posed by different uses?

    💡 The financial services industry has unique explainability and fairness


    challenges arising from compliance and ethical considerations in credit
    decisions. Compliance with fair lending laws and producing specific
    reasons for denying credit is one of the most important use cases of
    machine learning explainability. Typically financial institutions leverage
    model-agnostic methods to generate explanations for individual
    predicitons.

    Question 4
    How do financial institutions using AI manage risks related to data quality and
    data processing? How, if at all, have control processes or automated data quality
    routines changed to address the data quality needs of AI? How does risk

    FDIC 3
    management for alternative data compare to that of traditional data? Are there
    any barriers or challenges that data quality and data processing pose for
    developing, adopting, and managing AI? If so, please provide details on those
    barriers or challenges.

    💡 Data quality monitoring automatically monitors machine learning models


    in production and notifies you when data quality issues arise. Machine
    Learning models in production have to make predictions on real-life data
    that is not carefully curated like most training datasets. If the statistical
    nature of the data that your model receives while in production drifts
    away from the nature of the baseline data it was trained on, the model
    begins to lose accuracy in its predictions. The financial institution can
    use in-house and off-the-shelf data quality monitoring products
    available at online cloud providers which can ensure the quality and
    accuracy of input data used in machine learning models.

    Question 5

    Are there specific uses of AI for which alternative data are particularly effective?

    💡 ?

    Question 6
    Overfitting “Overfitting” can occur when an algorithm “learns” from idiosyncratic
    patterns in the training data that are not representative of the population as a
    whole. Overfitting is not unique to AI, but it can be more pronounced in AI than
    with traditional models. Undetected overfitting could result in incorrect predictions
    or categorizations.
    How do financial institutions manage AI risks relating to overfitting? What barriers
    or challenges, if any, does overfitting pose for developing, adopting, and
    managing AI? How do financial institutions develop their AI so that it will adapt to
    new and potentially different populations (outside of the test and training data)?

    FDIC 4
    💡 The bias-variance tradeoff is important in all machine learning models
    and it is not limited to financial institutions. Theoretical and applied
    methodologies to address bias-variance trade-off is well known:
    - Adding more training data
    - Add regularization
    - Early stopping
    - Feature selection
    - Decreasing model size and model capacity

    In order to adapt the machine learning model to a dynamic population,


    developers must consider model drift, data drift, and concept drift. To
    address this issue the models must be periodically re-fit to new data
    address these issues.

    Question 8
    How do financial institutions manage AI risks relating to dynamic updating?
    Describe any barriers or challenges that may impede the use of AI that involve
    dynamic updating. How do financial institutions gain an understanding of whether
    AI approaches producing different outputs over time based on the same inputs are
    operating as intended?

    FDIC 5
    💡 Data processing and feature engineering are important parts of any
    machine learning life cycle. The importance of robust infrastructure to
    ensure fresh and accurate features increases when machine learning
    models need frequent updating. In recent years the 'Feature Store'
    concept in data architecture that motivated by these challenges. Some
    of the benefits of feature stores for machine learning models are:

    1. Track and share features between data scientists including a version-


    control repository
    2. Process and curate feature values while preventing data leakage
    3. Ensure parity between training and inference data systems
    4. Serve features for ML-specific consumption profiles including model
    training, batch and real-time predictions
    5. Accelerate ML innovation by reducing the data engineering process
    from months to days
    6. Monitor data quality to rapidly identify data drift and pipeline errors
    7. Empower legal and compliance teams to ensure compliant use of data
    8. Bridging the gap between data scientists and data & ML engineers
    9. The lower total cost of ownership through automation and
    simplification
    10. Faster Time-To-Market for new model-driven products
    11. Improved model accuracy: the availability of features will improve
    model performance
    12. Improved data quality via data →feature → model lineage

    Machine learning models produce outputs that reflect the correlations


    between input variables and target variables in the training set.
    Producing different outputs over time is not inherently problematic if it is
    reflective of the true relationship between inputs and targets.
    Developers should implement monitoring processes to detect any drift in
    data or concept drift.

    Question 10

    FDIC 6
    Please describe any particular challenges or impediments financial institutions
    face in using AI developed or provided by third parties and a description of how
    financial institutions manage the associated risks.

    Please provide detail on any challenges or impediments. How do those challenges


    or impediments vary by financial institution size and complexity?

    FDIC 7

    You might also like