Responsible AI metrics

Responsible AI Data Management

Maria Prokofieva

Lead ML Engineer

AI project lifecycle

lifecycle

Responsible AI Data Management

Responsible AI project

  • Legally compliant
  • Fair and diverse
  • Practices are transparent, accountable, and secure

 

Model fairness:

  • Fair, unbiased, and has equitable outcomes for everyone

 

lifecycle

Responsible AI Data Management

Protected characteristics

  • Groups likely to be treated unfairly and face discrimination
  • Defined by protected characteristics:
    • Race
    • Ethnicity
    • Gender
    • Socioeconomic background

A group of people

Responsible AI Data Management

Data acquisition

  • Equal outcomes
  • Demographic disparity
  • Laws and regulations

People looking through documents

Responsible AI Data Management

AI in facial recognition

  • High accuracy

BUT

  • Fail to capture specific ethnicities or genders

WHY...

Lack of:

  • Data availability
  • Diversity
  • Representation

 

faces

Responsible AI Data Management

Equal outcomes and demographic disparity

Equal outcomes: benefits are equal across groups

Conditional demographic disparity: differences between groups

  • Use descriptive statistics and distributions to assess data diversity
  • Corrective measures: weighting and balancing
  • Revisit after modeling
  • Keep track of tests

Equity

Responsible AI Data Management

Modeling

  • Equal performance

For example, medical diagnoses:

  • Some more common in protected groups

 

  • Evaluate false negatives, false positives, and accuracy
  • Explainability:
    • Local Interpretable Model-agnostic Explanation (LIME)
    • Shapley Additive Explanation (SHAP)

 

Bullseye with missed shots on target

Responsible AI Data Management

Deployment and monitoring

  • Model drift:
    • Changes to model performances over time

 

  • Monitor distributions
  • Technical performance metrics
  • Adjust model
  • Keep track!
Responsible AI Data Management

Applying metrics

  • Understand the protected characteristics
  • Many more metrics exist!
  • Always consult appropriate legal and domain experts
  • Conduct privacy and security checks

Responsible data dimensions

Responsible AI Data Management

Let's practice!

Responsible AI Data Management

Preparing Video For Download...