MIT 6.S191: AI Bias and Fairness

  • humans label and categorize sensory inputs to simplify groups
  • we tend to notice atypical categories
  • bias and stereotypes happen when we make decisions based on particular categories/labels
  • Algorithmic Bias is tied to LLM image classification and recognition based on data its fed
    • predicted classes define how models are classifying people and making decisions
  • Bias in AI Stages (basic list of bias taxonomy)
    • Data: respect to labels
    • Model: tracking, uncertainty and metrics
    • Training: perpetuated loops for training
    • Evaluation: checking subgroups in QA
    • Deployment: models performing differently in context than from testing
    • Interpretation: Human analysis error

Data Bias

  • Selection Bias: data is not randomized properly (eg. classification imbalance)
  • Reporting Bias: information given is not likely or un-vetted (eg. news reports)
  • Sample Bias: particular instances are checked more often (eg. hair and skin color for facial recognition)

Interpretation Bias

  • How humans can perpetuate these problems of bias

  • Correlation Fallacy: correlation does not equal causation

  • Overgeneralization: general conclusions tied to a small data set

  • Automation: trusting AI overlooks versus human evaluation

  • distribution shift - training models on a curated data set, but undermining its training on other areas (eg. profiles of mugs, but not training on other angles/full mugs)

    • eg. training model on typical western grocery spices versus typical eastern grocery spices → leads to blind spots of bias towards the most prominent data set
  • Creating an improper ratio of data points creates a class imbalance

    • Model is trying to optimize its classification accuracy and will pass based on bias of larger data set
    • eg. Health care. Brain tumor is relatively rare in a large group at 0.003% so model may optimize to reach this number with all data sets even if untrue.

Mitigating class imbalance

  • batch selection - incremental updates per batch are made to classifier during learning
  • example weighting - higher frequency results are weighed less than rare ones in order to keep their considerations relatively equal.
    • latent distribution allows for heavier weighted results to move bell curve up to be less radical
  • checking bias/fairness can be done through understanding classification evaluation
    • 1. desegregated evaluation - evaluate performance based on multiple subgroups (eg. color, shape, size against others of the same class)
    • 2. intersectional evaluation - compare desegregated groups against each other (eg. color AND shape)