Representation via Representations: Improving Generalization to Unseen Populations
summary
Proposes training an encoder+predictor+decoder with an adversary that forces representations to remove which-study signals so predictors rely on invariant features across hospitals. Uses simulated genomics-like data (linear, interaction, threshold functions) and shows a sweet-spot fairness coefficient: increasing adversarial strength reduces error on a held-out study until over-penalizing hurts. Works well versus logistic and random forest for few training studies, but with many (≈10) training studies random forest sometimes matches or exceeds performance. Emphasizes no target-population data required.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:42Outline
- 1:15Hope for precision medicine
- 2:01Replication of biomedical predictors is worrying
- 4:22Validity of biomedical predictors is worrying
- 5:04Fairness implications
- 5:58Why don't predictors generalize?
- 9:08Learning representations that censor sensitive info Goat learn a mapping from data to representations that
- 10:31Adversarial Learning of Fair Representations
- 13:27Hypothesis
- 14:27Differences with domain adaptation
- 18:53Return to why predictors don't generalize
- 20:27Representation via Representations (test time)
- 20:47Multi-study loss function
- 21:41Synthetic data generation
- 24:44Functions simulated
- 25:44Synthetic data results - linear functions
- 27:44Linear with an interaction term
- 28:27Logical OR of two linear functions
- 29:39Adding many more studies
- 32:06Conclusion
related talks
19mAdversarial Fairness in the Labor Market
42mBeyond Adversarial Approaches on Distributional Robustness
31mIndividual Gaps in Prediction Generalization
13mCan AI design a fairer future?
30mMachine Learning Techniques for Class Imbalances & Adversaries
47mEthical Machine Learning
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
42mFairness in Machine Learning: From Theory to Practice