A Discrepancy-Based Theory of Adaptation
summary
Introduces labeled and unlabeled discrepancy as the central divergence, derives finite-sample estimation via Rademacher complexity and gives explicit bounds, and analyzes sample reweighting via empirical discrepancy minimization (including a difference-of-convex formulation for squared loss). Shows the generalization bound balances empirical weighted loss, sum and L2 norm of weights, and discrepancy; implements a reweighting algorithm and reports experiments where it outperforms several baselines and is empirically robust when labeled target data is scarce.
AI-written summary · about
chapters (12)
- 0:00Workshop introduction
- 5:03Speaker introduction
- 6:58Motivation and overview
- 11:24Defining the adaptation problem
- 14:08Divergence and discrepancy
- 17:02Properties of discrepancy
- 22:25Bounds and estimations
- 28:17Advanced theoretical bounds
- 34:40Wasserstein and other metrics
- 37:20Theoretical guarantees
- 43:28Limitations and reweighting
- 45:40Algorithms and conclusions
related talks
01h 15mMeasures of relatedness in domain adaptation and related problems.
01h 23mGeneralization II
01h 01mDistribution Shift as Underspecification, and What We Might Do About It
01h 03mOverview of Statistical Learning Theory Part 2
27mPreventing Overfitting in Adaptive Data Analysis via Stability
45mMulticalibration, Universal Adaptability and Causality
01h 01mRobustness, Stochastics, Uncertainty 2
28mData Driven Convergence in Statistical Estimators