Measures of relatedness in domain adaptation and related problems.
summary
Proposes a concrete quantity called the transfer exponent and a broader modulus of continuity that map source excess risk to target excess risk, proves minimax rates that combine nP and nQ (rates ~ (nP^{1/ρ}+nQ)^{-1/(2−β)}), shows adaptive procedures via penalized ERM or weak confidence set reductions, proves unlabeled data gives no worst‑case gain, and gives impossibility results: multi‑source and model‑selection settings can forbid adaptive gains unless extra structure (causal or representation) is assumed.
AI-written summary · about
chapters (6)
- 0:00Transfer learning basics
- 13:21Relatedness measures limits
- 26:05Transfer exponent measure
- 49:32Modulus of continuity approach
- 1:08:58Theory limits and negatives
- 1:14:10Q&A
related talks
54mA Discrepancy-Based Theory of Adaptation
41mTransfer learning via local convergence rates of the nonparametric least squares estimator
45mTBD
42mBypassing the Impossibility of Online Learning Thresholds: Unbounded Losses and Transductive Priors
45mMulticalibration, Universal Adaptability and Causality
29mOn Learning Distributions from their Samples
01h 23mGeneralization II
34mDistributional Robustness, Learning, and Empirical Likelihood