English

Testable Learning with Distribution Shift

Data Structures and Algorithms 2024-05-22 v2 Machine Learning

Abstract

We revisit the fundamental problem of learning with distribution shift, in which a learner is given labeled samples from training distribution DD, unlabeled samples from test distribution DD' and is asked to output a classifier with low test error. The standard approach in this setting is to bound the loss of a classifier in terms of some notion of distance between DD and DD'. These distances, however, seem difficult to compute and do not lead to efficient algorithms. We depart from this paradigm and define a new model called testable learning with distribution shift, where we can obtain provably efficient algorithms for certifying the performance of a classifier on a test distribution. In this model, a learner outputs a classifier with low test error whenever samples from DD and DD' pass an associated test; moreover, the test must accept if the marginal of DD equals the marginal of DD'. We give several positive results for learning well-studied concept classes such as halfspaces, intersections of halfspaces, and decision trees when the marginal of DD is Gaussian or uniform on {±1}d\{\pm 1\}^d. Prior to our work, no efficient algorithms for these basic cases were known without strong assumptions on DD'. For halfspaces in the realizable case (where there exists a halfspace consistent with both DD and DD'), we combine a moment-matching approach with ideas from active learning to simulate an efficient oracle for estimating disagreement regions. To extend to the non-realizable setting, we apply recent work from testable (agnostic) learning. More generally, we prove that any function class with low-degree L2L_2-sandwiching polynomial approximators can be learned in our model. We apply constructions from the pseudorandomness literature to obtain the required approximators.

Keywords

Cite

@article{arxiv.2311.15142,
  title  = {Testable Learning with Distribution Shift},
  author = {Adam R. Klivans and Konstantinos Stavropoulos and Arsen Vasilyan},
  journal= {arXiv preprint arXiv:2311.15142},
  year   = {2024}
}

Comments

To appear in The 37th Annual Conference on Learning Theory (COLT 2024)