PAC learnability of a concept class under non-atomic measures: a problem by Vidyasagar
Vladimir Pestov

TL;DR
This paper characterizes when a concept class is PAC learnable under all non-atomic measures, introducing a new combinatorial parameter called VC dimension modulo countable sets, which generalizes classical VC dimension.
Contribution
It provides a necessary and sufficient condition for distribution-free PAC learnability under non-atomic measures using the new VC dimension modulo countable sets, extending classical VC theory.
Findings
Finiteness of classical VC dimension is sufficient but not necessary for learnability.
Learnability under non-atomic measures does not imply uniform Glivenko-Cantelli property.
Introduces VC dimension modulo countable sets as a key parameter for characterization.
Abstract
In response to a 1997 problem of M. Vidyasagar, we state a necessary and sufficient condition for distribution-free PAC learnability of a concept class under the family of all non-atomic (diffuse) measures on the domain . Clearly, finiteness of the classical Vapnik-Chervonenkis dimension of is a sufficient, but no longer necessary, condition. Besides, learnability of under non-atomic measures does not imply the uniform Glivenko-Cantelli property with regard to non-atomic measures. Our learnability criterion is stated in terms of a combinatorial parameter which we call the VC dimension of modulo countable sets. The new parameter is obtained by ``thickening up'' single points in the definition of VC dimension to uncountable ``clusters''. Equivalently, $\VC(\mathscr C\modd\omega_1)\leq…
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMachine Learning and Algorithms · Computability, Logic, AI Algorithms · Algorithms and Data Compression
