Extending meta‑complexity–based agnostic learning beyond P/poly

Determine whether the result of Goldberg and Kabanets (2023)—that easiness of certain meta‑complexity problems implies agnostic learnability of P/poly in polynomial time using random examples over polynomially‑samplable distributions—extends to typical circuit class restrictions such as ACC^0, TC^0, or NC^1. Establish precise conditions under which analogous implications hold for these restricted circuit classes.

Background

Goldberg and Kabanets (2023) connect meta‑complexity to learning by showing that, under certain easiness assumptions, P/poly is agnostically learnable in polynomial time from random examples over polynomially samplable distributions.

The paper raises whether such implications can be generalized from P/poly to more structured, widely studied circuit classes (e.g., ACC^0, TC^0, NC^1), whose learnability is a central topic in computational learning theory.