Agnostic extension of contaminated PAC learning results

Extend the PAC learning results in the iterative contamination model to the agnostic setting, where the true concept may not belong to the hypothesis class and labels may be noisy beyond the model-induced contamination, and establish corresponding learning guarantees.

Background

The PAC learning analysis in the paper is conducted in the realizable setting, proving upper bounds for two algorithms and a lower bound showing repeated ERM can stall under contamination.

The authors explicitly list expanding these results to the agnostic setting as an open problem, which would require handling irreducible label noise and model misspecification in addition to recursive contamination.