Assess binomial-bagging combinations of approximate RERMs for robust generalization

Determine whether combining approximate robust empirical risk minimizers produced by adversarial training through methods inspired by binomial bagging improves robust generalization or mitigates robust overfitting in modern artificial-intelligence systems.

Background

The theoretical results use robust empirical risk minimizers as a black-box component and show that binomial bagging of such predictors achieves optimal distribution-free PAC rates. In practical machine-learning systems, however, robust empirical risk minimization is typically approximated by adversarial training rather than solved exactly.

The paper identifies an unresolved practical and theoretical question: whether aggregating these approximate adversarial-training models using binomial-bagging-inspired procedures can transfer the paper’s benefits to practical systems, particularly by improving robust generalization or reducing robust overfitting.