nonvacuous

Definition

Describes a theoretical guarantee that actually says something. A generalization bound is vacuous if it exceeds the trivial value — asserting an error rate below 100% is true and useless — so the bar is a bound tight enough to be informative about the model you actually trained. Obtaining nonvacuous bounds for heavily overparameterized networks was a long-standing open problem, and the PAC-Bayes route is where most progress has come from.