Image credit: UnsplashAbstract
In this work, we study the generalization properties of infinite ensembles of infinitely-wide neural networks. Amazingly, this model family admits tractable calculations for many information-theoretic quantities. We report analytical and empirical investigations in the search for signals that correlate with generalization.
Type
Publication
Proceedings of the Symposium on Advances in Approximate Bayesian Inference
