Technical perspective: Why don't today's deep nets overfit to their training data?
Sign Up to like & getrecommendations! Published in 2021 at "Communications of the ACM"
DOI: 10.1145/3446773
Abstract: that a traditional measure, Rademacher complexity, is high for the deep net architecture. Subsequent work has explored the authors’ suggestion that the training algorithm (a variant of gradient descent) plays a powerful role in how… read more here.
Keywords: today deep; overfit training; deep nets; nets overfit ... See more keywords