Why you don't overfit, and don't need Bayes if you only train for one epoch

Laurence Aitchison

Here, we show that in the data-rich setting where you only train on each datapoint once (or equivalently, you only train for one epoch), standard "maximum likelihood" training opti…