I think what’s happening is that memorizing the training data takes only a small portion of the capacity of the model. So the remainder starts getting used to build better and better models of the…

Posters: eshear. eshear

2 likes recorded by the source corpus.