ARXIV2017

We propose a strategy to assess overfitting using gradient and weight statistics in neural networks. This makes the use of a separate validation set unnecessary.