Early Stopping without a Validation Set

TL;DR

This paper introduces a new early stopping method that eliminates the need for a validation set by using local gradient statistics, applicable to various models including neural networks.

Contribution

It proposes a novel early stopping criterion based on gradient statistics that removes the reliance on a validation set for model training.

Findings

Effective in least-squares and logistic regression

Works well with neural networks

Achieves comparable performance without validation set

Abstract

Early stopping is a widely used technique to prevent poor generalization performance when training an over-expressive model by means of gradient-based optimization. To find a good point to halt the optimizer, a common practice is to split the dataset into a training and a smaller validation set to obtain an ongoing estimate of the generalization performance. We propose a novel early stopping criterion based on fast-to-compute local statistics of the computed gradients and entirely removes the need for a held-out validation set. Our experiments show that this is a viable approach in the setting of least-squares and logistic regression, as well as neural networks.