Early Stopping without a Validation Set
This addresses the issue of data efficiency for machine learning practitioners by removing the need for a validation split, though it appears incremental as it builds on existing early stopping methods.
The paper tackles the problem of needing a validation set for early stopping by proposing a new criterion based on gradient statistics, which eliminates the need for a held-out validation set. The experiments show this approach works for least-squares and logistic regression, as well as neural networks.
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.