Hot training errors predict cold test errors in the interpolation regime
Erfan Mirzaei ⋅ Andreas Maurer ⋅ Massimiliano Pontil
Abstract
We study the generalization of iterative learning algorithms, which follow the negative gradient of the empirical error and inject noise at every step. Identifying the quantity of injected randomness with temperature, we demonstrate theoretically and experimentally that the training errors of hotter limiting distributions are predictive of the test errors of colder limiting distributions.
Chat is not available.
Successful Page Load