Distributional Generalization: A New Kind of Generalization

arXiv2009.08092MaRDI QIDQ6349293

Preetum Nakkiran, Yamini Bansal

Publication date: 17 September 2020

Abstract: We introduce a new notion of generalization -- Distributional Generalization -- which roughly states that outputs of a classifier at train and test time are close *as distributions*, as opposed to close in just their average error. For example, if we mislabel 30% of dogs as cats in the train set of CIFAR-10, then a ResNet trained to interpolation will in fact mislabel roughly 30% of dogs as cats on the *test set* as well, while leaving other classes unaffected. This behavior is not captured by classical generalization, which would only consider the average error and not the distribution of errors over the input domain. Our formal conjectures, which are much more general than this example, characterize the form of distributional generalization that can be expected in terms of problem parameters: model architecture, training procedure, number of samples, and data distribution. We give empirical evidence for these conjectures across a variety of domains in machine learning, including neural networks, kernel machines, and decision trees. Our results thus advance our empirical understanding of interpolating classifiers.

Has companion code repository: https://github.com/kuangliu/pytorch-cifar

This page was built for publication: Distributional Generalization: A New Kind of Generalization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6349293)