Reviewed Version (pdf): https://openreview.net/references/pdf?id=jTQFicoaG5
Keywords: understanding deep learning, generalization, interpolating methods, empirical investigation
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 understanding of interpolating classifiers.
One-sentence Summary: We introduce a new notion of generalization ("Distributional Generalization"), to characterize empirical observations of interpolating classifiers.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics