Understanding the Failure Modes of Out-of-Distribution Generalization

TL;DR

This paper investigates why machine learning models fail to generalize out-of-distribution, identifying fundamental geometric and statistical failure modes through theoretical analysis and dataset modifications.

Contribution

It uncovers two fundamental failure modes in OOD generalization caused by spurious correlations, supported by theoretical analysis and experimental dataset modifications.

Findings

Identifies geometric and statistical failure modes in OOD generalization.

Demonstrates these failure modes can be isolated in neural network training.

Provides dataset modifications to study failure modes in practice.

Abstract

Empirical studies suggest that machine learning models often rely on features, such as the background, that may be spuriously correlated with the label only during training time, resulting in poor accuracy during test-time. In this work, we identify the fundamental factors that give rise to this behavior, by explaining why models fail this way {\em even} in easy-to-learn tasks where one would expect these models to succeed. In particular, through a theoretical study of gradient-descent-trained linear classifiers on some easy-to-learn tasks, we uncover two complementary failure modes. These modes arise from how spurious correlations induce two kinds of skews in the data: one geometric in nature, and another, statistical in nature. Finally, we construct natural modifications of image classification datasets to understand when these failure modes can arise in practice. We also design experiments to isolate the two failure modes when training modern neural networks on these datasets.