Understanding the failure modes of out-of-distribution generalization

This repository contains the code used for the following paper:

"Understanding the failure modes of out-of-distribution generalization"
Vaishnavh Nagarajan, Anders Andreassen, Behnam Neyshabur.
International Conference on Learning Representations (ICLR), 2021.

Disclaimer: this is not an official Google product.

Experiments

The code in experiments.ipynb allows you to interactively run experiments from Appendix E. These are experiments demonstrating the effects of geometric skews and statistical skews in a real-world dataset in the context of a max-margin classifier and a gradient-descent-trained linear classifier.