A call for better unit testing for invariant risk minimisation

TL;DR

This paper critically examines the IRMv1 framework for invariant risk minimization, revealing its instability and scaling issues, and emphasizes the need for rigorous unit testing to improve its reliability and generalization.

Contribution

It provides a controlled study showing IRMv1's potential instability and scaling problems, advocating for better unit testing in IRM research.

Findings

IRMv1 can be unstable under small changes to the optimal regressor

Scaling issues affect IRMv1's performance and stability

Rigorous unit testing is essential for measuring progress in IRM

Abstract

In this paper we present a controlled study on the linearized IRM framework (IRMv1) introduced in Arjovsky et al. (2020). We show that IRMv1 (and its variants) framework can be potentially unstable under small changes to the optimal regressor. This can, notably, lead to worse generalisation to new environments, even compared with ERM which converges simply to the global minimum for all training environments mixed up all together. We also highlight the isseus of scaling in the the IRMv1 setup. These observations highlight the importance of rigorous evaluation and importance of unit-testing for measuring progress towards IRM.