A call for better unit testing for invariant risk minimisation

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. Invariant risk minimization (IRM) (Arjovsky et al., 2020) is a machine learning framework whose primary goal is to learn invariances across multiple training environments.