An Empirical Study of Invariant Risk Minimization
This is an incremental study testing a theoretical framework for robustness in machine learning, relevant for researchers in domain generalization.
The paper empirically investigates Invariant Risk Minimization (IRM) to understand its performance in learning predictors invariant to spurious correlations, finding that IRMv1 improves with wider spurious correlation variations and works in text classification.
Invariant risk minimization (IRM) (Arjovsky et al., 2019) is a recently proposed framework designed for learning predictors that are invariant to spurious correlations across different training environments. Yet, despite its theoretical justifications, IRM has not been extensively tested across various settings. In an attempt to gain a better understanding of the framework, we empirically investigate several research questions using IRMv1, which is the first practical algorithm proposed to approximately solve IRM. By extending the ColoredMNIST experiment in different ways, we find that IRMv1 (i) performs better as the spurious correlation varies more widely between training environments, (ii) learns an approximately invariant predictor when the underlying relationship is approximately invariant, and (iii) can be extended to an analogous setting for text classification.