The workshop brings together domain experts and researchers to facilitate discussions and forge collaborations on problems with spurious correlations, and instability of machine learning models. Models built without accounting for spurious correlations often break when deployed in the wild, despite excellent performance on benchmarks. In particular, models can learn to rely on apparently unnatural or irrelevant features. Such examples abound in recent literature:

1. In detecting lung disease from chest X-rays, models rely on the type of scanner and marks that technicians use in specific hospitals, instead of the physiological signals of the disease [1, 2]

2. In Natural Language Processing, when reasoning whether a premise entails a hypothesis, models rely on the number of shared words rather than the subject’s relationship with the object. [3]

3. In precision medicine, polygenic risk scores for diseases like diabetes and breast cancer rely on genes prevalent mainly in people of European ancestry, and are not as accurate in other populations [4].

Extensive work on resolving problems akin to spurious correlations has sprung up in several communities. These include works on invariance constraints and graph-based methods rooted in Causality, methods to avoid discrimination of compromised subgroups in Algorithmic Fairness, and stress testing procedures to discover unexpected model dependencies in reliable ML. Yet there is little consensus on best practices, useful formal frameworks, rigorous evaluations of models, and fruitful avenues for the future.

We invite work addressing all aspects of ML in the presence of spurious correlations, from formalization to deployment.