Combining Diverse Feature Priors
This work addresses the challenge of model robustness and generalization for machine learning practitioners, though it appears incremental as it builds on existing feature prior methods.
The paper tackles the problem of improving model generalization by combining models trained with diverse feature priors, showing that this approach reduces overlapping failure modes and enhances resilience to spurious correlations, leading to better generalization.
To improve model generalization, model designers often restrict the features that their models use, either implicitly or explicitly. In this work, we explore the design space of leveraging such feature priors by viewing them as distinct perspectives on the data. Specifically, we find that models trained with diverse sets of feature priors have less overlapping failure modes, and can thus be combined more effectively. Moreover, we demonstrate that jointly training such models on additional (unlabeled) data allows them to correct each other's mistakes, which, in turn, leads to better generalization and resilience to spurious correlations. Code available at https://github.com/MadryLab/copriors