Combining Diverse Feature Priors

Saachi Jain, Dimitris Tsipras, Aleksander Madry

Keywords: robustness, spurious correlations, feature priors

Abstract: 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 explicit 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.

One-sentence Summary: We explore how a diverse set of feature priors can be leveraged to improve model generalization.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2110.08220/code)