Timezone: »

Combining Diverse Feature Priors
Saachi Jain · Dimitris Tsipras · Aleksander Madry

Tue Jul 19 03:30 PM -- 05:30 PM (PDT) @ Hall E #508

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.

Author Information

Saachi Jain (Massachusetts Institute of Technology)
Dimitris Tsipras (Stanford University)
Aleksander Madry (MIT)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors