Combining Diverse 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 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

References in corpus (10)

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• Certified Adversarial Robustness via Randomized Smoothing
• Adversarial Examples Are a Natural Consequence of Test Error in Noise
• In Defense of Pseudo-Labeling: An Uncertainty-Aware Pseudo-label Selection Framework for Semi-Supervised Learning
• Noise or Signal: The Role of Image Backgrounds in Object Recognition
• Frustratingly Simple Domain Generalization via Image Stylization
• Informative Dropout for Robust Representation Learning: A Shape-bias Perspective
• Are Perceptually-Aligned Gradients a General Property of Robust Classifiers?
• Diverse Ensembles Improve Calibration
• Out-distribution aware Self-training in an Open World Setting