Self-Supervised Dynamic Networks for Covariate Shift Robustness

TL;DR

This paper introduces Self-Supervised Dynamic Networks (SSDN), a method that dynamically adjusts model weights at test-time using self-supervision to improve robustness against covariate shifts like noise and illumination changes.

Contribution

The paper proposes SSDN, a novel input-dependent dynamic network that predicts main network weights at test-time using self-supervision, enhancing covariate shift robustness.

Findings

SSDN significantly outperforms comparable methods on image classification under covariate shifts.

The approach demonstrates conceptual and empirical advantages in handling input nuisances.

Self-supervision enables the network to adapt dynamically to test-time covariate changes.

Abstract

As supervised learning still dominates most AI applications, test-time performance is often unexpected. Specifically, a shift of the input covariates, caused by typical nuisances like background-noise, illumination variations or transcription errors, can lead to a significant decrease in prediction accuracy. Recently, it was shown that incorporating self-supervision can significantly improve covariate shift robustness. In this work, we propose Self-Supervised Dynamic Networks (SSDN): an input-dependent mechanism, inspired by dynamic networks, that allows a self-supervised network to predict the weights of the main network, and thus directly handle covariate shifts at test-time. We present the conceptual and empirical advantages of the proposed method on the problem of image classification under different covariate shifts, and show that it significantly outperforms comparable methods.