Self-Supervised Dynamic Networks for Covariate Shift Robustness
This work addresses robustness issues in AI applications for scenarios with common input variations, representing an incremental advancement over existing self-supervised approaches.
The paper tackles the problem of covariate shift in supervised learning, where input variations like noise or illumination degrade test-time accuracy, by proposing Self-Supervised Dynamic Networks (SSDN), which uses a self-supervised network to predict weights for the main network, resulting in significant performance improvements over comparable methods in image classification tasks.
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.