Test-Time Training with Self-Supervision for Generalization under Distribution Shifts
This addresses the challenge of generalization for predictive models when training and test data distributions differ, with incremental contributions to robustness in image classification.
The paper tackles the problem of improving model performance under distribution shifts by introducing Test-Time Training, which updates model parameters on each unlabeled test sample using self-supervision before prediction, leading to improvements on diverse image classification benchmarks.
In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a prediction. This also extends naturally to data in an online stream. Our simple approach leads to improvements on diverse image classification benchmarks aimed at evaluating robustness to distribution shifts.