Chen Liang ; Zhang Yong ; Song Yibing ; Shan Ying ; Liu Lingqiao

Abstract

The main challenge in domain generalization (DG) is to handle thedistribution shift problem that lies between the training and test data. Recentstudies suggest that test-time training (TTT), which adapts the learned modelwith test data, might be a promising solution to the problem. Generally, a TTTstrategy hinges its performance on two main factors: selecting an appropriateauxiliary TTT task for updating and identifying reliable parameters to updateduring the test phase. Both previous arts and our experiments indicate that TTTmay not improve but be detrimental to the learned model if those two factorsare not properly considered. This work addresses those two factors by proposingan Improved Test-Time Adaptation (ITTA) method. First, instead of heuristicallydefining an auxiliary objective, we propose a learnable consistency loss forthe TTT task, which contains learnable parameters that can be adjusted towardbetter alignment between our TTT task and the main prediction task. Second, weintroduce additional adaptive parameters for the trained model, and we suggestonly updating the adaptive parameters during the test phase. Through extensiveexperiments, we show that the proposed two strategies are beneficial for thelearned model (see Figure 1), and ITTA could achieve superior performance tothe current state-of-the-art methods on several DG benchmarks. Code isavailable at https://github.com/liangchen527/ITTA.