In this paper, we propose a new test-time learning (TTT) method, Synchronizing Tasks for Test-time Training (S4T), to solve the generalization problem of neural networks with domain changes in situations where multiple tasks must be performed. We found that existing TTT methods suffer from the problem of asynchronous task behaviors in which the adaptation steps for optimal performance on multiple tasks do not match each other. S4T's core idea is to predict task relations with domain changes and synchronize tasks during test time. We experimentally demonstrate that it outperforms existing TTT methods on various benchmarks.
