SOURCE-FREE CONTROLLED ADAPTATION OF TEACHERS FOR CONTINUAL TEST-TIME ADAPTATION

In real-world deployments, models often face a sequence of evolving data distributions (continual domain shift). Existing continual test-time adaptation (CTTA) methods often rely on fixed momentum in teacher-student frameworks, which can lead to error accumulation when the model drifts. Furthermore, many state-of-the-art approaches require access to source data or statistics, which is often prohibited by privacy or storage constraints. This paper asks: how can we adapt models continuously without source data while maintaining stability against distribution shifts?

The authors propose DMSE (Dynamic Momentum and Source Estimation), a source-free framework. First, they replace the fixed momentum used in standard mean-teacher updates with a dynamic momentum parameter. This parameter is adjusted based on the prediction entropy of incoming test batches: lower entropy (higher confidence) triggers more aggressive updates from the student to the teacher, while higher entropy (lower confidence) increases the teacher's stability. Second, they eliminate the need for source data by using the weights of the pre-trained classifier as initial class prototypes. These prototypes are then dynamically updated using confident pseudo-labeled samples from the target domain to maintain alignment as the domain evolves.

The study demonstrates that a fixed momentum is sub-optimal because different types of distribution shifts require different degrees of model adaptation. By dynamically modulating the momentum and using classifier-derived prototypes, the DMSE framework effectively reduces error accumulation. Extensive experiments on benchmark datasets show that DMSE outperforms existing state-of-the-art methods, including those that require access to source data, proving that source-free adaptation can be both effective and robust in continual settings.