Adaptive and Balanced Re-initialization for Long-timescale Continual Test-time Domain Adaptation

TL;DR

The paper addresses the challenge of maintaining model performance in long-timescale continual test-time domain adaptation (CTTA) under non-stationary environments. It introduces Adaptive-and-Balanced Re-initialization (ABR), which leverages the trajectory of label flips as a signal to trigger adaptive, balanced re-initialization of model weights, incorporating a shrink-restore update between source and adapted weights. Key contributions include establishing label flip as a predictor of long-term loss, detailing the adaptive trigger based on the slope of the label-flip trajectory, and demonstrating substantial improvements over prior CTTA methods across CIN-C, CIN-3DCC, and CCC benchmarks, with zero-tuning of reset intervals. The method is practical and can be integrated with existing CTTA approaches to enhance robustness in non-stationary test environments.

Abstract

Continual test-time domain adaptation (CTTA) aims to adjust models so that they can perform well over time across non-stationary environments. While previous methods have made considerable efforts to optimize the adaptation process, a crucial question remains: Can the model adapt to continually changing environments over a long time? In this work, we explore facilitating better CTTA in the long run using a re-initialization (or reset) based method. First, we observe that the long-term performance is associated with the trajectory pattern in label flip. Based on this observed correlation, we propose a simple yet effective policy, Adaptive-and-Balanced Re-initialization (ABR), towards preserving the model's long-term performance. In particular, ABR performs weight re-initialization using adaptive intervals. The adaptive interval is determined based on the change in label flip. The proposed method is validated on extensive CTTA benchmarks, achieving superior performance.

Figures (5)

• Figure 1: Long-timescale CTTA. The source model is adapted to a large number of continually changing domains. The timestep corresponds to the number of test-time samples.
• Figure 2: Performance in long-timescale CTTA. As shown in the sub-figure below, test data continually changes from one domain to another. The classification accuracies of RPL rusak2021selflearning and EATA niu2022efficient on the CCC dataset press2024rdumb are plotted in the upper and middle sub-figures, respectively. As shown in the figure, methods with a reset policy, represented by the blue line, achieve superior performance compared to those without a reset policy, represented by the orange line. Moreover, this advantage becomes more distinct in the later run. The dotted green line represents the reset time point.
• Figure 3: Model long-term performance under different reset timings. Using RPL rusak2021selflearning, we trigger the reset using four randomly chosen timings (the blue line). We also use the no-reset policy for comparison (the orange line). The dotted green line represents the reset time point. We can see from the figure that these four reset timings do not yield better performance, indicating that suitable re-initialization times are needed.
• Figure 4: Label flip and long-term performance in long-timescale CTTA. The black dot box highlights the timing of re-initialization.
• Figure 5: The framework of the proposed Adaptive-and-Balanced Re-initialization (ABR). When the model continuously adapts in non-stationary domains, balanced re-initialization is triggered adaptively to mitigate long-term performance loss. For the balanced re-initialization, we use the strategy called "shrink-restore" to keep parts of previously learned knowledge of the adapted model while restoring knowledge from the source model.