Maintain Plasticity in Long-timescale Continual Test-time Adaptation

TL;DR

This work addresses the problem of sustaining long-term adaptability in continual test-time adaptation (CTTA) by focusing on plasticity, the model's ability to learn from new distributions over time. It introduces Adaptive Shrink-Restore (ASR), a policy that adaptively reinitializes model weights based on the trajectory of label flips, using a shrink-and-restore scheme to preserve prior knowledge while injecting necessary plasticity. The authors demonstrate that plasticity loss correlates with label-flip fluctuations and show that ASR improves mean accuracy and robustness on CCC, CIN-C, and CIN-3DCC benchmarks, while being compatible with various CTTA methods. The proposed approach offers a practical, generalizable solution for sustaining continual adaptation in non-stationary test environments, with potential impact on deployments requiring long-running robustness.

Abstract

Continual test-time domain adaptation (CTTA) aims to adjust pre-trained source models to perform well over time across non-stationary target 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 with preserved plasticity over a long time? The plasticity refers to the model's capability to adjust predictions in response to non-stationary environments continually. In this work, we explore plasticity, this essential but often overlooked aspect of continual adaptation to facilitate more sustained adaptation in the long run. First, we observe that most CTTA methods experience a steady and consistent decline in plasticity during the long-timescale continual adaptation phase. Moreover, we find that the loss of plasticity is strongly associated with the change in label flip. Based on this correlation, we propose a simple yet effective policy, Adaptive Shrink-Restore (ASR), towards preserving the model's plasticity. In particular, ASR does the weight re-initialization by the adaptive intervals. The adaptive interval is determined based on the change in label flipping. Our method is validated on extensive CTTA benchmarks, achieving excellent performance.

Figures (5)

• Figure 1: Illustration of plasticity loss in adaptation: Test data continuously changes from one domain to another, and the dotted line represents the reset point for CTTA methods. Upon resetting, we observe that a CTTA model with re-initialization (represented by the blue line) initially experiences a significant performance drop. However, after adapting to sufficient samples, the model quickly recovers and fits the data better. In contrast, the CTTA model without resetting (represented by the yellow line) shows a reduced ability to adapt with consistently superior performance later in the long run.
• Figure 2: A simple test of different TTA methods on the part of the CCC-medium dataset for plasticity loss. We observe that all tested methods experience some degree of accuracy degradation in the long run compared to the periodically reset model, indicating a decrease in plasticity over time. The green line represents the fixed reset interval, where the model is reinitialized for adaptation. In these tests, EATA and Tent exhibit a greater loss in plasticity, whereas RPL shows a relatively slower degradation. Most importantly, this is a common and significant issue that needs to be addressed in many CTTA methods.
• Figure 3: Model plasticity under different plasticity preservation policy timings: Using RPL rusak2021if, we test four different timings to trigger the reset interval policy. Two re-initialization policies, reset and no-reset, are used here for better comparison. It can be observed that if the reset occurs too early, there is minimal plasticity loss. However, resetting the adaptation model loses all the previous knowledge gained, leading to an initial drop in performance. In these cases, the drop in adaptation accuracy after the reset outweighs the minor plasticity loss in adaptation.
• Figure 4: The framework of our proposed ASR: when continuously adapting in non-stationary domains, reinitialization is performed adaptively to recover lost plasticity in the adaptation process caused by overfit and error accumulation. During weight reinitialization, we use a warm-up strategy called "shrink-restore" to preserve previously learned knowledge of the adapted model while adding necessary plasticity from the source model.
• Figure 5: The proposed adaptive trigger policy is used to determine the suitable point for initialization. When significant fluctuations in adaptation are detected, the model is reinitialized to restore its learning capacity.