Continual Active Adaptation to Evolving Distributional Shifts

Building neural network models that are adaptable to evolving data distributions without suffering catastrophic forgetting is important for real-world deployment in many applications. In real-world setting, the observed data distribution changes over time due to non-stationary environment. In this paper, we consider the problem of evolving covariate shift and propose source-free active adaptation method to fine-tune the neural networks to continually evolving data without catastrophic forgetting. We evaluate the model performance with respect to adaptation as well as forgetting under sequential evolution of data based on fifteen different common corruptions and perturbations from CIFAR10-C related to shift in lighting, weather, noise etc. We demonstrate the proposed method improves model accuracy to the continually evolving data by 21.3% on an average over the different covariate shifts without catastrophic forgetting.