FATA: Focal-Adjusted Test-Time Adaptation Under Data Imbalance

Robust models should exhibit strong performance even in the presence of data imbalance problems. Solutions such as test-time adaptation(TTA) are currently proposed to address the challenge of inconsistent distribution of training and test data. However, most existing TTA methods do not consider the model's performance in the presence of data imbalance. In this paper, we propose a method of optimizing the loss function using focal loss while using a predictive entropy minimization optimization strategy for updating the parameters of the BN-layer affine transform in a cycle, i.e.,Focal-Adjusted Test-Time Adaptation (FATA), to alleviate the problem of data class imbal-ance. We selected the CIFAR10-C and CIFAR100-C datasets as experimental objects and used them to create class imbalance data. Among them, the CIFAR10-C dataset contains ten classes, and we set different data volume shares according to the class labels; the CIFAR100-C dataset contains ten major classes and 100 minor classes, and we set different data volume shares for each major class. The experimental results show that compared with the source pre-trained model and the traditional cross-entropy loss, the optimized method significantly improves the generalization performance and robustness of the model when dealing with the data imbalance problem. The research in this paper provides new ideas and techniques for solving the data imbalance problem, which is of great significance for further improving the performance of deep learning models in practical applications.