Towards Test Time Adaptation via Calibrated Entropy Minimization

Robust models must demonstrate strong generalizability, even amid environmental changes. However, the complex variability and noise in real-world data often lead to a pronounced performance gap between the training and testing phases. Researchers have recently introduced test-time-domain adaptation (TTA) to address this challenge. TTA methods primarily adapt source-pretrained models to a target domain using only unlabeled test data. This study found that existing TTA methods consider only the largest logit as a pseudo-label and aim to minimize the entropy of test time predictions. This maximizes the predictive confidence of the model. However, this corresponds to the model being overconfident in the local test scenarios. In response, we introduce a novel confidence-calibration loss function called Calibrated Entropy Test-Time Adaptation (CETA), which considers the model's largest logit and the next-highest-ranked one, aiming to strike a balance between overconfidence and underconfidence. This was achieved by incorporating a sample-wise regularization term. We also provide a theoretical foundation for the proposed loss function. Experimentally, our method outperformed existing strategies on benchmark corruption datasets across multiple models, underscoring the efficacy of our approach.