Robust fine-tuning aims to ensure performance on out-of-distribution (OOD) samples, which is sometimes compromised by pursuing adaptation on in-distribution (ID) samples. However, another criterion for reliable machine learning -- confidence calibration has been overlooked despite its increasing demand for real-world high-stakes applications, e.g., autonomous driving. We raise concerns about the calibration of fine-tuned vision-language models (VLMs) under distribution shift by showing that naive fine-tuning and even state-of-the-art robust fine-tuning hurt the calibration of pre-trained VLMs, especially on OOD datasets. We first show the OOD calibration error is bounded from above with ID calibration errors and domain discrepancy between ID and OOD. From this analysis, we propose CaRot, a calibrated robust fine-tuning method that incentivizes ID calibration and robust prediction across domains to reduce the upper bound of OOD calibration error. Extensive experiments on three types of distribution shifts (natural, synthetic, and adversarial) on ImageNet-1K classification demonstrate the effectiveness of CaRot across diverse environments. We justify the empirical success of CaRot through our theoretical analysis.

翻译：暂无翻译