… calibration as a promising solution because we find that the only major damage in the fine-tuned
model is the biased logits towards fine-tuning classes. In other words, if the fine-tuned …

… issue, we propose a regularized fine-tuning method. Our method introduces two types of
regularization for better calibration: (1) On-… On the other hand, if we use NBAUCC, we will have …

… maps, largely facilitating optimizing the downstream encoder towards the downstream task;
2) SC greatly reduces the semantic drift and the classification bias is thus alleviated when …

… fine-tuning, when the value function is pre-trained with Cal-QL, and show that enforcing
calibration (… [52], but studies the impact of calibration on fine-tuning. We also remark that we …

… For each task, we consider in-domain as well as challenging … 1) when used out-of-the-box,
pretrained models are calibrated … calibration error (ECE), both averaged across 5 fine-tuning …

… While state-of-the-art fine-tuning techniques mainly focus on how to impose regularization
when data … We also introduce the calibration option if source validation data or the confidence-…

… We observe a similar phenomenon when fine-tuning with … When pre-trained models are
adapted to a specific distri… Calibrated ensembles: A simple way to mitigate ID-OOD accuracy …

… fitting than the standard fine-tuning approaches, and can be used even on a very small
calibration set; and (2) … If the system at hand supports more extensive fine-tuning process and the …

… When fine-tuning on ImageNet we also evaluate on … Calibration. While ensembles improve
model calibration (Guo et … As hyperparameters can also have an effect on calibration, we …

… informed design decisions when calibration of the fine-tuned model is a major concern. …
and fine-tuning. In the following subsections, we provide some background on calibration and …