Paper ID: 2311.01723

Towards Calibrated Robust Fine-Tuning of Vision-Language Models

Changdae Oh, Hyesu Lim, Mijoo Kim, Dongyoon Han, Sangdoo Yun, Jaegul Choo, Alexander Hauptmann, Zhi-Qi Cheng, Kyungwoo Song

Improving out-of-distribution (OOD) generalization through in-distribution (ID) adaptation is a primary goal of robust fine-tuning methods beyond the naive fine-tuning approach. However, despite decent OOD generalization performance from recent robust fine-tuning methods, OOD confidence calibration for reliable machine learning has not been fully addressed. This work proposes a robust fine-tuning method that improves both OOD accuracy and calibration error in Vision Language Models (VLMs). Firstly, we show that both types of errors have a shared upper bound consisting of two terms of ID data: 1) calibration error and 2) the smallest singular value of the input covariance matrix. Based on this insight, we design a novel framework that conducts fine-tuning with a constrained multimodal contrastive loss enforcing a larger smallest singular value, which is further aided by the self-distillation of a moving averaged model to achieve well-calibrated prediction. Starting from an empirical validation of our theoretical statements, we provide extensive experimental results on ImageNet distribution shift benchmarks that demonstrate the effectiveness of our method.

Submitted: Nov 3, 2023