Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models

2024-10-03

AI Summary

This work identifies a confidence calibration trade-off between base and novel classes in vision-language models (e.g., CLIP) after prompt tuning: novel classes become overconfident while base classes suffer from underconfidence—stemming from divergent shifts in label text embeddings. To address this, we propose Dynamic Outlier Regularization (DOR), a method that selectively constrains only the text feature shifts of novel classes while preserving full optimization freedom for base-class text embeddings. DOR achieves selective calibration in the text embedding space via dynamic vocabulary sampling and feature-bias minimization regularization. Evaluated across multiple few-shot transfer benchmarks, DOR significantly reduces both Expected Calibration Error (ECE) and Brier score for both base and novel classes, outperforming CoOp and KgCoOp in calibration fidelity without compromising classification accuracy.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Confidence calibration is critical for the safe deployment of machine learning models in the real world. However, such issue in vision-language models like CLIP, particularly after fine-tuning, has not been fully addressed. In this work, we demonstrate that existing prompt tuning methods usually lead to a trade-off of calibration between base and new classes: the cross-entropy loss in CoOp causes overconfidence in new classes by increasing textual label divergence, whereas the regularization of KgCoOp maintains the confidence level but results in underconfidence in base classes due to the improved accuracy. Inspired by the observations, we introduce Dynamic Outlier Regularization (DOR) to ensure the confidence calibration on both base and new classes after fine-tuning. In particular, we propose to minimize the feature deviation of novel textual labels (instead of base classes) sampled from a large vocabulary. In effect, DOR prevents the increase in textual divergence for new labels while easing restrictions on base classes. Extensive experiments demonstrate that DOR can enhance the calibration performance of current fine-tuning methods on base and new classes.

Problem

• Addressing miscalibration in vision-language models after fine-tuning
• Balancing confidence calibration between base and new classes
• Improving calibration via Dynamic Outlier Regularization (DOR)

Innovation

• Dynamic Outlier Regularization for calibration
• Minimize novel label feature deviation
• Balance base and new class confidence