RaVL: Discovering and Mitigating Spurious Correlations in Fine-Tuned Vision-Language Models

TL;DR

RaVL is a novel method that identifies and reduces spurious correlations in fine-tuned vision-language models by focusing on local image features, significantly improving zero-shot robustness across diverse models and domains.

Contribution

RaVL introduces a region-level clustering and a region-aware loss to discover and mitigate spurious correlations in VLMs, advancing fine-grained robustness techniques.

Findings

RaVL improves spurious correlation discovery by 191% over baselines.

RaVL enhances worst-group accuracy by 8.2%.

Qualitative results confirm effective mitigation of spurious correlations.

Abstract

Fine-tuned vision-language models (VLMs) often capture spurious correlations between image features and textual attributes, resulting in degraded zero-shot performance at test time. Existing approaches for addressing spurious correlations (i) primarily operate at the global image-level rather than intervening directly on fine-grained image features and (ii) are predominantly designed for unimodal settings. In this work, we present RaVL, which takes a fine-grained perspective on VLM robustness by discovering and mitigating spurious correlations using local image features rather than operating at the global image level. Given a fine-tuned VLM, RaVL first discovers spurious correlations by leveraging a region-level clustering approach to identify precise image features contributing to zero-shot classification errors. Then, RaVL mitigates the identified spurious correlation with a novel region-aware loss function that enables the VLM to focus on relevant regions and ignore spurious relationships during fine-tuning. We evaluate RaVL on 654 VLMs with various model architectures, data domains, and learned spurious correlations. Our results show that RaVL accurately discovers (191% improvement over the closest baseline) and mitigates (8.2% improvement on worst-group image classification accuracy) spurious correlations. Qualitative evaluations on general-domain and medical-domain VLMs confirm our findings.