Debiasing CLIP: Interpreting and Correcting Bias in Attention Heads

2025-05-23

AI Summary

Multimodal models like CLIP exhibit strong zero-shot capabilities but suffer from inherent biases due to spurious correlations—e.g., between background cues or gender attributes and target classes—arising from attention mechanisms. To address this, we propose Locate-Then-Correct (LTC), a contrastive debiasing framework. First, leveraging mechanism-driven interpretability analysis of attention heads, LTC precisely identifies bias-correlated heads. Then, it applies zero-shot, fine-tuning-free interventions: targeted head pruning and task-relevant head amplification, coupled with orthogonal projection for feature decoupling and fusion. To our knowledge, LTC is the first post-hoc, training-free zero-shot debiasing method. On multiple intrinsic bias benchmarks, it improves worst-group accuracy by over 50%, substantially outperforming existing non-training-based approaches. Comprehensive visualization and ablation studies validate both the accuracy of bias-head identification and the efficacy of the correction mechanism.

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Multimodal models like CLIP have gained significant attention due to their remarkable zero-shot performance across various tasks. However, studies have revealed that CLIP can inadvertently learn spurious associations between target variables and confounding factors. To address this, we introduce extsc{Locate-Then-Correct} (LTC), a contrastive framework that identifies spurious attention heads in Vision Transformers via mechanistic insights and mitigates them through targeted ablation. Furthermore, LTC identifies salient, task-relevant attention heads, enabling the integration of discriminative features through orthogonal projection to improve classification performance. We evaluate LTC on benchmarks with inherent background and gender biases, achieving over a $>50%$ gain in worst-group accuracy compared to non-training post-hoc baselines. Additionally, we visualize the representation of selected heads and find that the presented interpretation corroborates our contrastive mechanism for identifying both spurious and salient attention heads. Code available at https://github.com/wj210/CLIP_LTC.

Problem

• Identifies spurious attention heads in CLIP models
• Mitigates biases via targeted ablation and orthogonal projection
• Improves worst-group accuracy in biased benchmarks

Innovation

• Identifies spurious attention heads via mechanistic insights
• Mitigates bias through targeted ablation technique
• Improves classification via orthogonal projection integration