Abstract
Vision-Language Pre-training (VLP) models have achieved state-of-the-art performance in numerous cross-modal tasks. Since they are optimized to capture the statistical properties of intra- and inter-modality, there remains risk to learn social biases presented in the data as well. In this work, we (1) introduce a counterfactual-based bias measurement CounterBias to quantify the social bias in VLP models by comparing the [MASK]ed prediction probabilities of factual and counterfactual samples; (2) construct a novel VL-Bias dataset including 24K image-text pairs for measuring gender bias in VLP models, from which we observed that significant gender bias is prevalent in VLP models; and (3) propose a VLP debiasing method FairVLP to minimize the difference in the [MASK]ed prediction probabilities between factual and counterfactual image-text pairs for VLP debiasing. Although CounterBias and FairVLP focus on social bias, they are generalizable to serve as tools and provide new insights to probe and regularize more knowledge in VLP models.
Introduction
Vision-Language Pre-training (VLP) models, such as ALBEF and ViLT, serve as the engines behind modern cross-modal applications like image retrieval and captioning. However, these models are typically trained on massive, uncurated web datasets, leading them to inherit and potentially amplify human social biases. For instance, a model might disproportionately associate "shopping" with women or "doctor" with men.
A significant hurdle in addressing this issue is the difficulty of measurement. While text-based bias is easily probed by swapping pronouns, visual bias is harder to isolate. Finding "natural" image pairs that differ only in the subject's gender while keeping backgrounds, lighting, and clothing identical is nearly impossible. These confounding variables make it difficult to determine if a model's prediction stems from a social stereotype or a background correlation.
Methodology: CounterBias and FairVLP
To overcome the challenge of visual confounders, the authors propose a framework based on counterfactual reasoning.
CounterBias Measurement
The core innovation is the use of model-specific adversarial attacks to generate counterfactual pairs. Starting with an original image (e.g., a woman), a tiny, imperceptible perturbation is added to "trick" the model into perceiving the subject as a different gender (e.g., a man). This creates a pair of images that are pixel-identical in their background but represent different "bias concepts" to the model.
By applying Masked Language Modeling (MLM), researchers can observe how the model’s probability of predicting a target word (e.g., "[MASK] is shopping") shifts between the factual image and its counterfactual counterpart. A large discrepancy indicates a high level of bias.
FairVLP Elimination
Building on this measurement, FairVLP integrates a fairness constraint directly into the pre-training objective. The method minimizes the Kullback-Leibler (KL) divergence between the prediction distributions of factual and counterfactual pairs. This forces the model to ensure that its activity predictions remain consistent, regardless of the gender detected in the visual or textual input.
The VL-Bias Dataset
The study introduces VL-Bias, a benchmark consisting of 24,000 image-text pairs. This dataset covers 52 human activities and 13 occupations, providing a robust foundation for evaluating gender bias. Standard datasets like MS-COCO and Flickr30K were also utilized for initial pre-training and downstream task evaluation.
Analysis of Social Bias in VLP
The research yielded several key insights into how bias manifests in multimodal models:
- Visual Dominance: Gender bias was found to be significantly stronger in the visual modality than in language. This suggests that images contain complex statistical correlations that VLP models are highly sensitive to.
- Cross-Modal Exacerbation: VLP models often exhibit stronger bias than the unimodal BERT models they are built upon, indicating that learning cross-modal representations can amplify stereotypes.
- Architectural Differences: Dual-stream models (like ALBEF) tend to show "cross-modal conformity," where visual and textual biases align. Single-stream models (like ViLT) exhibit more inconsistent bias patterns.
Key Results
The proposed FairVLP method demonstrated significant success in reducing bias without compromising performance:
- Bias Reduction: On the VL-Bias dataset, FairVLP reduced multimodal bias (B@V-L) from 13.02% to 6.97% for activities and from 14.84% to 7.74% for occupations.
- Performance Stability: On the Flickr30K retrieval task, FairVLP maintained a 90.8% Recall@1, matching the performance of the biased vanilla model. This indicates that debiasing does not necessarily lead to a "tax" on accuracy.
Conclusion and Future Work
While this work focuses on gender bias, the counterfactual approach is theoretically generalizable to other social categories like race or age. Furthermore, the authors suggest that this technique could be used to identify and regularize other "spurious correlations," such as a model's over-reliance on background cues (e.g., associating "horse" exclusively with "grass"). Future research will likely explore these broader applications to enhance the general robustness and fairness of AI systems.
Summary updated