ViG-Bias: Visually Grounded Bias Discovery and Mitigation

TL;DR

ViG-Bias tackles hidden biases in visual recognition by leveraging visual explanations to guide cross-modal bias discovery and mitigation. It introduces a modest yet effective mechanism, using a mapping $h$ to highlight regions via GradCAM heatmaps and to pre-process inputs for Visually Grounded variants of DOMINO, FACTS, and Bias-to-Text. Across CelebA, Waterbirds, and NICO++, ViG-Bias consistently improves bias discovery precision ($\text{Precision@k}$) and mitigation metrics (average and worst-group accuracy), and shows robustness to different CAM methods and threshold choices. By aligning explanations with spurious features, ViG-Bias enhances both detection and correction of biases, offering a practical, interpretable boost to existing bias-aware frameworks. Limitations include dependence on localized cues and the challenge of non-local biases, with future work aimed at broader integration with diverse bias detectors and mitigation strategies.

Abstract

The proliferation of machine learning models in critical decision making processes has underscored the need for bias discovery and mitigation strategies. Identifying the reasons behind a biased system is not straightforward, since in many occasions they are associated with hidden spurious correlations which are not easy to spot. Standard approaches rely on bias audits performed by analyzing model performance in pre-defined subgroups of data samples, usually characterized by common attributes like gender or ethnicity when it comes to people, or other specific attributes defining semantically coherent groups of images. However, it is not always possible to know a-priori the specific attributes defining the failure modes of visual recognition systems. Recent approaches propose to discover these groups by leveraging large vision language models, which enable the extraction of cross-modal embeddings and the generation of textual descriptions to characterize the subgroups where a certain model is underperforming. In this work, we argue that incorporating visual explanations (e.g. heatmaps generated via GradCAM or other approaches) can boost the performance of such bias discovery and mitigation frameworks. To this end, we introduce Visually Grounded Bias Discovery and Mitigation (ViG-Bias), a simple yet effective technique which can be integrated to a variety of existing frameworks to improve both, discovery and mitigation performance. Our comprehensive evaluation shows that incorporating visual explanations enhances existing techniques like DOMINO, FACTS and Bias-to-Text, across several challenging datasets, including CelebA, Waterbirds, and NICO++.

Figures (6)

• Figure 1: The four groups in the CelebA (left) and Waterbirds (right) datasets are based on combinations of the spurious attribute and the label. The groups highlighted in green have the most training samples, whereas the groups highlighted in red have the least training samples.
• Figure 2: Visual explanation methods (e.g. GradCAMSelvaraju_2019 for the spurious attribute blond) help with identifying spurious correlations. We use the Intersection-over-Union (IoU) metric to measure the percentage of overlap between a binary image representing the spurious feature and the image we get after applying a visual explanation method.
• Figure 3: The B2T framework treats visual biases as language, allowing to (1) uncover new biases by interpreting keywords and (2) mitigate biases in models by leveraging these identified keywords. Through B2T, spurious correlations between attributes like female and blond are uncovered. To enhance this process, we suggest integrating a visual explanation component before keyword generation. This addition aims to improve the relevance of identified keywords and assess the effectiveness of using these keywords in debiasing.
• Figure 4: Our objective is to pinpoint slices of data where a spurious correlation exists between a task-irrelevant attribute (such as gender) and the label of interest. For instance, in the given example, women who are not blond correspond to a bias-conflicting slice, while not blond men form a bias-aligned slice. FACTS amplifies correlations, aiming to establish a straightforward bias-aligned hypothesis. Then, it applies a visual explanation to the dataset using the amplified model. Finally, correlation-aware slicing is executed, a process in which clustering is conducted within the bias-amplified feature space. To enhance this process, we suggest integrating a visual explanation component $h$ before keyword generation.
• Figure 5: Ablating the GradCAM Selvaraju_2019 mask threshold $\tau$ on the CelebA dataset for bias discovery (a) and mitigation (b). The main hyperparameter that is important for our model is the parameter $\tau$ that integrates the optimal threshold for the mask. We optimize this parameter depending on the dataset, however, our method performs better than the baseline in almost all the different choices of $\tau$. Empirically, as a rule of thumb, setting $\tau=0.5$ is a good starting point.