Abstract
Deep learning models often inherit and amplify biases present in training data, leading to significant performance disparities across different demographic groups. This paper introduces a novel framework designed to identify and mitigate these biases by locating "weakspots" in a model's decision boundaries. By integrating latent space analysis, metadata extraction, and targeted data augmentation using web search and generative models (e.g., DALL-E 2, Stable Diffusion), the proposed method strategically patches these weak regions. Experimental results on the ImageNet People SubTree demonstrate a 77.3% reduction in gender accuracy disparity and a 4% improvement in overall classification accuracy.
Introduction
A persistent challenge in computer vision is the tendency of models to rely on spurious correlations—such as associating a "potted plant" with a "gardener" rather than the person's physical features. These failures are often concentrated in specific regions of the model's latent space, termed "weak decision boundaries." Simply balancing datasets with more images is frequently insufficient to resolve these localized failures. This work proposes a systematic pipeline to identify these "pivotal images" where the model is confused and procure targeted data to strengthen the decision boundary.
Methodology
The authors propose a multi-stage pipeline to identify and repair model weakspots, as illustrated in Figure 1.
1. Identifying Weakspots
The system utilizes Facebook AI Similarity Search (FAISS) to explore the latent space for misclassified instances within high-perplexity neighborhoods. These instances, labeled "pivotal images," serve as anchors for the repair process.
2. Metadata Extraction and Explainability
To diagnose the root cause of failure, the framework extracts rich metadata from pivotal images. This involves object segmentation via Mask R-CNN, scene recognition using Places365, and the generation of Grad-CAM heatmaps to visualize model attention. As shown in Figure 2, this process reveals whether the model is focusing on irrelevant background objects or spurious features.
3. Prompt Engineering and Data Augmentation
Captions for pivotal images are generated using ViT-GPT2 and refined by incorporating the extracted scene metadata and specific class labels. These high-fidelity prompts are then used to retrieve or generate "neutralizing" images via Google Image Search, DALL-E 2, and Stable Diffusion. This targeted augmentation aims to break spurious correlations by providing counter-examples.
4. Retraining
The original training set is augmented with these targeted samples, and the classifier (ResNet50) is finetuned to achieve more robust decision boundaries.
Experimental Results
The framework was evaluated on a subset of the ImageNet People SubTree, focusing on 40 occupational classes. Key findings include:
- Fairness Improvements: Gender accuracy disparity was reduced by 77.30%. For the "person of color female doctor" demographic, which initially had a dismal 27.78% accuracy, the disparity was reduced by nearly 50%.
- Accuracy Gains: Overall test accuracy increased from 80.12% to 84.09%, demonstrating that bias mitigation does not necessarily trade off with performance.
- Boundary Robustness: Post-intervention analysis showed a significant reduction in the number of identified weakspots, indicating cleaner separation in the latent space.
Conclusion and Future Work
This research demonstrates that targeted generative "patching" is a highly effective strategy for improving model robustness and fairness. While the current approach requires some human intervention to validate spurious correlations, future work will focus on making the process more autonomous and proactive, preventing bias during the initial training phase rather than relying solely on post-hoc corrections.
Summary updated