Abstract
Machine learning (ML) technologies are increasingly deployed in sensitive applications, yet they often exhibit significant performance disparities across different demographic groups. This paper addresses the issue of "population bias," where discriminative models perform poorly on underrepresented or vulnerable populations. We propose a novel framework that identifies weak decision boundaries and repairs them through targeted data augmentation. By leveraging explainability tools, web search, and generative models (DALL-E 2 and Stable Diffusion), we surgically enhance training datasets. Our approach, demonstrated on the ImageNet People Subtree, achieves a 77.30% reduction in gender accuracy disparity and improves overall model accuracy by approximately 4%, demonstrating that fairness and performance can be improved simultaneously.
Introduction
As businesses increasingly integrate image classification into critical decision-making processes, the ethical implications of model bias have become a primary concern. A recurring problem in computer vision is the failure of models to generalize across diverse populations, often resulting in lower accuracy for groups such as female doctors of color. These failures are frequently rooted in "weak decision boundaries"—regions in the model's latent space where the classifier lacks sufficient representative data to distinguish between classes accurately.
Traditional mitigation strategies often rely on "blanket" data balancing, which involves adding more images to minority classes. However, this approach often fails to address the underlying cause of bias: spurious correlations. For example, a model might associate a "potted plant" with a "gardener" rather than a "nurse," leading to misclassification regardless of the subject's identity. This research proposes a more surgical intervention to identify and "heal" these specific weakspots.
Methodology
1. Weakspot Detection and the "Pivotal Image"
To identify where the model struggles, we utilize FAISS (Facebook AI Similarity Search) to analyze the model’s latent space. By clustering misclassified images, we can identify a "pivotal image"—a sample at the center of a failure cluster that represents a specific type of model confusion. This allows us to target the exact scenarios where the decision boundary is most fragile.
2. Explainability and Metadata Extraction
Once a weakspot is identified, we must understand why the model failed. We employ a suite of explainability tools:
- GradCAM: Generates heatmaps to visualize which parts of the image the model prioritized.
- Places365: Provides scene recognition to determine if the background context is influencing the label.
- Mask R-CNN: Performs object segmentation to identify specific items in the frame that might be causing spurious correlations (e.g., a wooden structure leading to a "carpenter" label instead of "lifeguard").
3. Targeted Data Augmentation
Using the insights from the explainability phase, we generate captions for the pivotal images using ViT-GPT2. These captions are then enriched with metadata (e.g., gender, race, setting) to create high-precision prompts for data procurement:
- Web Search: We query search engines for real-world examples that match the identified gaps.
- Generative AI: For rare or counter-stereotypical intersections where web search may fail (e.g., a male nurse in a specific domestic setting), we use DALL-E 2 and Stable Diffusion to synthesize the exact training samples needed to break spurious correlations.
4. Fine-Tuning
The original model is then retrained on the augmented dataset, which includes these strategically crafted samples alongside the original training data.
Experimental Setup
The framework was evaluated using the ImageNet People SubTree. We refined the initial 2,832 categories down to 40 occupation-based classes (e.g., doctor, military officer, carpenter) that were deemed "safe and imageable." The final dataset consisted of approximately 64,516 training images and 6,278 annotated test images, specifically curated to evaluate performance across diverse demographics.
Results and Discussion
Performance and Fairness Gains
The proposed method yielded significant improvements in both equity and raw performance:
- Gender Disparity: We achieved a 77.30% reduction in gender accuracy disparity.
- Overall Accuracy: Model accuracy rose from 80.12% to 84.09%, an increase of roughly 4%.
- Vulnerable Populations: In a specific case study of female doctors of color, accuracy improved from a baseline of 27.78% to significantly higher levels following the targeted augmentation.
Structural Improvements
Beyond metrics, post-training analysis revealed a notable enhancement in the classifier's decision boundaries. The latent space exhibited fewer weakspots and increased separation between classes, suggesting that the model had learned more robust features rather than relying on environmental cues or stereotypes.
The Power of Surgical Intervention
A key insight of this work is that massive datasets are not always the answer. By adding a relatively small number of strategically chosen images (representing only a 3.32% increase in total dataset size), we achieved results that blanket data increases often fail to provide. Generative AI proved particularly useful as a "gap-filler" for scenarios where search engines are limited by their own inherent biases.
Conclusion and Future Work
This research demonstrates that leveraging generative models and web search can effectively mitigate population bias in image classifiers. While the current framework requires a human-in-the-loop to identify which correlations are "spurious," future work will focus on automating the detection of socially harmful biases. Additionally, we aim to transition from post-hoc "fixes" to proactive training methodologies that prevent weak decision boundaries from forming during the initial learning phase. Although showcased on occupational classes, this technique is extendable to a wide range of domains where fairness and robustness are paramount.
Summary updated