๐๐๐๐ซ๐๐ฌ๐ฌ๐ข๐ง๐ ๐๐๐๐ค ๐๐๐๐ข๐ฌ๐ข๐จ๐ง ๐๐จ๐ฎ๐ง๐๐๐ซ๐ข๐๐ฌ ๐ข๐ง ๐๐ฆ๐๐ ๐ ๐๐ฅ๐๐ฌ๐ฌ๐ข๐๐ข๐๐๐ญ๐ข๐จ๐ง ๐๐ฒ ๐๐๐ฏ๐๐ซ๐๐ ๐ข๐ง๐ ๐๐๐ ๐๐๐๐ซ๐๐ก ๐๐ง๐ ๐๐๐ง๐๐ซ๐๐ญ๐ข๐ฏ๐ ๐๐จ๐๐๐ฅ๐ฌ #IJCAI2023
Preetam Dammu, Yunhe Feng and Chirag Shah
In an era where machine learning (ML) technologies are becoming more prevalent, the ethical and operational issues surrounding them cannot be ignored. Here’s how we tackled this challenge:
๐ก The Problem:
ML models often don’t perform equally well for underrepresented groups, placing vulnerable populations at a disadvantage.
๐ Our Solution:
We leveraged web search and Generative AI to improve the robustness and reduce bias in discriminative ML models.
๐ Methodology:
1. We identified weak decision boundaries for classes representing vulnerable populations (e.g., female doctor of color).
2. We constructed search queries for Google and generated text for creating images with DALL-E 2 and Stable Diffusion.
3. We used these new training samples to reduce population bias.
๐ Results:
1. Achieved a significant reduction (77.30%) in the model’s gender accuracy disparity.
2. Enhanced the classifier’s decision boundary, resulting in fewer weak spots and better class separation.
๐ Applicability:
Although demonstrated on vulnerable populations, this approach is extendable to a wide range of problems and domains.