Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jul 2·5 min read
ForAug: Mitigating Biases in Image Classification via Controlled Image Compositions
TL;DR
Image classification systems often exhibit biases, leading to inaccurate results across different demographics. ForAug is a new technique that creates controlled image compositions to reduce these biases.
✦ Why It Matters
Engineers can implement ForAug to enhance fairness in their image classification models, improving overall system reliability.
Key Takeaways
- Utilize ForAug to control object position and scale in training images.
- Apply segmentation and inpainting models for effective image factorization.
- Conduct targeted diagnostic tests to identify and reduce biases.
- Leverage public code and datasets to replicate and build upon findings.
How It Works
ForAug separates images into foreground objects and backgrounds, allowing for controlled recombination. By manipulating object position, scale, and background identity, it breaks the typical correlations that lead to biases in model training.
This is achieved using off-the-shelf segmentation and inpainting models, which extract the foreground and create neutral backgrounds for diverse combinations.