Core Concepts
Quality-Diversity Generative Sampling (QDGS) improves fairness and accuracy in training classifiers with balanced synthetic datasets.
Abstract
Quality-Diversity Generative Sampling (QDGS) is a model-agnostic framework that focuses on protecting quality and diversity when generating synthetic training datasets. By using prompt guidance, QDGS optimizes a quality objective across measures of diversity for synthetically generated data without fine-tuning the generative model. The framework aims to create intersectional datasets with a combined blend of visual features, such as skin tone and age, to improve fairness while maintaining accuracy on facial recognition benchmarks. QDGS has shown promising results in debiasing color-biased shape classifiers and improving accuracy on dark-skinned faces in facial recognition tasks.
Stats
QDGS increases the proportion of images recognized with dark skin tones from 9.4% to 25.2%.
QDGS achieves the highest average accuracy across facial recognition benchmarks.
QDGS can repair biases in shape classifiers up to ≈ 27% improvement.
Models pretrained with QD15/50 achieve higher accuracies for dark-skinned faces compared to those pretrained with Rand15/50.
Pretraining with QDGS improves performance on dark-skinned faces.
Quotes
"QDGS is a model-agnostic framework that uses prompt guidance to optimize a quality objective across measures of diversity for synthetically generated data."
"QDGS has the potential to improve trained classifiers by creating balanced synthetic datasets."
"We propose exploring the latent space to identify and generate underrepresented attribute combinations."