Provable Fairness Repair for Deep Neural Networks
Researchers have developed a new "fairness layer" that can be integrated into deep learning models to ensure specific fairness criteria are met. This layer works by appending to the model's output and uses a differentiable optimization approach. An accompanying online primal-dual inference algorithm provides aggregate fairness guarantees even for streaming predictions with very small batch sizes. AI
IMPACT Introduces a novel method for embedding fairness constraints directly into deep learning models, potentially improving ethical AI development.