Divergence Regularization

Divergence regularization is a technique used to improve the robustness and generalization of machine learning models by incorporating measures of dissimilarity between probability distributions into the learning process. Current research focuses on applying various f-divergences (including Kullback-Leibler, Jensen-Shannon, and total variation) and Wasserstein distances within diverse applications such as adversarial training, reinforcement learning, and model alignment (e.g., aligning language models to human preferences). This approach enhances model performance, particularly in handling noisy or uncertain data, improving fairness, and mitigating the impact of distribution shifts, leading to more reliable and effective machine learning systems across various domains.