Uniform convergence bounds via PAC-Bayes and Wasserstein distances

In machine learning, practitioners may encounter overfitting when a model performs well on the training set but poorly on the task represented by the test set. One way to assess overfitting is through generalization bounds, which provide upper bounds on a model's performance for unseen tasks. In this talk, I will first review some basics of machine learning and discuss two types of bounds introduced in the literature: PAC-Bayesian bounds and uniform convergence bounds. Although these two types exhibit distinct natures, I will introduce a new approach to obtain generalization bounds that combines their strengths.