Deep Learning Theory credit:

4 Hours

A rigorous mathematical course covering foundational analyses of the approximation, optimization, and generalization properties of Deep Neural Networks. Topics include: constructive and non-constructive approximations with one hidden layer; benefits of depth; optimization in the NTK regime; maximum margin optimization outside the NTK regime; Rademacher complexity, VC dimensino, and covering number bounds for ReLU networks. Evaluation is primarily based on homeworks, with a smaller project component. The course goal is to prepare students perform their own research in the field. Prerequisite: Basic linear algebra, probability, proof-writing, and statistics required. Real analysis recommended.