Towards Continuous Mathematical Models for the Analysis of Classes of Deep Neural Networks
摘要
Our goal is to develop and formulate continuous mathematical models for the analysis of deep neural networks, in order to a) provide a convergence analysis, b) conduct an optimization analysis with respect to the optimal choice and computation of parameters, c) study the role of overparametrization. The tools used here are control theory, Hamilton-Jacobi-Bellman equations, Barron functions as well as different optimization algorithms.