Dimension reduction in physical and data sciences

Global convergence in neural network optimization

Eric Vanden-Eijnden

New York University

I will discuss two questions related to neural networks used in machine learning: (i) how accurate is the function approximation they provide and (ii) how trainable are they? These questions will be investigated by mapping the parameters of the neural network to a system of interacting particles. I will also show how these findings can be used to accelerate the training of networks and optimize their architecture, and discuss what these results imply for applications in scientific computing.