Neural network learning: A shallow overview with deep dives

Stephan Wojtowytsch

Assistant Professor of Mathematics

Texas A&M University

 In supervised machine learning in general and deep learning in particular, we use functions in parametrized classes to interpolate given labels on finite data sets with the goal of extracting an underlying pattern for future use on previously unseen data. In this talk, we will consider general ideas before considering two specific problems in greater detail: (1) The trade-off between the approximation power of a function class and its ability to learn from finite amounts of data and (2) aspects of 'training' a neural network through the lens of stochastic optimization.