Deep Learning Theory

This project develops mathematical theory for deep learning, critical in making these enormously successful machine learning methods more broadly applicable, efficient, interpretable, safe, and reliable. Concretely, we seek to clarify the interplay between the representational power of artificial neural networks as parametric sets of hypotheses, the properties and consequences of the parameter optimization procedures that are employed in order to select a hypothesis based on data, and the performance of trained neural networks at test time on new data.