[50 seconds] Universal Approximation with Deep Narrow Networks

The classical universal approximation theorem (dating back to ~1999) is a foundational result, giving (part of) an answer to the question "why do neural networks work?" This is a brief introduction to our recent paper which proves some natural "dual" theorems on universal approximation, and in particular highlights a difference between deep and shallow neural networks.

This presentation was prepared as part of the acceptance of our paper to the Conference on Learning Theory 2020.