Representation Power of Neural Networks

Date

October 29, 2015

Speaker

Matus Telgarsky

Affiliation

University of Michigan

Overview

This talk will survey a variety of classical results on the representation power of neural networks, and then close with a new result separating shallow and deep networks: namely, there exist classification problems where any shallow network needs exponentially as many nodes to match the accuracy of certain deep or recurrent networks

Speakers

Matus Telgarsky

Matus Telgarsky is a postdoc in EECS at University of Michigan.