Guest
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.
