Friday, 8 September 2017
TITLE: What are deep networks good for?
SPEAKER: Brendan McCane, Department of Computer Science
WHEN: Friday 08 September, 1:00pm
WHERE: Room G34, Owheo Building
Deep networks and deep learning have recently been winning both academia and industry. But the majority of work has been empirical. Since both deep networks and shallow networks are universal function approximators, why should deep networks do better than shallow counterparts? In this talk I will show that a particular deep network architecture is provably more efficient at approximating radially symmetric functions than the best known shallow networks. Then I will show how such an architecture can be used to approximate Gaussian kernel support vector machines. The main result is of theoretical interest only, but don’t worry, proofs will be outlined using diagrams and waving of hands only.