A mathematical theory of deep neural networks – Prof. Helmut Bolcskei, ETH Zurich

When:
May 22, 2020 @ 12:00 pm – 1:00 pm
2020-05-22T12:00:00+01:00
2020-05-22T13:00:00+01:00
Where:
Department of Engineering
LT1
Contact:
Dr Ramji Venkataramanan

During the past decade deep neural networks have led to spectacular successes in a wide range of applications such as image classification and annotation, handwritten digit recognition, speech recognition, and game intelligence. In this talk, we describe efforts to
develop a mathematical theory that can explain these impressive practical achievements and possibly guide future deep learning architectures and algorithms. Specifically, we develop the fundamental limits of learning in deep neural networks by characterizing what is possible in principle. We then attempt to explain the inner workings of deep generative networks and of scattering networks. A brief survey of recent results on deep networks as solution engines for PDEs is followed by considerations of interesting open problems and philosophical remarks on the role of mathematics in AI research.

Leave a Reply

Your email address will not be published. Required fields are marked *