On Successes and Failures of Deep Learning

In the last 5 years, Deep Learning has become the go-to solution for a broad range of applications, often outperforming state-of-the-art. In the first part of the talk I will describe the success of deep learning for autonomous driving. The second part of the talk deals with our current theoretical understanding of deep learning. I will argue that in order to make progress, it is important, for both theoreticians and practitioners, to gain a deeper understanding of the difficulties and limitations associated with common approaches and algorithms. Finally, I will leverage the insight from failures of deep learning to present new results on the cruciality of convolutions for the success of gradient based deep learning.