Description
Deep neural networks are complex functions composed of a large number of simple units called neurons. Their remarkable success in machine learning, where they excel in high-dimensional problems when they have a large number of parameters, has challenged classical theories of learning. Understanding how learning emerges from the interaction of millions of neurons presents a deep theoretical challenge for mathematicians, physicists and statisticians. In this talk, I will give a brief introduction into what neural networks are and how they learn. Then I will argue that a modern theory of deep learning will have to explain how learning emerges from the interplay of network architecture, the learning algorithm, and the structure of the training data.
