Talk
Loss surface of deep and wide neural networks
- Quynh Nguyen (University Saarbruecken)
Abstract
What are possible properties of the loss function and how they might change with the size of the network are among the fundamental questions that one could study to improve our understanding of deep learning. In this talk, I will present some results that we have obtained along this direction, including a topological analysis of sub-level sets, critical points with global optimality, and then show how to leverage this insight to get a convergence result for gradient descent.