AIAI Seminar-Monday 22 April-Talk by Fengxiang He

Title: Building Up Deep Learning Theory: Perspectives of Optimisation and Loss Landscape

Abstract: Deep learning has long been criticised as a “black-box” model. A major problem is why deep neural networks are over-parameterised, but can have good generalisability. In this talk, I will give a brief overview of efforts at building up deep learning theory, and my works from the perspectives of optimisation dynamics and the algebraic/geometric structures of deep neural networks’ loss landscape.