Introduction
I am second year Dphil student under supervision of Prof. Ard Louis.
As with the group, I have interest in how deep learning generalizes in overparameterised regime. We know that most parameters do not contribute to the functions of interest, but we cannot find the "good" generalising function if we train without them. I believe these parameters are required because they help in creating the inductive bias of DNNs. The connection between parameterisation and inductive bias on functions is already well-established by information geometry. Unfortunately, manifold of DNNs are not Riemannian and there are singularities in Fisher Information Matrix. I wish to follow the work of Prof.Sethna on sloppy models and combine them with work of Watanabe on singular learning theory to explain the inductive bias of deep neural networks.
If you are reading this, and you have interest in the field, please email me straight away for collaboration. If you happen to be working on blackhole thermodynamics or Ruppeiner geometry, it is closely related to my work, so please do get in touch.
Short CV:
BA physics in Oxford University
MaST part3 in Cambridge University
RDBMS Software engineer at TmaxData
Machine Learning researcher/engineer in Naver