Machine Learning Landscapes

This project will involve the characterisation of machine learning landscapes for neural networks. The theory and numerical methods developed for exploring molecular energy landscapes will be applied to the non-convex landscapes defined by the loss function of an artificial neural network. In particular, we will investigate analogues of the density of states to define thermodynamic quantities and rate constants. The resulting insight will be used to define combinations of solutions that provide superior predictive power for applications ranging from molecular geometry optimisation to clinician diagnostic support.