The Physics of Machine Learning on Oscillator Ising Machines

A physical approach to machine learning, along with the associated optimisation processes, has often been fruitful both in terms of developing new fundamental knowledge as well as building upon and improving existing methods. This is very much topical to date, with theoretical physics being expected to help develop a statistical theory of learning, and with renewed efforts to tackle optimisation questions with physics-inspired approaches being called for.

Recently we have seen the emergence of specialised analogue hardware that can realise a model widely studied in physics and machine learning: Ising machines. A particular implementation, oscillator Ising machines, uses the synchronisation of self-sustaining oscillators for a fast and energy efficient optimisation of an Ising energy function.

This project will examine the close connection between statistical physics, generative machine learning models and dynamical systems, and it will seek to discover the physical rules that govern efficient learning systems.