Two PhD Studentships – Neural network analysis of quantum processes
Supervisor: Prof Ilya Kuprov
This project is about training and reverse-engineering deep neural networks that process and interpret magnetic resonance data.
Earlier this year, we created a family of neural networks that recovered molecular distance distributions from electron spin resonance data. The networks reliably returned distance data for proteins, nucleic acids, photosynthetic reaction centres, and other systems of current chemical and biological interest. However… this was not supposed to be possible: the mathematical transformations in question are ill-posed and numerically unstable. It is utterly unclear how neural networks accomplish what is technically, without regularisation, an impossible mathematical operation.
The objective of this project is to find out. You will make use of one of the biggest supercomputers in the UK to run neural network training against quantum mechanical simulations of spin dynamics in biological systems. The resulting networks will be taken apart with the purpose of finding out how exactly they work. The conclusions will have significance across physical sciences – at the moment, the internal functioning of artificial neural networks is largely a mystery.
These are 4-year studentships; they are open to UK and EU nationals, and include all applicable university fees, as well as a tax-free stipend of £15,009 per year. It is likely that the students will make a major contribution to both quantum theory and artificial intelligence by exploring, developing, and analysing neural networks that process spectroscopic data. This project is a collaboration with ETH Zurich (Prof Gunnar Jeschke) and Weizmann Institute (Prof Daniella Goldfarb), and will involve visits to both of these institutions.
University of Southampton supercomputing facilities are among the best in the country – Iridis5 cluster is the most powerful academic supercomputer in the UK. It has 20,000 CPU cores as well as dedicated NVidia Tesla V100 compute nodes that this project will use.
To start the application process, please send a CV to Prof Ilya Kuprov (firstname.lastname@example.org). Deadline: 31 May 2019; earlier submissions would be much appreciated for logistical reasons.