This project will investigate some foundational questions regarding the quantification of the complexity of neural networks and developing new tools to assist architecture discovery. Recent advances in topological data analysis give new possible directions to investigate at the boundary of AI and Mathematics. The work will concentrate on the systematic development of numerical characteristics that capture the essence of the complexity of both the problem and of the proposed network architecture. These ideas will be tested on various specific data sets. A second part of the project will address the question of a proper topological description of time dependent data. Both these strands will be unified and implemented algorithmically. The project is designed as a pump priming activity that will lead to follow on projects and grant proposals.
Principal Investigator: Professor Jacek Brodzki (Southampton)