Bath Engineering & Design Studentships: Knowledge Representation in Bio-Inspired Neural Networks

This project is one of a number in competition for a studentship from the Faculty of Engineering and Design at the University of Bath. If successful, this studentship is expected to commence 04 October 2027.

An alternative start date may be possible if agreed with your intended supervisors and the Doctoral College.

Artificial Intelligence (AI) currently dominates most modern computing tasks, predominantly through their scaled-up versions deep neural networks (DNNs). Despite the constant improvement in efficiency, the data and power demands of AI are ever growing due to their very fast-paced development.
This project focuses on bio-inspired knowledge representation in AI models. For example, we may want to replace an existing face profile in a facial recognition system with a different face profile without retraining the entire model. To do this efficiently, one needs to identify the specific parts of that model representing the information of interest. This challenge is closely related to abstraction and concept learning, which are central to human generalization but remain less explicit and interpretable in most machine learning models [1]. 

The project will look at models such as spiking neural networks (SNNs) and consider the problems of knowledge representation and localised model editing. SNNs have been considered the third generation of ANNs which, besides being more biologically realistic, are characterised by data sparsity and power efficiency [2]. The ability of today’s hardware to emulate biological neurons is constantly improving [3]. Being able to characterise the representation of knowledge in such networks has a strong impact on efficient AI as well as enabling better hypotheses for knowledge representation in the brain. The project is expected to lead to publications in top AI and Computational Neuroscience conferences and journals.