Numerical Analysis for Stable AI

About the Project

Recent empirical evidence shows that many large-scale AI systems suffer from two clear drawbacks. First, they are unstable, in the sense of being vulnerable to small changes in the input data or the system parameters. In other words, they can be fooled by adversarial attackers. Second, they reflect, and sometimes amplify, any bias that is present in the training data. This project will study these two issues from a mathematical perspective, with the aim of (a) providing insights into why these effects arise and (b) where possible, suggesting mitigation strategies. Techniques involved will include numerical analysis, applied linear algebra, applied statistics, optimization and high dimensional analysis. Supporting scientific computing experiments in Pytorch, or a related system, will also form a key part of the project.

This studentship will form a key part of the wider endeavour “Numerical Analysis for Stable AI” (NumAStAI) supported by an Advanced Grant from the European Research Council. The overall team will include two post-doctoral research assistants and one other PhD student. The team will benefit from interactions with many visiting experts, and there is scope for conference travel and external visits. The topic of this studentship overlaps with interdisciplinary concerns around ethics, regulation and privacy, and there will be opportunities to interact with University of Edinburgh colleagues in the Centre for Technomoral Futures and the Generative AI Lab.

Requirements: A good honours-level degree in mathematics, computer science, or a related area. Skills in scientific programming.

Application enquiries: Please contact Prof Des Higham d.j.higham@ed.ac.uk for more information.