Formal Methods for Safe Artificial Intelligence

The School of Computer Science at the University of Birmingham offers one PhD position in formal methods for safe artificial intelligence.

Safe artificial intelligence addresses the problem of automatically ensuring that learning systems satisfy correctness specifications. For deep learning systems which are based on neural networks, computer-aided formal verification is crucial; this is because (1) they often are not human interpretable, (2) they change behaviour continuously and automatically, and (3) they are being increasingly applied to safety critical domains. Exemplars are robotics and autonomous driving contexts, where deep learning systems are being embedded in decision-making components. Verifying the correctness of these systems is an open and important problem in artificial intelligence, highly relevant to industry and academia.

Computer-aided verification of deep learning systems connects formal methods, machine learning and software engineering, and builds upon symbolic verification methods such as model checking, abstract interpretation, and satisfiability modulo theories as well as novel machine learning methods for safe AI. The candidate will contribute to open problems in the area. Specific research questions include but are not limited to, developing methods to ensure that cyber-physical systems with deep learning components are safe [1] and ensuring that software and hardware algorithmic components are correct [2]. This research is at the intersection of verification and machine learning and spans the analysis of software and cyber-physical systems and the analysis of learning systems.

Overall, the project aims at developing verification, synthesis, and machine learning methods that ensure that digital, physical, probabilistic systems interacting with deep learning and neural network are formally guaranteed to be safe.

The candidate will study and work in a stimulating environment within the school and the university and will have the opportunity to research theoretical and/or experimental aspects of formal verification of learning systems, investigate connections between theoretical computer science (e.g., logic, automata) and machine learning, and explore applications to science and engineering (e.g., control, robotics). The project spans the theory and practice of computer science and requires

– analytical skills (strong background in mathematical modelling, logic, and algorithms design) and

– programming skills (willingness to learn verification and machine learning toolchains, ability to develop software prototypes).

In short, the candidate should have a strong interest for both maths and code.

We invite home and international candidates with First or Upper Second Class Honours undergraduate degree (or an international equivalent) in computer science, electrical engineering, mathematics, physics, or any other academic background that is relevant to our research portfolio and the intended research of the candidate.

We aim for our PhD student cohorts to reflect the diversity of our society and are dedicated to ensuring equal opportunities for all applicants. We encourage applicants from ethnic minorities, under-represented groups, individuals with disabilities, and neurodiverse candidates. We provide support to students to help them adapt to their unique personal circumstances through options like part-time and split-site study opportunities.