Research
At the Centre for AI Fundamentals, we specialise in machine learning and collaborate with experts across various fields to address real-world challenges. We are actively engaged with an expanding network of leading researchers in several key disciplines, working together on a range of cross-cutting themes to drive innovation and discovery.
Our research themes
AI for science
Through research into new methodologies for AI and machine learning, we're seeking to help translate scientific discovery into real-world applications. The Centre will be at the interface between AI methods and science and engineering, via virtual (simulation based) laboratories that will apply ML-based probabilistic modelling, simulator-based inference, digital twins and collaborative AI for human-AI teamwork. A wide range of disciplines and sectors are anticipated to collaborate with this work.
Current highlights:
- Virtual Laboratories: Transforming research with AI
- Investigating the ability of PINNs to solve Burgers’ PDE near finite-time blowup
- Towards Size-Independent Generalization Bounds for Deep Operator Nets
Decision making in machine learning systems
For scientific systems that produce huge data volumes, so-called “big science” (e.g., SKA, CERN), AI-driven decisions are increasingly necessary to replace human decisions at multiple points within large scientific analyses and other areas such as facility operations. Our team will research automated AI approaches that can ensure such systems combination is robust, safe and accurate. Decision making with AI needs to be interpretable and explainable to facilitate interrogation of decision processes such that trust can be built by the human, and it is essential for understanding and meeting ethical and legal implications.
Current highlights:
- SMACv2: An Improved Benchmark for Cooperative Multi-Agent Reinforcement Learning
- Comparing the Efficacy of Fine-Tuning and Meta-Learning for Few-Shot Policy Imitation
- Imitating Human Behaviour with Diffusion Models.
Decision making with humans in the loop
A human user, in many cases, is unable to fully specify all details a computer systems would require. By jointly modelling the machine learning task with humans in the loop, a system’s decision making can be improved over time. Through this AI technologies will be more efficient in addressing key challenges such as experimental design from limited data as well as promoting trust in AI-enabled systems. Decision making with Humans in the Loop (DMHL) is a key theme in researching the fundamentals of AI.
Current highlights:
Theory of machine learning
We are uncovering the core mathematical principles that underpin machine learning. Theory research can yield overarching ideas that one can rely on for building deployable intelligent machines. Focus areas include: developing generalisation bounds for conventional deep-neural systems and operator learning setups; establishing novel principles to provably design algorithms for training neural nets and do reinforcement learning; and addressing the mystery of why the size of neural architectures is such a critical factor behind performance - particularly for (operator) neural nets that solve (systems of) partial differential equations.
Current highlights:
- Size Lowerbounds for Deep Operator Networks
- Global Convergence of SGD For Logistic Loss on Two Layer Neural Nets
- Trust Region Bounds for Decentralized PPO Under Non-stationarity.
Uncertainty in complex systems
Human and AI collaborative decision making requires principled uncertainty quantification. Our researchers seek to develop leading-edge methods in probabilistic machine learning, leveraging uncertainty in a statistical manner to drive the exploration of new parameter spaces and promote scientific discovery. Focusing both on the methodological and theoretical aspects, our research aims to help any field where decision-making is critical. Uncertainty quantification and modelling underpins two further themes on decision making.
Current highlights:
- Multi-output prediction of dose-response curves enable drug repositioning and biomarker discovery
- Spatio-Angular Convolutions for Super-resolution in Diffusion MRI
- Adjoint-aided inference of Gaussian process driven differential equations.
How we apply our research
Outcomes of our leading-edge work on the fundamentals of AI will be applied in varied domains. Examples include:
- Engineering Biology > Manchester Institute for Biotechnology
- Health > Christabel Pankhurst Institute and Cancer Beacon
- Materials Science > Henry Royce Institute
- Physics and Astronomy > Department of Physics and Astronomy
- Robotics > Centre for Robotics and AI