We innovate in the models and methods that form the foundations of AI systems, making both theoretical and practical contributions across a wide range of areas – machine learning, natural language processing, reasoning, computer vision, robotics, adaptive systems, trustworthy AI, and areas that cross over between them.
Our research spans across multidisciplinary research
Visual understanding is one of the oldest and most central problems in AI. Core tasks like detection, segmentation, and 3D reconstruction remain far from solved, while the rise of large-scale vision and multi-modal models has opened entirely new questions about what machines can and cannot see. Our work spans both layers: advancing foundational visual tasks while probing the capabilities and limits of modern visual models.
Computer Vision is critical for any system that processes images, from personal cameras to medical scans and industrial robots. Our research enables new, more efficient, and more reliable systems, expanding what is possible across a range of industries.
Associate Professor Weidong Cai, Associate Professor Tongliang Liu, Professor Zhiyong Wang, Associate Professor Chang Xu, Dr William Zhi
Large Language Models and other methods from Natural Language Processing are rapidly changing the world. Our work explores how to improve these models, how to understand them, and how to effectively use them as part of larger systems. This involves everything from manipulation of the internal structure of models up to building user interfaces that leverage their unique abilities.
Our research yields advances in LLMs, endowing them with new capabilities, opening new areas of applications, and enabling more efficient training and inference.
Dr Jonathan K. Kummerfeld, Associate Professor Josiah Poon
AI systems and machine learning models are increasingly used in open-ended tasks and deployed in uncertain environments, to handle data and complexity at scales hard to fathom, and to make decisions — from recommendation systems to robots interacting with humans in a manufacturing plant, to self-driving cars in a city. Our research aims to not only advance our fundamental understanding of what can be achieved and what performance to aim for, but, crucially, to provide rigorous, principled approaches to design these systems in ill-understood or highly constrained settings, or with noisy, sensitive or even untrusted data. One aspect of our work involves understanding how to specify decision-making tasks so that humans can understand them and machines can find policies that guarantee them.
Our research informs the design of future AI systems and helps us understand AI today. Understanding is crucial for building systems that people can trust, while exploration of fundamental ideas will help us overcome the limitations of current technology.
Dr Clément Canonne, Professor Irena Koprinska, Associate Professor Tongliang Liu, Professor Mikhail Prokopenko, Dr Sasha Rubin, Dr William Zhi
Distributed Intelligence studies how intelligent behaviour can emerge from many devices, AI models, robots, and agents working together. Rather than concentrating intelligence in a single system, the goal is to enable learning, reasoning, and decision-making to be distributed across the network edge. The research spans distributed AI, Edge AI, federated learning, collaborative inference, and multi-agent systems, with a focus on efficiency, adaptability, and resilience.
Our work improves the scalability, resilience, and capabilities of AI systems that work across a set of devices, such as a set of agents collaborating on a task, or a model learned from data across many phones.
Professor Mikhail Prokopenko, Associate Professor Nguyen Tran
Trustworthy AI asks a deceptively simple question: Can we rely on a model when it matters most? Real-world deployments expose systems to distribution shifts, adversarial inputs, corruption of training data, out-of-distribution data, and novel anomalies (i.e., unknown-unknowns) that current models handle poorly, often failing silently and confidently. Our work spans this landscape, from provably making models robust during training or inference or detecting when a model operates beyond its competence, to building more robust representations and evaluation frameworks that stress-test generalisation rather than reward benchmark performance.
Our research ensures that AI systems can be relied upon, even in unforeseen scenarios or when compromised. This is critical when these systems move from the lab into practical use, when they face new challenges that were not seen in training. Rather than giving a confident incorrect answer, our work leads to systems that can accurately communicate with users, leading to better trust in systems and their capabilities.
Dr Clément Canonne, Dr Jonathan K. Kummerfeld, Associate Professor Tongliang Liu, Dr Sasha Rubin, Associate Professor Suranga Seneviratne
