PhD in Improving Generalization Through Advanced Learning Strategies
Achieving robust generalisation remains a central challenge in artificial intelligence, significantly impacting the practical deployment of machine learning models. Current approaches predominantly focus on finding flat minima in the loss landscape, yet the precise mechanisms enabling robust generalisation remain unclear. This project proposes to investigate an alternative formulation to this challenge: rather than merely pursuing flat minima, we investigate the mechanisms that differentiate generalisable features from those that lead to overfitting.
The project goal is to explore generalisation formulations via structured perturbation methods. Through deliberate perturbations, our methodology aims to distinguish between generalisable and non-generalisable features, effectively promoting robust performance by enabling models to escape the pitfalls of spurious correlations and noisy data.
This project will establish a new generation of optimisation algorithms based on structured forgetting applicable not only to neural networks but to all gradient-descent based approaches. It will study how selective forgetting enhances generalisation, robustness, and efficiency across multiple applications. This research will address fundamental questions about the nature of intelligence and learning, pushing the boundaries of what is possible in AI.
This project bridges foundational theoretical insights with practical algorithmic innovations to develop more robust, efficient, and trustworthy AI systems, marking a pivotal step toward general and robust intelligence characterised by human-like adaptability. The ideal candidate will have a strong background in mathematics, computer science, and machine learning, with a keen interest in theoretical research and its practical applications. This is an opportunity to contribute to a groundbreaking research agenda with the potential for high-impact publications and to shape the future of artificial intelligence.
Supervision
You will be supervised by Dr. Gabriel Oliveira (Personal Website). His research focuses on Robust Machine Learning for real-world applications, with expertise spanning computer vision, time series prediction, robotics and federated learning. Dr. Gabriel Oliveira has published in top-tier ML, Computer Vision and Robotics venues, including ICLR, ICML, ICCV, TMLR and IJRR. His work has been widely cited, demonstrating a strong impact on Robust ML.
During the PhD study, you will receive comprehensive research training, including:
Regular one-to-one meetings to guide your research direction, ensure steady progress, and refine your problem-solving skills.
Access to cutting-edge computational resources, including Bristol University’s Isambard AI GPU cluster and state-of-the-art facilities.
The ideal candidate should have knowledge of ML and Math. Interested applicants are strongly advised to contact Dr. Oliveira (gabriel.leivasoliveira@bristol.ac.uk) to discuss more details.
Apply by November 17th.
