In the real world, decisions are typically made under a degree of uncertainty; we rarely have access to perfect information or complete knowledge about a problem. Humans learn to be more cautious under ambiguous information, but in many cases, we struggle to teach machines to behave in the same way. Machine learning (ML) systems are now ubiquitous in society and are becoming ever more responsible for decisions that impact human safety. Examples of ML making life critical decisions include medical imaging diagnosis systems and self-driving cars. To ensure the safety of these decisions, it is essential that these models can be imbued with a meaningful understanding of uncertainty.
This research project will advance our understanding of how to effectively and efficiently produce and make use of uncertain predictions from ML models for computer vision and medical image analysis tasks. We will develop state of the art methods for creating novel probabilistic ML models, which are capable of accurately characterising the inherent uncertainty in the problems they are tasked with. We will also investigate how this information can best be communicated to humans and incorporated into subsequent processes.
The specific application areas for this projectare flexible and will depend on the student’s interest. Some potential topics include:magnetic resonance image analysis, shape analysis, image registration and tracking, segmentation, image editing and depth estimation.
This is a computational project, which would suit a student with good mathematical and programming skills and a keen interest in probabilistic machine learning and computer vision. Students will be expected to present their work at top-tier computer vision, medical image analysis or machine learning venues such as CVPR, ICCV/ECCV, MICCAI, NeurIPS etc.