Learning adaptive sensing and active learning

In this two-part thesis, we develop methods for using machine learning to learn adaptive/active sensing and active learning methods. In Part 1, we learn sensing strategies specifically for accelerating Magnetic Resonance Imaging (MRI). In Part 2, we develop methods for learning active learning on a variety of tasks.
Specifically, in Chapter 2, we use reinforcement learning to train one of the first deep active learned sensing strategies for MRI subsampling that can adapt to the anatomy of the patient. MRI is a widely-used non-invasive medical imaging technique whose primary disadvantage is long scanning times that lead to expensive procedures and low patient throughput. Sensing strategies such as ours can drive the acceleration of MRI scans by selecting sets of especially useful measurements during a scan, mitigating this issue. In Chapter 3, we extend our method to the more clinically relevant parallel MRI setting, by incorporating a sampling operation into an iterative reconstruction network. Additionally, our adapted method allows for joint training of the adaptive sensing strategy and the MRI reconstruction model, due to a simplification of the training process.
Next, we explore the topic of learning active learning strategies. Active learning involves a process in which the learner – here, a machine learning model – itself takes part in deciding which examples it should learn from. Our work focuses on scenarios where we train a separate model to help guide this learning process. In Chapter 4, we develop a proof-of-concept system for learning general active learning strategies that can adapt to variations in the specified data setting. In Chapter , we work on a related problem in the context of simulator optimisation, where we train a policy to assist in the optimisation by guiding both the optimisation itself and the training of a surrogate model.