IFML Seminar: 04/18/25 - Learning to Solve Imaging Inverse Problems without Ground Truth

Abstract: Ill-posed inverse problems appear in many critical scientific imaging scenarios where the goal is to reconstruct cleaner images faster from fewer measurements. While deep learning methods reconstruct high quality images quickly, they require ground truth to train, which is impractical or impossible to obtain in real problems. In this talk I will introduce various recent self-supervised paradigms that learn to reconstruct without ground truth, and show some recent results on real imaging ground-truth-free problems including accelerated dynamic MRI reconstruction and superresolution for remote sensing. Finally, I will demonstrate how you can use our Python library, DeepInverse, to train your own image reconstruction models using state-of-the-art deep learning methods.

Bio: Andrew Wang is a PhD student at the Institute for Imaging, Data and Communications in the School of Engineering at the University of Edinburgh, supervised by Prof. Mike Davies. His research focuses on self-supervised learning for imaging inverse problems, and is a lead developer of the DeepInverse Python library. Andrew welcomes connections with researchers working in inverse problems, computer vision, self-supervised learning, geometric deep learning, medical imaging, and remote sensing. Reach out via LinkedIn or email.