Ozan Öktem, Professor, KTH Royal Institute of Technology in Stockholm

Structured Learning for Inverse Problems (I)

10月13日 (周二) 10:00-11:00，蓝白楼311

Structured learning is an emerging research area within statistical learning that focuses on injecting inductive bias into the learning in order to improve upon robustness, interpretability and generalisation. This has been important when adapting machine learning to applications in science, engineering, and medicine. Challenges in these fields have attributes that make them very different in nature to computer science applications where machine learning has traditionally found success. The talk will provide an overview of structured learning from the viewpoint of solving large-scale ill-posed inverse problems. This includes usage of domain adapted deep neural network architectures obtained by unrolling a suitable iterative scheme (so they can account for the forward operator and its adjoint) and using a domain adapted loss (task adapted reconstruction). We will next showcase this framework of structured deep learning for limited angle tomographic image reconstruction where the compute pipeline also encodes the microlocal canonical relation. The latter is a handcrafted description from mathematical analysis of how singularities in tomographic data are related singularities in the image.

Bio: Prof. Öktem is a chaired professor of mathematics at KTH Royal Institute of Technology in Stockholm, with over 13 years of experience working as an applied mathematician in industry before returning to academia in 2009. His research focuses on theory and algorithms for severely ill-posed inverse problems in tomographic imaging. The work lies at the intersection of mathematical analysis, differential geometry, statistics, and machine learning. Since 2016, he has led work in combining model-based approaches with deep neural networks for uncertainty quantification and task-adapted reconstruction in large-scale inverse problems, driven by applications such as cryo-EM, fluorescence microscopy, low-dose CT and PET/CT, X-ray phase-contrast tomography, and seismic tomography.