BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Machine learning tools for large tomographic inverse problems with
  limited training data - Thomas Blumensath (University of Southampton)
DTSTART:20230331T085000Z
DTEND:20230331T094000Z
UID:TALK198283@talks.cam.ac.uk
DESCRIPTION:X-ray tomographic inverse problems with limited measurements r
 equire strong non-linear constraints. Increasingly\, machine learning tool
 s are used to learn these constraints from large collections of representa
 tive training data. When using X-ray tomography to image manufactured comp
 onents\, it is often beneficial to target the training data to the specifi
 c application\, as this can lead to very strong constraints that will allo
 w image reconstruction even if significant amounts of measurements are mis
 sing. There are however two fundamental problems with this approach for re
 al applications. Firstly\, it is often difficult to collect sufficient tra
 ining data to train the most advanced machine learning models. Secondly\, 
 the inverse problem is extremely large\, with billions of measurements use
 d to estimate 3D images with billions of voxels. This further restricts th
 e models that can be trained and used on most computing hardware. We here 
 report on the use of block based 3D image models and show how they can be 
 trained on a single 3D image. This approach can be used for image de-noisi
 ng as well as a building block in an unrolled optimisation algorithm to so
 lve the tomographic inverse problem.&nbsp\;
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR