Simon Arridge
University College London
"Some Aspects of Modeling Error and Bayesian Methods in Ill-Posed Inverse Problems"
Inverse Problems in imaging are typically ill-posed and in some cases severely so. A typical approach involves a numerical model for the forward problem that is fit to data, plus regularization to control instability. If the numerical model is not very accurate, additional errors are included which may be greater than the measurement noise. One recent approach to solving this problem uses a combined model of measurement and modeling errors. To construct such a model involves sampling on an appropriate prior distribution. In this talk I will show some recent examples applied to diffuse optical tomography, which is non-linear and exponentially ill-posed. By taking account of the modeling error statistics, significant accelerations can be achieved.
Host: Guillaume Bal
Please note: the University is officially closed on Tuesday, Nov. 3, for the election day holiday. To enter the Mudd Building and room 214 Mudd, please contact Prof. Bal ahead of time to make arrangements.
|
