Selected publications
For a full list, see my CV (
PDF) or the
list on
Google Scholar.
-
Alternating dual updates algorithm for X-ray CT reconstruction on the GPU
Madison G. McGaffin, Jeffrey A. Fessler
IEEE Transactions on Computational Imaging, September 2015
(
PDF)
Slides from Fully3D 2015 conference (
PDF)
This paper demonstrates a way to use the proximal point
algorithm with stochastic group coordinate ascent to "split"
different parts of a cost function from one another (e.g.,
a data fidelity term and a noise-reducing regularizer). This
trick also allows one to design algorithms that consider only a
portion of the data at a time (as in stochastic gradient methods),
but provides some convergence theory without relying on relaxation.
-
Edge-preserving image denoising via group coordinate descent on the GPU
Madison G. McGaffin, Jeffrey A. Fessler
IEEE Transactions on Image Processing, April 2015 (
PDF)
Slides from SPIE Computational Imaging 2014 (
PDF)
This paper presents an intuitive algorithm for doing
edge-preserving image denoising using group coordinate descent.
The trick is to select the groups of pixels that are uncoupled
by the cost function; these pixels can be simultaneously updated
on a GPU. Because group coordinate descent requires storing
no additional variables besides the problem parameters and the
variable being optimized, this algorithm requires very little
memory and so is well-suited to the GPU.
-
Algorithmic design of majorizers for large-scale inverse problems
Madison G. McGaffin, Jeffrey A. Fessler
Preprint on arXiv (
PDF)
This preprint considers finding a matrix \(D\) such that \(K'DK
\succeq H\), where \(H\) is a given large positive semidefinite
matrix. This problem could easily be solved with semidefinite
programming techniques, if not for needing to manipulate and
factor enormous matrices. We present an approach to find
majorizers solutions that are approximately minimum Frobenius
norm. The method only requires storing vectors and computing
matrix-vector products. Our hope is that this opens the door
to more exotic and power matrix majorizers that would be
difficult to design by hand.
Dissertation
X-ray CT Image Reconstruction on Highly-Parallel Architectures