Unrolled deep networks for sparse signal restoration
Abstract
This paper addresses the problem of sparse signal recovery
by deep unrolling approaches. The task of sparse restoration
from linearly transformed and noisy observations occurs
in many applied fields, including geoscience, biochemistry,
remote sensing, and biomedical data processing, and was
thoroughly studied in literature. Recently, novel approaches
based on ‘deep unrolling’ or ‘deep unfolding’, have been
introduced. They consist in creating deep neural networks
inspired from iterative algorithms initially built for penalized
loss minimization. The iterations of the algorithm are recast
as neural network layers. The use of deep learning frameworks
ensures an efficient implementation and the possibility
to learn the algorithm native hyperparameters, through the
minimization of a task-oriented loss. However, for a given
application, choosing an adequate iterative scheme to unroll
and fine-tuning the architecture parameters remains a challenging
task. In this work, our goal is to present a comprehensive
comparative study of deep unrolled approaches deployed
for sparse signal reconstruction. Three architectures are introduced
and compared through a motivating application,
arising in analytical chemistry. A reproducible Github code
is provided.
Domains
Mathematics [math]Origin | Files produced by the author(s) |
---|