Dynamic Deep Learning Rollout Sensing and Processing for Interpretability
Abstract
Deep neural networks provide unprecedented performance gains in many real world problems. Despite these gains, future development and practical deployment of deep networks is hindered by their black-box nature, i.e., lack of interpretability, and by the need for very large training sets. An emerging technique called algorithm unrolling o ers promise in eliminating these issues by providing a concrete and systematic connection between iterative algorithms and deep neural networks. We aim to develop novel deep network architectures that are inspired by iterative algorithms for two signifcant problem domains- 1.) inverse problems in signal and image processing { notably blind signal-image deconvolution, and 2.) radio frequency (RF) sensing and processing namely radar signal processing problems such as waveform design and optimization. For image deconvolution, we first present an iterative algorithm based on half-quadratic splitting (HQS) that may be considered as a generalization of the traditional total-variation regularization method in the gradient domain. We then unroll the algorithm to construct a neural network that once learned, processes `blurred signals-images to output both the de-convolved image as well as an estimate of the underlying kernel. Generalized transform domains can be learned by optimizing HQS algorithm parameters using training data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 22, 2024
- Source ID
- FA95502310009
Entities
People
- Vishal Monga
Organizations
- Air Force Office of Scientific Research
- Pennsylvania State University
- United States Air Force