8th International Conference on
Computational Harmonic Analysis

September 12-16, 2022

Ingolstadt, Germany

"Deep unfolding for analysis Compressed Sensing: does redundancy affect the generalization ability?"

Kouni, Vasiliki

In this paper, we examine an ADMM-based deep unfolding network for analysis Compressed Sensing, dubbed U-ADMM-DAD net; the latter jointly learns a decoder for Compressed Sensing and a redundant analysis operator for sparsification. We compare U-ADMM-DAD net to a synthesis-sparsity-based unfolding network -- serving as a baseline -- on real-world speech data. Our experimental results demonstrate that the redundancy of the learnable sparsifier affects the generalization ability of U-ADMM-DAD net, which outperforms the baseline in terms of both reconstruction and generalization error.

« back