2020 Volume 14 Issue 1 Pages 60-72
Deep learning techniques can be used for not only learning deep neural networks but also internal parameter optimization of differentiable iterative algorithms. By embedding learnable parameters into an excellent iterative algorithm, we can construct a flexible derived algorithm with data-driven learnability. This approach is called deep unfolding. We present an overview of deep unfolding and its features, focusing on sparse signal recovery algorithms. In the first half of this paper, examples of deep unfolding including a sparse signal recovery algorithm, TISTA, will be presented. We observed the convergence acceleration phenomenon for deep unfolding-based algorithms. In the second half of the paper, our theoretical results (spectral radius control based on the Chebyshev step) for convergence acceleration are outlined.
