DSP-Inspired Deep Learning: A Case Study Using Ramanujan Subspaces

Abstract

Can Deep Learning be used to augment DSP techniques? Algorithms in DSP are typically developed starting from a mathematical model of an application. In some cases however, simplicity of the model can result in deterioration of performance when there is a severe modeling mis-match. This paper explores the idea of implementing a DSP technique as a computational graph, so that hundreds of parameters can jointly be trained to adapt to any given dataset. Using the specific example of period estimation by Ramanujan Subspaces, significant improvement in estimation accuracies under high noise and very short datalengths is demonstrated.1