LMS to Deep Learning: How DSP Analysis Adds Depth to Learning

Abstract

Difficulty in analyzing deep learning systems is preventing association of its parameters and state outputs with elements that are derived from theoretic considerations. Medical and criminal justice communities are excited by the possibilities nevertheless reluctant to adopt machine learning for fear of errors and bias. There is also concern among educators that mystifying learning may have long-term adverse pedagogical implications on human learning. Deep learning techniques have not received enthusiastic attention in the realm of communication systems, either. This is in part due the effectiveness of traditional analytical solutions that render classification error rates larger that 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> unacceptable. Understanding of deep networks for communications, and ability to perform comparison tests can be useful in developing scalable methods to understand them in more complex scenarios. This paper contributes a perspective and analysis with focus on Nyquist and non-Nyquist pulse shapes for bandlimited channels. It is hoped that this fundamental presentation motivates wider consideration of neural networks and deep learning for demodulation.