A Photonic System for Wideband Online Principal Component Analysis using Unsupervised Learning

Abstract

Principal Component Analysis is a statistical technique used for dimensionality reduction that can also be used for decorrelating inputs. This work discusses the significance and applicability of PCA to wideband radio frequency signals. While RF communication is often approached in the domain of electronic systems, we argue that the superior bandwidth performance, low noise and fan-in scalability associated with optics necessitate the use of a photonic system. However, while many photonics based systems can handle wideband signals at radio frequencies, they lack unsupervised learning capabilities due to the absence of a fast feedback mechanism, rendering them incapable of performing PCA online.In this thesis, we propose the design of a photonic system and implement an iterative learning algorithm that uses unsupervised learning to tune system parameters fast enough for real time analysis in dynamic environments, overcoming these challenges, and enabling the online principal component analysis of wideband signals. This is demonstrated on four partially correlated channels carrying 13-GBd signals over optical fibres, with the iterative control performed on a readily available and easily programmable FPGA in the form of a Programmable System-on-a-Chip. This work also discusses the applications of this technique to wideband signals at radio frequencies, particularly in the field of communication, and potentially path-breaking implications to open problems such as blind source separation.