Algorithms and Architectures for High Performance Kernel Adaptive Filtering

Abstract

In recent years, machine learning algorithms have been taking over traditional programming approaches and are being used to great effect in a broad range of applications. In this thesis, we look at kernel adaptive filters (KAFs): a class of non-linear adaptive filters which utilise a Mercer kernel function. These algorithms have been shown to provide high accuracy in a number of applications, including those that would require extremely high data rates, such as channel equalisation, extremely low latency, such as financial market prediction, and many other applications from embedded to cloud computing. The purpose of this thesis is to determine whether or not it is feasible to apply KAFs to such a range of applications. To do this, we propose new KAF algorithms which are hardware friendly in nature. We also explore exotic computing platforms from field programmable gate arrays (FPGAs) to distributed computing settings. We show that with these techniques, orders of magnitude increases in performance can be achieved and as such, KAFs can be applied to a wide range of problems.