Collaborative adaptive filtering for machine learning

Abstract

Quantitative performance criteria for the analysis of machine learning architectures and algorithms have long been established. However, qualitative performance criteria, which identify fundamental signal properties and ensure any processing preserves the desired properties, are still emerging. In many cases, whilst offline statistical tests exist such as assessment of nonlinearity or stochasticity, online tests which not only characterise but also track changes in the nature of the signal are lacking. To that end, by employing recent developments in signal characterisation, criteria are derived for the assessment of the changes in the nature of the processed signal. Through the fusion of the outputs of adaptive filters a single collaborative hybrid filter is produced. By tracking the dynamics of the mixing parameter of this filter, rather than the actual filter performance, a clear indication as to the current nature of the signal is given. Implementations of the proposed method show that it is possible to quantify the degree of nonlinearity within both real- and complex-valued data. This is then extended (in the real domain) from dealing with nonlinearity in general, to a more specific example, namely sparsity. Extensions of adaptive filters from the real to the complex domain are non-trivial and the differences between the statistics in the real and complex domains need to be taken into account. In terms of signal characteristics, nonlinearity can be both split- and fully-complex and complex-valued data can be considered circular or noncircular. Furthermore, by combining the information obtained from hybrid filters of different natures it is possible to use this method to gain a more complete understanding of the nature of the nonlinearity within a signal. This also paves the way for building multidimensional feature spaces and their application in data/information fusion. To produce online tests for sparsity, adaptive filters for sparse environments are investigated and a unifying framework for the derivation of proportionate normalised least mean square (PNLMS) algorithms is presented. This is then extended to derive variants with an adaptive step-size. In order to create an online test for noncircularity, a study of widely linear autoregressive modelling is presented, from which a proof of the convergence of the test for noncircularity can be given. Applications of this method are illustrated on examples such as biomedical signals, speech and wind data.