Detection of Small Embedded Signals in Noise using Probabilistic Independent Component Analysis.pdfVIP

DETECTION OF SMALL EMBEDDED SIGNALS IN N OISE USING PROBABILISTIC INDEPENDENT COMPONENT ANALYSIS Stephen Roberts Rizwan Choudrey Pattern Analysis Research Group, University of Oxford, UK. October 30, 2001 Abstract Many communications problems involve the separation of signals in the presence of noise. In this paper we consider the use of ICA to recover very weak modulated signals. We show that a prob- abilistic approach gives considerably better results compared to alternative, kurtosis-based, algorithms. Introduction Independent Component Analysis (ICA) has been widely used in data analysis and decomposition in recent years (see, for example, [1] for an overview). ICA typically aims to solve the blind source sep- aration problem in which a set of unknown sources are mixed in some way to form the data that are observed, the classic example being that of multiple speakers in a room with several microphones. Some ICA algorithms exploit the fact that the probability density functions (pdfs) of mixtures of sources are more Gaussian (normally distributed) than the source pdfs themselves and hence forming source esti- mates whose pdfs are maximally non-Gaussian may achieve the goal of ICA [2, 3]. Alternatively, ICA may be seen as a generative model in a probabilistic framework (see, e.g. [4, 5, 6]). This has the advan- tage that, in principle, fully Bayesian models may be formulated in which priors exist over all the model parameters. This enables, for example, model selection (to determine e.g. how many sources there are) and the principled handling of uncertainty and noise. Recently the use of fully Bayesian, very flexible ICA models has been realised [7, 8]. 1 Theory N In general, we have a set of observation vectors,