Unsupervised Learning and Blind Source Separation
Francis F. Li, Trevor J. Cox · 2019
Unsupervised learning is often used in signal processing information extraction, dataset dimensionality reduction, signal separation, and many more situations. This chapter considers two important unsupervised neural network models, namely principal component analysis (PCA) and independent component analysis (ICA) neural networks. ICA-based blind signal separation and de-convolution techniques show potential usefulness in modelling the cocktail party effect and separating signals that are in the same frequency range but statistically independent. PCA can reduce variable numbers while maintaining the most important feature of observed data. ICA neural networks, evolved from PCA networks, are unsupervised models for separating statistically independent signals and were applied to speech separation problems by A. J. Bell and T. J. Sejnowski. The properties of input data that the neural network focuses on and the compressed formats that the neural network takes depend on the learning strategy and the network architecture used. The chapter discusses the envelope detection and envelope spectrum estimation algorithms are identical to the ones.