12.4.1 Independent Component Analysis - Pattern Recognition and Machine Learning

Описание: In this section, we make a brief venture out beyond linear-Gaussian latent variable models and discuss independent component analysis (ICA). Motivated by demixing independent conversations from microphone signals carrying their mixtures, in ICA we look for latent variables that are not just decorrelated, but independent. However, independence alone isn't enough, as the rotational redundancy present in the latents that PCA extracts demonstrates. For ICA to extract the latents (up to permutations, etc.) we additionally require that their marginal distributions have heavy tails.