Multivariate Density Factorization for Independent Component Analysis: An Unsupervised Artificial Neural Network Approach
Abstract
We propose a novel homogenous nonlinear self-organising network which employs solely computationally simple hebbian and anti-hebbian learning, in approximating a linear independent component analysis (ICA). The learning algorithms diagonalise the transformed data covariance matrix and approximate an orthogonal rotation which maximises the sum offourth order cumulants. This provides factorisation of the input multivariate density into the individual independent latent marginal densities. We apply this network to linear mixtures of data, which are inherently non-gaussian and have both Laplacian and bi-modal probability densities. We show that the proposed network is capable of factorising multivariate densities which are linear mixtures of independent latent playkurtic, leptokurtic and uniform distributions.
Cite
Text
Girolami and Fyfe. "Multivariate Density Factorization for Independent Component Analysis: An Unsupervised Artificial Neural Network Approach." Proceedings of the Sixth International Workshop on Artificial Intelligence and Statistics, 1997.Markdown
[Girolami and Fyfe. "Multivariate Density Factorization for Independent Component Analysis: An Unsupervised Artificial Neural Network Approach." Proceedings of the Sixth International Workshop on Artificial Intelligence and Statistics, 1997.](https://mlanthology.org/aistats/1997/girolami1997aistats-multivariate/)BibTeX
@inproceedings{girolami1997aistats-multivariate,
title = {{Multivariate Density Factorization for Independent Component Analysis: An Unsupervised Artificial Neural Network Approach}},
author = {Girolami, Mark and Fyfe, Colin},
booktitle = {Proceedings of the Sixth International Workshop on Artificial Intelligence and Statistics},
year = {1997},
pages = {223-230},
volume = {R1},
url = {https://mlanthology.org/aistats/1997/girolami1997aistats-multivariate/}
}