Nonlinear Blind Source Separation by Variational Bayesian Learning

摘要: Blind separation of sources from their linear mixtures is a well understood problem. However, if the are nonlinear, this problem becomes generally very difficult. This because both nonlinear mapping and underlying must be learned data in blind manner, highly ill-posed without suitable regularization. In our approach, multilayer perceptrons used as generative models for data, variational Bayesian (ensemble) learning applied finding sources. The technique automatically provides reasonable regularization paper, we first consider static mixing model, with successful application to real-world speech compression. Then discuss extraction dynamic processes, detection abrupt changes process dynamics. difficult test chaotic approach clearly outperforms currently available prediction change techniques. proposed methods computationally demanding, but they can problems higher dimensions than other existing approaches.