Modified procedures for teaching elementary neural network blind separation and proposed architecture on the standard elements of digital technology, allowing to solve the problem in the mode of self-real-time using a polynomial activation functions. He received further development of neural processing method of multivariate mixtures of signals based on the sharing of independent analysis and principal component analysis and show that it can be realized both within the basic single-layer architecture by alternate use of different learning algorithms and architectures in the framework of the type "bottleneck". Modified recursive procedures "whitening", separation and estimation of basis vectors of independent components, providing high speed of convergence, and characterized by simplicity of numerical implementation.
neural network, image components, analysis, synthesis, adaptation procedure
“Vosstanovlenie izobrazhenii s ispolzovaniem analiza glavnykh i nezavisimykh komponent”,
Information Processing Systems,