Neuro-ICA: Novel Learning Algorithms for  Feature Extraction

and Noise Removal

Project Outcomes

•   Development of a two-layer neural network ICA architecture maximizing the entropy of the outputs with logistic transfer function.

• Design of a neural architecture for nonlinear three-layer ICA network for estimating the respective basis vectors. They are counterparts of the PCA eigenvectors, but characterize the data in many cases better. The purpose of the three layers is: a) whitening of the input data for yielding good separation results; b) separation of independent sources (components); c) estimation of the basis vectors.

• Design, modification and implementation of nonlinear constrained Hebbian rules – one-unit case and multi-unit case, bigradient  learning rule of Wang etc. for learning with different criterion functions.

• Setting the experimental parameters: image data, sampling, data preprocessing, the algorithm and its parameters, estimation of statistical quantities.