Neural network nonlinear factor analysis of high dimensional binary signals.
Dušan Húsek, Hana Řezanková, Václav Snåšel, Alexander Frolov, Pavel Yu. Polyakov · 2005
Possible application of a new neural network suitable for binary factorization of signals of large dimension and complexity is introduced. We developed the new recall procedure of Hoppfield-like associative memory which allows search all attractors corresponding to factors (a true attractor). Necessary separation of spurious attractors is based on calculation of their Lyapunov function. Being applied to textual data the procedure allows to reveal groups of highly correlated words (factors) which frequently occur in documents jointly and represent topics of that documents. 1.