Fractal learning of fast orthogonal neural networks
Alexander Yu. Dorogov · Optical Memory and Neural Networks · 2012
A new method of learning fast one- and multiple-dimensional orthogonal transformations is considered. Tunable orthogonal transformations are regarded as special neural networks. The learning takes a finite number of steps. The learning algorithm does not have the error feedback and is absolutely stable. The method is based on fractal filtering of signals and images. Linguistic models are used to determine the topology and structure of fast transformations. Examples are given.