Piecewise-linear classifiers, formal neurons and separability of the learning sets

The design of piecewise-linear classifiers from formal neurons is considered. The design classifiers are based on hierarchical, multilayer neural networks. The described procedure allows to find both the structure of network (the numbers of layers and neurons) and weights of single neurons. The main principle of the synthesis procedure is to preserve separability of learning sets during data compression by successive neural layers. Different procedures aiming at improving the network compression ability are also considered.