artificial neural networks, modular architecture, comparative study

The majority of this paper relies on some forms of automatic decomposition tasks into modules. Both described methods execute automatic neural network modularization. Modules in neural networks emerge; we do not build them straightforward by penalizing interference between modules. The concept of emergence takes an important role in the study of the design of neural networks. In the paper, we study an emergence of modular connectionist architecture of neural networks, in which networks composing the architecture compete to learn the training patterns directly from the interaction of reproduction with the task environment. Network architectures emerge from an initial set of randomly connected networks. In this way can be eliminated connections so as to dedicate different portions of the system to learn different tasks. Mentioned methods were demonstrated for experimental task solving.