ABSTRACT

The paper presents an architecture of a modular neural network for. recognition of patterns of high similarity. It consists of a recurrent single-layer network and a feedforward multilayer network. The idea of the single-layer network has been inspired by architectures proposed by McClelland and coauthors (McClelland and Rumelhart (1986, 1988), . McClelland et al. (1988), O'Reilly and McClelland (1994) and has been modified so that pattern separation is performed by the single-layer module in terms of increased Euclidean distance between the teaching patterns at the output of the first module for fasterfaster learning by the second. The paper presents the equilibrium and stability conditions of the first module of the network a well as tests showing that the modular network performs best for input data of high similarity in comparison with a feedforward network employing backpropagation. The proposed neural architecture is currently being employed in human-computer interaction context for recognition of individual user profiles in adaptive interface systems.

KEYWORDS: neural networks, pattern recognition