How the internal state subnetwork works

This paper presents a neural network architecture that performs pattern classification using constructional simple and compact form of recurrent connections. The work was motivated by the desire to decrease computational complexity and to maintain a greater degree of modularity in the neural network design. The main use of the network can be expected in various speech applications, in which the context information is crucial. A new method to extend the design of the multilayer perceptron topology is introduced. The method uses an additional recurrent subnetwork module between two subsequent and processing layers of the feedforward network. The analysis has shown that the perceptron with a subnetwork module can be efficiently used for speech signal processing where it gives matching results with the standard time delay neural network (TDNN). The network is tested for two trajectory classification problems and has shown good results