Design and training for combinational neural-logic systems

Abstract

This paper presents the combinational neural-logic system. The basic components, i.e., the neural-logic-AND, -OR, and -NOT gates, will be proposed. As different applications have different characteristics, a traditional neural network with a common structure might not handle every application well if some network connections are redundant and cause internal disturbances, which may downgrade the training and network performance. In this paper, the proposed neural-logic gates are the basic building blocks for the applications. Based on the knowledge of the application and the neural-logic design methodology, a combinational neural-logic system can be designed systematically to incorporate the characteristics of the application into the structure of the combinational neural-logic system. It will enhance the training and network performance. The parameters of the combinational neural-logic system will be trained by the genetic algorithm. To illustrate the merits of the proposed approach, the combinational neural-logic system will be realized practically to recognize Cantonese speech commands for an electronic book.