Optimal method for growth in dynamic self organizing learning systems

Self-organization in learning systems refers to the ability of the system to adapt and respond data as it is presented without outside intervention. The ability to self-organize is desirable and is critical in realizing on-line and real-time adaptive systems for applications including control systems, navigation, vision, and speech. In this paper, we focus on self-organizing learning systems which utilize the addition and subtraction of receptor nodes or neurons in some way. Typically, these algorithms store error information and use it to modify the neural network dynamically so that this error will be decreased. Examples of these learning systems include self-organizing maps, growing cell structures, and dynamic cell structures. We describe current methods for growing the number nodes in the case of the Dynamic Cell Structures neural network and discuss issues via examples that could lead to potentially incorrect data representation during the implementation of the algorithm. A new algorithm is provided that overcomes the observed flaw and enables these learning systems to grow and operate in an optimal and robust manner. The analysis of the proposed optimal growing algorithm indicates that this modified algorithm is more reliable for use in on-line and real-time adaptive systems.