Biologically inspired neural fuzzy computational models

Abstract

In many practical applications, the complete set of training exemplars is usually unavailable to construct a classification or forecasting system. In other practical applications, a static system is rendered ineffective by the changing property of the data source. Therefore, the presentation of information becomes sequential. Adaptation to the system and forecasting by the system have to be alternatively carried out. However to apply existing batch learning computational models to these applications proves unfeasible. In addition, learning from a dynamic data stream poses several difficulties. They are: (1) the limitation on the use of past explicit information poses difficulties in extending existing learning techniques; (2) the limited prior knowledge poses difficulties in designing a learning paradigm that can work without much prior information; (3) the sensitivity of learning methodology on the choice of the training parameters; and (4) robustness of learning procedure against noisy or spurious data.