Realtime classification with structure adapting neuro-fuzzy systems

Abstract

Neural networks with a structure adaptation capability that are equivalent to fuzzy systems are investigated with the goal of designing hardware architectures for application in time critical classification problems. The consecutive steps adopted to build a realtime prototype of a structure adapting neural network include the modification of learning algorithms making them suitable for hardware implementation, the development of a hardware architecture, and the system implementation and test. Structure adaptation must be integrated into the learning algorithm implementable in hardware. Fusion of information hidden in data with the existing knowledge expressed in rules should also be possible. Three types of structure adapting algorithms have been mapped onto the proposed hardware architecture. Performance of the system implementation is evaluated comparing the speeds using software implementations for benchmarks and real applications.