High reliability or fault tolerance is one of the most important properties of artificial neural network systems. Whether it can be realized and how it is realized have very important in- fluence on the learning ability, expandation ability and normal operation of the corresponding network systems. In this paper some new neurobiological discoveries on the neural network system of human brain are stated, which is used as the prototype and the basis of the research and design on artificial neural network systems. This paper pointed out that all the artificial neural network sytems and their constructed components, such as calculating neurons and their weighted connections, are degradated sys terns or degradated components, therefore there exists the possibility of degradation failure. Whereas the degradation of systems is related with that of their constructed components. The learning abilities of artificial neural network systems should include the normal work of sys terns even under not ideal operation conditions where there exists component failure or circum stancial disturbance. Therefore the research and design on artificial neural network systems are extremely necessary and very important. Additionally. an advanced neural network sys tern should have strong plasticity and expandation ability, i.e. the trained network can have correct response to the new samples which do not appear in the training sample set, therefore to suit the complex applications. The multiple layer perceptron neural network model based on back propagation algorithms is the most extensively applied model, and is generally acknowledged as one kind of high fault tolerant model. But the analysis and simulation given in this paper pointed out that the fault tolerance of BP model neural network systems mainly emtxxties in their network con struction, where the neighboring layers are interconnected and there exist multiple parallel hid den units, and the time redundancy of the back propagation algorithms, which is actually a kind of software redundancy. Therefore once such kind of network is constructed, the fault tolerance of the corresponding systems is only and mainly a kind of software fault tolerance. So if some components of the systems are failure, the recovery of the nornal learning abilities of the systems is by extending the sample training time of the systems. But such discovery is very limited and is time wasting. There:Fore some further and comprehensive research and de sign on the reliability and fault tolerance of artificial neural network systems is needed to corn pensate the insufficiency of the single software fault tolerance. The simulation and the analysis of the system output error in this paper show that the fail ure of different components can have different influence on the learning abilities of the sys terns, based on which the defination of :important calculating neurons and their weighted
修改评论