Fault-tolerant design of mutually coupled neural networks based on the duplication of neurons with the higher functionality

摘要

Since the integration technology has been well developed to implement a large number of operational Units in a LSI chip, it is becoming possible to use such neuro-chips for highly parallel information processing. However we may encounter problems in implementing neural networks in neuro-chips, because constituent components of a neural network are complicated and further a neural network has numerous coupling elements among its constituent components. Thus, we cannot ignore the possibility that neuro-chips may suffer the occurrence of their faults. In this study, we propose a new simplified model of neural networks, in which each neuron with the higher functionality is duplicated to enhance the fault tolerance.