Recently, various neural network ( NN ) LSIs have been developed with practical applications of NN. In real-time handling of large problems in fields such as image and speech signal processing, however, the integration and processing speed of these LSIs are still insufficient. For higher integration of a digital NN LSI, a method of simplifying neuron models was examined. A new neuron model with weights expressed with power of two has been proposed as a simplified model. In this model, multiplications of the input signal and the weight are replaced with shift operations. Therefore, the circuitry scale reduction is achieved. Character recognition and learning simulations were carried out to examine the validity of the simplified neuron model. As a result, it was found that the simplified model had the character recognition ability almost equal to these of conventional models. The learning simulation revealed that the amount of time necessary to converge with the simplified model were longer than these with conventional ones. Finally, a simplified neuron whose application was specified to the recognition was designed. The total pattern area of one neuron was 1.47×1.66mm². The new structure, by which the network structure of the NN LSI could be changed has been proposed and used for the design of neurons.
�@