A pulse model neuron with high accurate calculation and its application to two-dimensional binary classification

In order to scale down circuit size for the hardware realization of neural network systems, there has been much research on the conventional pulse model neuron, in which a conventional numerical calculation is replaced by a stochastic calculation. In the conventional stochastic calculation, a linear feedback shift register (LFSR) has been used as a random number generator. In general, a good randomness is required for an exact calculation result. However, since random numbers generated by LFSR are lacking in randomness, there is a problem which affects the accuracy of the conventional stochastic calculation which uses LFSR to generate random numbers. In order to improve the accuracy of the conventional calculation, we propose to use GLFSR (generalized LFSR) which generates random numbers which are rich in randomness. The accuracy of the stochastic calculation is evaluated using a root-mean-square error (RMS) and a maximum error. As a result, the proposed stochastic calculation, which uses GLFSR as the random number generator, shows a highly accurate calculation result. Moreover, when it is applied to a neural network to solve a two-dimensional binary classification problem, the network shows better classification results than those of the neural network which uses the conventional stochastic calculation.