Implementation of a stochastic neural network system

摘要

In this paper, we discuss the effectiveness of stochastic logic for implementation of neural networks, and show the design of a such network with DBM learning function. One can convert analogue quantity to pulse firing rate by stochastic logic. In a stochastic logic system, the number of transistor in a chip can be reduced greatly since various complex operation can be done with basic logic gate. Furthermore, we incorporate a non-monotonic function as an activation function in this system and confirm higher learning ability compared to a usual one.