Analog/digital circuit simplification for Hopfield neural network

Abstract

Circuit realization of neural networks is a significant approach in neuromorphic computing. Researchers have simplified the circuit for single neuron model, but one for neural network is rarely reported, yet. This paper presents a ReLU-type Hopfield neural network (RHNN) model using simple ReLU function instead of traditional hyperbolic-type function as the activation function. RHNN with three neurons is focused on. Its boundedness and stability are confirmed theoretically, and complex chaotic dynamics are simulated numerically. Further, the printed circuit board (PCB)-based analog circuit is fabricated, and the RHNN circuit has 49.3 % fewer analog components than the Tanh-type HNN circuit. Meanwhile, with logical shift method, an efficient low-cost multiplierless digital circuit is developed on field-programmable gate array (FPGA) platform. Experimental results manifest that resource consumptions of one are less than that of IP core-based digital implementation. Particularly, the RHNN model is well applied to image encryption for satisfying requirements on transmission security.