ANALOG HARDWARE IMPLEMENTATIONS OF ARTIFICIAL NEURAL NETWORKS

摘要

There are several possible implementations of artificial neural network that are based either on software or hardware systems. Software implementations are rather inefficient due to the fact that the intrinsic parallelism of the underlying computation is usually not taken advantage of in a mono-processor kind of computing system. Existing hardware implementations of ANNs are efficient as the dedicated datapath used is optimized and the hardware is usually parallel. Hardware implementations of ANNs may be either digital, analog, or even hybrid. Digital implementations of ANNs tend to be of high complexity, thus of high cost, and somehow imprecise due to the use of lookup table for the activation function. On the other hand, analog implementation of ANNs are generally very simple and much more precise. In this paper, we focus on possible analog implementations of ANNs. The neuron is based on a simple operational amplifier. The reviewed implementations allow for the use of both negative and positive synaptic weights. An alternative implementation permits the realization of the training process.