Weight Partitioning for Dynamic Fixed-Point Neuromorphic Computing Systems

Abstract

Neuromorphic computing systems consist of neurons and synapses with limited programmability, and neural networks are modified to be mapped for such a system. In order to map a perceptron with a large number of connections into a hardware neuron with a fixed, small number of synapses, it is decomposed into a tree of perceptrons, which substantially affects the neuron usage and predictive performance. In this paper, we propose two decomposition algorithms that take advantage of plastic connections and the dynamic scaling capability of neurons. One algorithm based on sorting considers the neuron usage first, and the other algorithm based on packing considers the predictive performance first. Our experimental results on two popular deep convolutional neural networks showed that the sorting-based algorithm substantially improved the accuracy at no cost of neurons compared to a previous work, and the packing-based algorithm improved it even further at a small cost of neurons.