Spiking Neural Network With Weight-Sharing Synaptic Array for Multi-input Processing

Abstract

A multi-input processing spiking neural network inference system (MSS) is proposed to enhance the parallel processing capabilities of the spiking neural network (SNN) inference relative to the conventional SNN inference. Processing multiple input samples using MSS with shared synaptic arrays for compute-in-memory drastically reduces the number of synaptic arrays assigned for parallel processing, thereby minimizing the effort required by parallel processing networks. A shared weight array is evaluated using the 4-bit quantization capabilities of a manufactured ferroelectric field-effect-transistor as a synaptic device. A batch action was implemented as a multi-input on a 3-layer fully connected network to verify the MSS. The benefits of MSS in energy consumption and area throughput are rigorously investigated and estimated based on the number of multi-input processing. It is proven that the simultaneously processing of multiple input samples using the proposed MSS boosts the energy and area efficiency by up to 9.12 and 242 times, respectively.