Precision Exploration of Floating-Point Arithmetic for Spiking Neural Networks

초록

In this paper, we explore the precision of various floating-point representations for energy-efficient spiking neural network (SNNs). The IEEE 754 based 32-bit single-precision floating-point and reduced precision floating-point formats are applied to the leaky integrate-and-fire (LIF) neuron of the SNN to investigate the impact the reduced precision on the accuracy performance. When adopted in an unsupervised two-layer SNN for the MNIST digit recognition application, the 16-bit floatingpoint formats can be used in training and inference of the SNN without any classification performance degradation. Additionally, our experimental result reveals that the floating-point format with 4-bit exponent and 6-bit mantissa is enough for the SNN training and inference and offers great area, power, and energy reductions compared to the others.

키워드