Year: 2022
Type: preprint
Abstract: Spiking neural networks (SNNs) have attracted interest in recent years due to their low energy consumption and the increasing need for more power in real-life ML-related applications. Having those bio-inspired networks on neuromorphic hardware for extra-low energy consumption is another exciting aspect of this technology. Furthermore, many works discuss the improvement of SNNs in terms of performance and hardware implementation. This paper presents a progressive layer-based compression approach applied to convolutional spiking neural networks trained with unsupervised STDP. Moreover, we study the effect of this approach when used with SpiNNaker. This approach, inspired by neuroplasticity, produces highly compressed networks (up to 90% compression rate) while preserving the same network performance or slightly improving it, as shown by experimental results using MNIST, FMNIST, Caltech face/motorbike, STL-10, and CIFAR-10 datasets.
Source: HAL
Authors: Elbez, Hammouda, Fatahi, Mazdak
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4307655332