Spikeq enables efficient low-bit memory use in spiking neural networks
SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks
Neural and Evolutionary ComputingArtificial IntelligenceComputer Vision and Pattern Recognition
Summary
Spiking neural networks can save lots of energy on special brain-inspired chips but need very tiny memory for their connections and signals. The authors found that traditional ways to reduce memory size don’t work well because they ignore how errors affect the network’s spiking activity. So they designed SpikeQ, which carefully picks memory values by watching how spikes change, keeping the network accurate even with very few bits. Tests show SpikeQ keeps accuracy high while using much less memory and energy compared to older methods.
What this means in practice
- •For neuromorphic hardware engineers: Improve memory and energy efficiency of spiking neural network models deployed on neuromorphic processors by using spike-guided quantization.
- •For embedded system developers: Reduce the memory footprint and power consumption of AI models in low-resource devices that use spiking neural networks for image analysis tasks.
Authors
Kai Sun, Peibo Duan, Zhanglu Yan, Guowei Zhang, Yongsheng Huang, Levin Kuhlmann
Abstract
On neuromorphic chips, spiking neural networks (SNNs) consume far less energy than artificial neural networks (ANNs), but the limited memory of these chips requires low-bit synaptic weights and membrane potentials. Existing post-training quantization (PTQ) methods provide them without retraining and, like those for ANNs, minimize numerical errors such as the mean squared error (MSE). In an SNN, however, a quantization error reaches later layers only through the spikes it flips. We find that errors of equal MSE can cause very different accuracy losses, whereas errors that flip the same fraction of spikes cause the same loss. We therefore propose SpikeQ, a spike-guided PTQ framework. For weights, we select the level of each weight by simulating the downstream neurons and minimizing their spike errors. For membrane potentials, most existing methods quantize the state after the reset, whereas neuromorphic chips compare the threshold with the membrane potential stored at a fixed bit width. We target this pre-fire quantization and build a reset-aware grid that places the threshold on a rounding boundary, so rounding never flips the spike at the current time step. We evaluate SpikeQ on seven benchmarks of image classification, semantic segmentation, and object detection. With 2-bit pre-fire membrane potentials, every existing grid collapses, whereas SpikeQ retains most of the full-precision accuracy. With both at 4 bits, the models stay within 2.59 points of full precision, shrink their memory footprint 7.8--7.9 times, and use 6.4--6.9 times less energy.