Quantization of Spiking Neural Networks Beyond Accuracy

Evan Gibson Smith; Jacob Whitehill; Fatemeh Ganji

cs.LGApr 15, 2026

Quantization of Spiking Neural Networks Beyond Accuracy

Authors: Evan Gibson Smith, Jacob Whitehill, Fatemeh Ganji

Organizations: Worcester Polytechnic Institute Worcester, Massachusetts, USA

Abstract

Quantization is a natural complement to the sparse, event-driven computation of Spiking Neural Networks, reducing memory bandwidth and arithmetic cost for deployment on resource-constrained hardware. However, existing SNN quantization evaluation focuses almost exclusively on accuracy, overlooking whether a quantized network preserves the firing behavior of its full-precision counterpart. We demonstrate that quantization method, clipping range, and bit-width can produce substantially different firing distributions at equivalent accuracy, differences invisible to standard metrics but relevant to deployment, where firing activity governs effective sparsity, state storage, and event-processing load. To capture this gap, we propose Earth Mover's Distance as a diagnostic metric for firing distribution divergence, and apply it systematically across weight and membrane quantization on SEW-ResNet architectures trained on CIFAR-10 and CIFAR-100. We find that uniform quantization induces distributional drift even when accuracy is preserved, while LQ-Net style learned quantization maintains firing behavior close to the full-precision baseline. Our results suggest that behavior preservation should be treated as an evaluation criterion alongside accuracy, and that EMD provides a principled tool for assessing it.

Figures & tables

Table 1. Weight quantization comparison across clipping ranges and architectures. At higher bit-widths, wide clipping ([-10,10]) achieves accuracy comparable to narrow ([-1,1]) but produces substantially higher EMD, revealing pathological firing dynamics invisible to accuracy alone. LQ-Net preserves dynamics across all configurations. At 2-bit with wide clipping, uniform quantization collapses in accuracy while LQ-Net remains robust. Best quantized accuracy per section is bolded . Columns: Acc = Accuracy (%), Fire = Mean Firing Rate, EMD = EMD from FP32, Dead = Dead Neurons (%)

CIFAR-100 (FP32: R18: 75.2%, R8: 72.3%)
		ResNet18				ResNet8
Range	Bits	Acc	Fire	EMD	Dead	Acc	Fire	EMD	Dead
Signed	2	67.1	.041	.0139	82.2	66.6	.063	.0283	67.7
Unsigned	2	71.9	.033	.0104	84.9	70.7	.053	.0247	73.0
Signed	4	74.8	.073	.0075	67.3	72.4	.091	.0132	42.2
Unsigned	4	73.3	.077	.0097	63.3	71.4	.083	.0143	47.7

Table 2. Membrane potential quantization across datasets. Unsigned ranges ([0,1]) provide more robust dynamics than signed ([-1,1]) across both CIFAR-100 and CIFAR-10, with lower EMD from the FP32 baseline at equivalent or higher accuracy. Best quantized accuracy per section is bolded. Columns: Acc = Accuracy (%), Fire = Firing Rate, EMD = EMD from FP32, Dead = Dead Neurons (%)

Figure 1. Firing rate distributions across quantization methods at 2-bits on SEW-ResNet18 for the CIFAR-100 task for the first spiking layer (sn1).

Figure 2. Mean firing rates across layers of trained networks with various weight quantization techniques for 2-bit SEW-ResNet18 trained on the CIFAR-100 task. Layer labels follow SEW-ResNet nomenclature: layerX.Y.sn1 denotes the spiking layer in block Y of stage X; the leftmost sn1 is the stem spiking layer.

Figure 3. Mean firing rates across layers of trained networks with membrane quantization techniques for 2-bit SEW-ResNet18 trained on the CIFAR-100 task. Layer labels follow SEW-ResNet nomenclature: layerX.Y.sn1 denotes the spiking layer in block Y of stage X; the leftmost sn1 is the stem spiking layer.

Figure 4. EMD from FP32, mean firing rate, and dead neuron percentage across bit-widths for uniform (STE) and learned (LQ-Net) weight quantization under narrow [-1,1] and wide [-10,10] clipping on SEW-ResNet18 (CIFAR-100).

Figure 5. EMD, dead neuron percentage, and mean firing rate across bit-widths for signed and unsigned membrane quantization on SEW-ResNet18 (CIFAR-100). All three metrics converge toward full-precision levels as precision increases, supporting EMD as a diagnostic for firing behavior recovery.

Neural Network Quantization

Neural network quantization represents weights, activations, and states with fewer bits to reduce memory use, computation, and deployment costs.

See all 394 papers

SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks
Oct 16, 2026Kai Sun, Peibo Duan, Zhanglu Yan +3Spiking Neural Networks Post-Training Quantization
Monash University · National University of Singapore · Northeastern University
PTQ4SNN: Membrane-Aware Post-Training Quantization for Spiking Neural Networks
Aug 7, 2026Hui Xie, Tong Shi, Haotong Qin +3Mixed-Precision Quantization 4-Bit Quantization
State Key Laboratory of Complex & Critical Software Environment, Beihang University · School of Artificial Intelligence, Beihang University · Center for Project-Based Learning, D-ITET, ETH Zurich +2
BASC : Behavior-Aligned Quantization and Pruning for Low-Bit Spiking Neural Networks
Aug 12, 2026Linliang Chen, Yan Zhong, Xin Liu +2Spiking Neural Networks Neural Network Quantization
Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China · School of Mathematical Sciences, Peking University, Beijing 100871, China · School of Electronic and Information Engineering; Beihang University, Beijing 100191, China +1

Oct 16, 2026J/K move · Enter open · S save

SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks

Kai Sun, Peibo Duan, Zhanglu Yan, Guowei Zhang+2

Monash University · National University of Singapore · Northeastern University

Spiking Neural Networks Post-Training Quantization Neuromorphic Computing Activation Quantization

Open paper AskPDF arXiv

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.

Abstract

Neural Network Quantization

Explore similar work

SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks

PTQ4SNN: Membrane-Aware Post-Training Quantization for Spiking Neural Networks

BASC : Behavior-Aligned Quantization and Pruning for Low-Bit Spiking Neural Networks