Noisy group neurons with synchronous resetting for high-performance spiking neural networks

Yajie Zhai; Yanmei Kang; Meng Li; Zigang Huang

cs.CVAug 18, 2026

Noisy group neurons with synchronous resetting for high-performance spiking neural networks

Authors: Yajie Zhai, Yanmei Kang, Meng Li, Zigang Huang

Organizations: Department of Applied Mathematics, School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an 710049, China · Center for Intersection of Mathematics and Life Sciences, Xi’an Jiaotong University, Xi’an 710049, China · The Key Laboratory of Biomedical Information Engineering of Ministry of Education, Institute of Health and Rehabilitation Science, School of Life Science and Technology, Xi’an Jiaotong University, Xi’an 710049, China · Research Center for Brain-inspired Intelligence, Xi’an Jiaotong University, Xi’an 710049, China

Abstract

Spiking neural networks (SNNs), characterized by bio-inspired neuronal dynamics and event-driven communication, have attained significant progress in recent years. Nevertheless, training deep SNNs remains challenging due to spatiotemporal information loss and gradient mismatching. To simultaneously address these issues, we propose a noisy group neuron (NGN) model, which incorporates population-level synchronous resetting and neural stochasticity as fundamental computational mechanisms. We then develop the NGN method as a framework that combines the NGN model with backpropagation learning based on mean-field dynamics. We demonstrate the advantages of the NGN method through theoretical analysis and experimental validation on CIFAR-10, CIFAR-100, Tiny-ImageNet, DVS-Gesture, N-Caltech101, and CIFAR10-DVS. The proposed approach achieves an accuracy of 87.35% on CIFAR10-DVS within 10 inference time steps. These results support NGN as a practical approach to high-performance neuromorphic computing.

Figures & tables

Neuron	Metric	$\sigma=0$	$\sigma=0.25$	$\sigma=0.5$	$\sigma=0.75$	$\sigma=1$
Net neuron 1	MI	0.30	0.43	0.47	0.44	0.37
Net neuron 1	Acc(%)	91.19	93.22	94.50	70.27	9.23
Block1.neuron1	MI	0.21	0.32	0.35	0.33	0.28
Block1.neuron1	Acc(%)	92.37	93.60	94.50	76.20	9.09
Block1.neuron2	MI	0.41	0.54	0.55	0.49	0.41
Block1.neuron2	Acc(%)	92.12	93.82	94.50	79.75	10.12

Table 1: The relationship between neuronal mutual information and network performance.

Figure 2: Stochastic resonance behavior of the NGN model characterized by the input-output mutual information $MI\left(I^{t,n-1},\bar{O}^{t,n}\right)$ under the recorded presynaptic Gaussian inputs and non-Gaussian inputs. The Gaussian presynaptic input (a) and the corresponding input-output mutual information versus noise intensity are shown in the first row under different group sizes or iteration: (b) $t=1$ ; (c) $K=4$ . Here, the theoretical prediction is shown in solid curves and the simulated results are shown in dashed curves. The shaded regions represent one standard deviation of the data. The non-Gaussian input (d) and the corresponding mutual information are shown in the second row under different group sizes or iteration steps: (e) $t=1$ ; (f) $K=4$ . The initial noise intensity for training ResNet-18 is $\sigma_{0}=0.5$ . The quantile-quantile (Q-Q) plot embedded in (a) and (d) signifies how the presynaptic input deviates from Gaussian distribution. The non-monotonic curves in (b), (c), (e) and (f) manifest the occurrence of stochastic resonance.

Figure 3: Flowcharts of the fully connected $N$ -layer SNNs: (a) a general procedure and (b) the NGN-based procedure. Forward spike propagation and backward loss propagation are symmetric in gradient learning. Here, $\frac{\partial L}{\partial w^{n}}$ represents the gradient of the loss $L$ with respect to the weights $w^{n}$ for $n=1,2,\ldots$ .

Figure 5: Comparison of gradient mismatching levels and training loss between the baseline LIF method and the NGN method with different sizes $K$ . (a-c) The logarithm of the gradient mismatching probability $\log D_{\varepsilon}(o,o_{0})$ calculated according to Eq. ( 24 ) for $\varepsilon=0.1$ , $0.25$ , and $0.4$ , respectively. (d-e) Training loss curves on CIFAR-10 and CIFAR-100 datasets.

Dataset	Method	Architecture	Time steps	Top1-Accuracy
CIFAR-10	STBP-tdBN [ 39 ]	ResNet-19	6	93.16
	TET [ 42 ]	ResNet-19	6	94.50
	IM-LOSS [ 12 ]	ResNet-19	2	93.85
	NSNN [ 21 ]	CIFAR Net	4	94.30
	Diet-SNN [ 45 ]	VGG-16	4	92.76
	CLIF+TET [ 46 ]	ResNet-18	8	96.69*

Table 2: Comparison of the NGN method (

K=16

) with several existing methods on static image datasets, with * denoting the cooperation of Cutout and/or data auto-augmentation techniques.

Dataset	Method	Architecture	Time steps	Top1-Accuracy
DVS-Gesture	NSNN+TET [ 21 ]	7B-Net	16	96.88
	SSNN [ 50 ]	VGG-9	8	94.91
	TET † [ 42 ]	SNN-5	10	95.46
	DSGM+DTAM [ 51 ]	SNN-5	10	96.69
	NGN+TET	SNN-5	10	97.88 $\pm$ 0.29
N-Caltech101	SSNN [ 50 ]	VGG-9	8	79.25

Table 3: Comparison of the NGN method with TET loss implementation with the existing methods on neuromorphic image datasets, with † denoting the implemented results by us.

CIFAR-10
Neuron	Accuracy (%)	Loss
Vanilla LIF	93.94	0.3853
CLIF	94.00	0.3807
Vanilla NLIF	93.69	0.3981
Bernoulli NLIF	92.87	0.4056
NGN	95.17	0.3385

Table 4: Test accuracy and loss of neuron models on CIFAR datasets using ResNet-18 with TET (

T=4

K=8

Variant	$K$	Memory (MB)	Time/epoch (s)	Accuracy (%)
LIF (Gaussian surrogate)	1	2235.8	79.2	94.23
Noisy LIF	1	2237.4	81.8	93.32
NGN with synchronous resetting	8	2253.7	129.0	95.16
NGN w/o shared resetting	8	8933.8	791.5	95.10

Table 5: Component ablation on CIFAR-10. Peak memory and time per epoch characterize training cost, and Accuracy reports the best test accuracy over 200 epochs.

Explore similar work

Advancing Direct Training for Spiking Neural Networks with Circulate-Firing Neurons and Learnable Gradients
May 14, 2026Feifan Zhou, Xiang Wei, Yang Liu +1Spiking Neural Networks Neural Network Training Dynamics
School of Artificial and Intelligence, Tianjin University, Tianjin 300072, China
Intrinsically Stable Spiking Neural Networks: Overcoming the Performance Barrier in the Absence of Batch Normalization
Jun 30, 2026Ruichen Ma, Xiaoyang Zhang, Jian Bai +5Spiking Neural Networks
University of Electronic Science and Technology of China (UESTC) · Beijing Institute of Remote-Sensing Equipment · Shenzhen Institute for Advanced Study, UESTC
Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing
May 12, 2026Alex Fulleda-Garcia, Saray Soldado-Magraner, Josep Maria Margarit-TauléSpiking Neural Networks Neuromorphic Computing
Instituto de Microelectrónica de Barcelona (IMB-CNM), Consejo Superior de Investigaciones Cientíﬁcas (CSIC) Cerdanyola del Vallès, Spain · Department of Neurobiology, University of California Los Angeles (UCLA) Los Angeles, USA · Joint Senior authors

May 14, 2026, cs.NEJ/K move · Enter open · S save

Advancing Direct Training for Spiking Neural Networks with Circulate-Firing Neurons and Learnable Gradients

Feifan Zhou, Xiang Wei, Yang Liu, Qiang Yu

School of Artificial and Intelligence, Tianjin University, Tianjin 300072, China

Spiking Neural Networks Neural Network Training Dynamics

Open paper AskPDF arXiv

Spiking Neural Networks (SNNs) have emerged with promising energy-efficient property, yet a substantial performance gap persists compared to Artificial Neural Networks (ANNs). This gap stems from at least two key limitations: first, conventional spiking neurons offer limited information representation capacity, underutilizing the rich dynamics of membrane potentials; second, fixed surrogate gradient (SG) functions across time steps leads to imprecise gradient propagation, impeding effective direct training. To address these two challenges, we propose a new direct training algorithm with three core innovations: first, a circulate-firing spiking neuron model that enhances information representation capacity by leveraging membrane potentials more effectively; second, a time-step-wise learnable surrogate gradient function, enabling accurate gradient estimation during backpropagation; third, a positive-negative balanced loss function to achieve equilibrium between positive and negative membrane potentials and further boost SNN performance. Extensive experiments demonstrate that our methods achieve competitive performance across multiple datasets. Our methods can generalize seamlessly to advanced architectures of Transformer, consistently outperforming existing methods. Our work highlights the effectiveness of further harnessing intrinsic membrane dynamics of SNNs for performance improvement, and thus open a new avenue for advancing high-performance spiking neural architectures.

Abstract

Explore similar work

Advancing Direct Training for Spiking Neural Networks with Circulate-Firing Neurons and Learnable Gradients

Intrinsically Stable Spiking Neural Networks: Overcoming the Performance Barrier in the Absence of Batch Normalization

Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing