MorphSNN: Adaptive Graph Diffusion and Structural Plasticity for Spiking Neural Networks

Yongsheng Huang; Peibo Duan; Yujie Wu; Kai Sun; Zhipeng Liu; Jiaxiang Liu; Guangyu Li; Changsheng Zhang

cs.NEMar 15, 2026

MorphSNN: Adaptive Graph Diffusion and Structural Plasticity for Spiking Neural Networks

Authors: Yongsheng Huang, Peibo Duan, Yujie Wu, Kai Sun, Zhipeng Liu, Jiaxiang Liu, Guangyu Li, Changsheng Zhang,

Organizations: School of Software, Northeastern University, China · Guangdong Institute of Intelligence Science and Technolog, China · The Hong Kong Polytechnic University, China · Monash University, Australia

Abstract

Spiking Neural Networks (SNNs) currently face a critical bottleneck: while individual neurons exhibit dynamic biological properties, their macro-scopic architectures remain confined within conventional connectivity patterns that are static and hierarchical. This discrepancy between neuron-level dynamics and network-level fixed connectivity eliminates critical brain-like lateral interactions, limiting adaptability in changing environments. To address this, we propose MorphSNN, a backbone framework inspired by biological non-synaptic diffusion and structural plasticity. Specifically, we introduce a Graph Diffusion (GD)mechanism to facilitate efficient undirected signal propagation, complementing the feedforward hierarchy. Furthermore, it incorporates a Spatio-Temporal Structural Plasticity (STSP) mechanism, endowing the network with the capability for instance-specific, dynamic topological reorganization, thereby overcoming the limitations of fixed topologies. Experiments demonstrate that MorphSNN achieves state-of-the-art accuracy on static and neuromorphic datasets; for instance, it reaches 83.35% accuracy on N-Caltech101 with only 5 timesteps. More importantly, its self-evolving topology functions as an intrinsic distribution fingerprint, enabling superior Out-of- Distribution (OOD) detection without auxiliary training. The code is available at anonymous.4open.science/r/MorphSNN-B0BC.

Figures & tables

Figure 1: The overall framework of MorphSNN. Input spikes flow from the direct encoding layer through $L$ stacked DGD-SNN layers to the final classifier. Top Left: The model exhibits robustness against noise, with orange (In-Distribution, ID) and blue (OOD) samples activating distinct graph structures for OOD detection. Top Center: Each node is implemented as a ConvBNSN triplet. Bottom Left: The GD executes undirected signal propagation using the Laplacian operator. Bottom Center: The STSP strengthens synapses via Hebbian plasticity while maintaining stability via homeostatic plasticity, driving the temporal evolution of the entire graph structure ( Right ).

Method	Architecture	$\mathbf{T}$	DVS-Gesture	CIFAR10-DVS	NCaltech101
MLF ( Feng et al., 2022 )	VGG-9	5	85.77	65.88	70.4
SSNN ( Ding et al., 2024 )	VGG-9	5	90.74	73.63	77.97
Spikingformer ( Zhou et al., 2023 )	Spikingformer-2-256	5	93.40 ∗	77.50 ∗	82.14 ∗
CogniSNN ( Huang et al., 2025b )	WS-RGA-7	5	95.41	79.00	78.62
MorphSNN	DGD-SNN-7-3	5	96.53 $\hskip 0.85005pt\scriptstyle\pm\hskip 0.8245pt0.11$	79.90 $\hskip 0.85005pt\scriptstyle\pm\hskip 0.8245pt0.13$	83.35 $\hskip 0.85005pt\scriptstyle\pm\hskip 0.8245pt0.18$
SEW-ResNet ( Fang et al., 2021a )	ConvNet	16	97.92	74.40	-

Table 1: Performance (%) comparison with state-of-the-art methods on DVS-Gesture, CIFAR10-DVS, and N-Caltech101. Bold and underline denote the best and second-best results, respectively. Results marked with ∗ denote our reproduction.

Method	$\mathbf{T}$	CIFAR-10	CIFAR-100	ImageNet
ResNet-19 (ANN)	1	94.97	75.35	-
Transformer (ANN)	1	96.73	81.02	-
MorphANN (Ours)	1	96.50	80.29	-
SNASNet ( Kim et al., 2022 )	5	93.64	73.04	-
Spikformer ( Zhou et al., 2022 )	4	95.51	78.21	73.38
Spikingformer ( Shi et al., 2024 )	4	95.81	78.21	74.79

Table 2: Comparison (%) on static CIFAR-10/100 and ImageNet.

Method	GD	STSP	Sym	DVS128	C10DVS	NCal101
Baseline	✗	✗	✗	92.01	78.90	81.77
Static	✓	✗	✓	95.44	79.90	83.35
Asym	✓	✓	✗	94.44	78.20	82.14
Full	✓	✓	✓	96.53	79.40	82.75

Table 3: Ablation studies (%) . Sym denotes symmetric.

Model	$\mathbf{T}$	Accuracy (%)
Event Frames ( Bi et al., 2020 )	8	57.9
Res-SNN18 + RM ( Yao et al., 2023b )	8	63.5
TIM ( Shen et al., 2024b )	10	63.8
MorphSNN (Static)	10	65.81
MorphSNN (Full)	10	67.12

Table 4: Comparison and STSP ablation on UCF101-DVS.

Figure 2: Robustness evaluation of Static and Full models on DVS-Gesture under varying interference intensities.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 3 — Figure 3: Hyperparameter sensitivity analysis on DVS-Gesture.

Stage	Neuromorphic Datasets			Static Image Datasets
	DVS-Gesture	C10DVS / NCal101	UCF101-DVS	CIFAR-10 / 100	ImageNet
	(Input: $128\times 128$ )	(Input: $128\times 128$ )	(Input: $64\times 64$ )	(Input: $32\times 32$ )	(Input: $224\times 224$ )
Encoder	{3x3, 32, AP}	{3x3, 156, AP}	{3x3, 156}	{3x3, 64}	{7x7, 64, S2, P3, AP}
1	$\left[\begin{array}[]{c}3\text{x}3,32\end{array}\right]\times 7$	$\left[\begin{array}[]{c}3\text{x}3,156\end{array}\right]\times 7$		$\left[\begin{array}[]{c}3\text{x}3,128\end{array}\right]\times 4$	$\left[\begin{array}[]{c}3\text{x}3,64\end{array}\right]\times 8$
2	$\left[\begin{array}[]{c}3\text{x}3,32\end{array}\right]\times 7$	$\left[\begin{array}[]{c}3\text{x}3,156\end{array}\right]\times 7$		$\left[\begin{array}[]{c}3\text{x}3,256\end{array}\right]\times 4$	$\left[\begin{array}[]{c}3\text{x}3,128\end{array}\right]\times 8$
3	$\left[\begin{array}[]{c}3\text{x}3,32\end{array}\right]\times 7$	$\left[\begin{array}[]{c}3\text{x}3,156\end{array}\right]\times 7$		$\left[\begin{array}[]{c}3\text{x}3,512\end{array}\right]\times 4$	$\left[\begin{array}[]{c}3\text{x}3,256\end{array}\right]\times 16$

Appendix

Table 6: Comparison of network architectures across all datasets

Dataset	Optimizer	Weight Decay	Batch Size	Epochs	Init. LR	Schedule
CIFAR-10	AdamW	0.05	64	400	$1\times 10^{-3}$	Cosine
CIFAR-100	AdamW	0.05	64	400	$1\times 10^{-3}$	Cosine
ImageNet	AdamW	0.05	256	300	$6\times 10^{-4}$	Cosine
DVS-Gesture	SGD	0	8	300	$1\times 10^{-3}$	Cosine
CIFAR10-DVS	SGD	0	8	192	$1\times 10^{-3}$	Step
N-Caltech101	SGD	0	8	192	$1\times 10^{-3}$	Step

Appendix

Table 7: Hyperparameter settings for different datasets.

Figure 4: Training Metrics over Epochs on DVS-Gesture.

Figure 5: Visualization of feature maps generated by directed connections (DC) and our graph diffusion (GD).

Figure 6: Visualization of dynamic topological evolution for a single sample in DVS-Gesture.

Figure 7: t-SNE visualization of ID (DVS-Gesture) and OOD (DVS-Lip) distributions between DGP and other baselines.

Panel of Figure 8 — Figure 8: Averaging Firing Rate of each node.

Method	T	Time	Param	Fire Rate	ACs	MACs	SOP Energy	Accuracy
Method	T	(s)	(M)	(%)	(M)	(M)	( $\mathbf{\mu}$ J)	(%)
MorphANN	1	32	11.03	-	0.29	400.20	1841.18	96.50
Baseline	4	70	10.64	10.07	68.24	1.77	69.558	92.01
Static	4	72	10.85	9.91	69.69	1.77	70.863	95.97
Full	4	121	11.54	9.86	68.10	3.69	78.264	95.54

Appendix

Table 8: Computation Cost of different models on CIFAR-10.

Method	Mem. Read & Write			Synaptic & Neuron Op. (mJ)	Addr. ( $\mathbf{\mu}$ J)	Total (mJ)
Method	Membrane Potential (mJ)	Parameters (mJ)	In / Out (mJ)	Synaptic & Neuron Op. (mJ)	Addr. ( $\mathbf{\mu}$ J)	Total (mJ)
MorphANN	0	39.8431	39.8192	1.8412	0.5563	81.5041
Baseline	13.2920	6.6460	0.0004	0.0696	59.9272	20.0676
Static	13.3569	6.6785	0.0005	0.0709	60.2313	20.1666
Full	13.5729	6.7214	0.0074	0.0691	60.3827	20.4313

Appendix

Table 9: Detailed energy consumption breakdown for different methods on CIFAR-10.

Spiking Neural Networks

Spiking neural networks process information through discrete, time-dependent neural events, aiming for efficient computation and biologically inspired learning.

See all 351 papers

Adaptive Spiking Neurons for Vision and Language Modeling
Apr 14, 2026Chenlin Zhou, Sihang Guo, Jiaqi Wang +5Spiking Neural Networks
School of Electronic and Computer Engineering, Peking University, Shenzhen, China · Harbin Institute of Technology, Shenzhen, China · School of Computer Science, Peking University, Beijing, China +1
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
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

Apr 14, 2026J/K move · Enter open · S save

Adaptive Spiking Neurons for Vision and Language Modeling

Chenlin Zhou, Sihang Guo, Jiaqi Wang, Dongyang Ma+4

School of Electronic and Computer Engineering, Peking University, Shenzhen, China · Harbin Institute of Technology, Shenzhen, China · School of Computer Science, Peking University, Beijing, China+1

Spiking Neural Networks

Open paper AskPDF arXiv

Regarded as the third generation of neural networks, Spiking Neural Networks (SNNs) have garnered significant traction due to their biological plausibility and energy efficiency. Recent advancements in large models necessitate spiking neurons capable of high performance, adaptability, and training efficiency. In this work, we first propose a novel functional perspective that provides general guidance for designing the new generation of spiking neurons. Following the insightful guidelines, we propose the Adaptive Spiking Neuron (ASN), which incorporates trainable parameters to learn membrane potential dynamics and enable adaptive firing. ASN adopts an integer training and spike inference paradigm, facilitating efficient SNN training. To further enhance robustness, we propose a specialized variant of ASN, the Normalized Adaptive Spiking Neuron (NASN), which integrates normalization to stabilize training. We evaluate our neuron model on 19 datasets spanning five distinct tasks in both vision and language modalities, demonstrating the effectiveness and versatility of the ASN family. Our ASN family is expected to become the new generation of general-purpose spiking neurons.

Abstract

Spiking Neural Networks

Explore similar work

Adaptive Spiking Neurons for Vision and Language Modeling

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

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