Neuronal Self-Adaptation Enhances Capacity and Robustness of Representation in Spiking Neural Networks

Zhuobin Yang; Yeyao Bao; Liangfu Lv; Jian Zhang; Xiaohong Li; Yunliang Zang

cs.LGMar 21, 2026

Neuronal Self-Adaptation Enhances Capacity and Robustness of Representation in Spiking Neural Networks

Authors: Zhuobin Yang, Yeyao Bao, Liangfu Lv, Jian Zhang, Xiaohong Li, Yunliang Zang

Organizations: Academy of Medical Engineering and Translational Medicine, Tianjin University, Tianjin, China.

Abstract

Spiking Neural Networks (SNNs) are promising for energy-efficient, real-time edge computing, yet their performance is often constrained by the limited adaptability of conventional leaky integrate-and-fire (LIF) neurons. Existing LIF models struggle with restricted information capacity and susceptibility to noise, leading to degraded accuracy and compromised robustness. Inspired by the dynamic self-regulation of biological potassium channels, we propose the Potassium-regulated LIF (KvLIF) neuron model. KvLIF introduces an auxiliary conductance state that integrates membrane potential and spiking history to adaptively modulate neuronal excitability and reset dynamics. This design extends the dynamic response range of neurons to varying input intensities and effectively suppresses noise-induced spikes. We extensively evaluate KvLIF on both static image and neuromorphic datasets, demonstrating consistent improvements in classification accuracy and superior robustness compared to existing LIF models. Our work bridges biological plausibility with computational efficiency, offering a neuron model that enhances SNN performance while maintaining suitability for low-power neuromorphic deployment.

Figures & tables

Panel of Figure 1 — Figure 1: Simulation comparison of the KvLIF model and conventional LIF variants regarding dynamic response range and noise robustness. LIF-S: LIF model with soft reset; LIF-H: LIF model with hard reset.

Figure 2: (a) Schematic of biological action potential regulation involving two distinct potassium channel subtypes. (b) Computational flow of the LIF model with soft reset. (c) Computational flow of the proposed KvLIF model.

Figure 3: Comparison of dynamic responses between the KvLIF model and two conventional LIF models.

Panel of Figure 4 — Figure 4: Comparative performance of KvLIF and conventional LIF models. (a) Gradient L2 norm distribution. (b) Performance retention as inference time steps reduce.

Neuron	0.04	0.08	0.12	0.16	0.20
LIF-S	65.58	53.96	40.81	29.74	21.57
LIF-H	71.60	61.23	44.35	30.57	21.33
GLIF	69.71	62.38	50.40	38.19	28.05
CLIF	75.57	66.71	60.71	40.83	29.93
ILIF	74.52	65.33	50.24	36.49	25.44
KvLIF	76.10	68.37	63.03	44.27	33.32

Table 3: Robustness test on CIFAR100 under different noise levels.

Neuron	0.10	0.20	0.30	0.40	0.50
LIF-S	90.63	84.03	77.78	69.10	66.32
LIF-H	77.78	68.40	60.07	55.56	52.08
GLIF	79.86	79.86	77.08	67.36	58.68
CLIF	89.58	82.64	75.69	63.89	57.29
ILIF	86.46	76.74	70.83	61.46	56.25
KvLIF	93.75	89.93	86.11	80.90	75.69

Table 4: Robustness test on DVS-Gesture under different noise levels.

Shunting	Adaptive
Settings		CIFAR10	CIFAR100	Tiny-ImageNet	CIFAR10-DVS	DVS-Gesture
Synaptic Inputs	Repolarization
✓	✓	95.55	79.30	64.82	79.50	98.26
✗	✓	95.46	79.22	64.23	78.60	97.57
✓	✗	94.26	77.43	62.82	77.60	97.22
✗	✗	94.09	77.05	62.70	77.20	96.18

Table 5: Ablation study of the proposed KvLIF neuron on different datasets.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Heatmap visualization of the encoding layer responses to different input samples.

Dataset	Bacth Size	Epoch	Learning Rate	Optimizer	Weight Decay	$\alpha$	$\beta$	$\gamma$
CIFAR10	128	200	0.1	SGD	5e-5	0.8	0.3	0.05
CIFAR100	128	300	0.1	SGD	5e-4	0.8	0.1	0.05
Tiny-ImageNet	256	300	0.1	SGD	5e-4	0.8	0.3	0.05
CIFAR10-DVS	128	200	0.05	SGD	5e-4	0.8	0.1	0.05
DVS-Gesture	16	150	5e-4	Adam	0.0	0.8	0.3	0.05

Appendix

Table 6: Training hyperparameter settings for various datasets.

Dataset	Architecture	Neuron	Timestep	Accuracy (%)
CIFAR10	ResNet-18	LIF-S	4 / 6	91.54 / 94.09
	ResNet-18	LIF-H	4 / 6	94.42 / 94.70
	ResNet-18	KvLIF	4 / 6	95.28 / 95.55
CIFAR100	ResNet-18	LIF-S	4 / 6	69.24 / 77.05
	ResNet-18	LIF-H	4 / 6	74.64 / 77.17
	ResNet-18	KvLIF	4 / 6	77.76 / 79.30

Appendix

Table 7: Performance comparison with conventional LIF models.

Panel of Figure 6 — Figure 6: Firing rate comparison across different datasets.

Dataset	Architecture	Neuron	T	ACs (M)	MACs (M)	SOP Energy ( $\mu$ J)
CIFAR10	ResNet-18	ReLU	1	0.00	549.13	2526.00
	ResNet-18	LIF-S	6	512.82	3.34	476.91
	ResNet-18	LIF-H	6	487.93	3.34	454.51
	ResNet-18	KvLIF	6	393.16	13.37	415.34
CIFAR100	ResNet-18	ReLU	1	0.00	549.18	2526.22
	ResNet-18	LIF-S	6	359.34	3.34	338.78

Appendix

Table 8: The energy consumption of synaptic operation on different datasets.

Neuron	0.04	0.08	0.12	0.16	0.20
LIF-S	90.65	86.74	79.91	70.96	60.75
LIF-H	93.72	90.87	86.38	79.59	71.82
GLIF	93.64	91.48	87.92	82.05	74.97
CLIF	94.44	92.46	88.40	82.69	74.98
ILIF	94.31	92.96	88.77	83.63	76.61
KvLIF	94.84	92.73	89.12	84.12	77.02

Appendix

Table 9: Robustness test on CIFAR10 under different noise levels.

Neuron	0.10	0.20	0.30	0.40	0.50
LIF-S	75.20	73.80	69.30	65.30	58.80
LIF-H	75.90	74.70	69.20	66.30	61.20
GLIF	70.20	61.40	51.00	40.60	31.30
CLIF	75.40	73.20	69.60	66.10	60.30
ILIF	75.80	75.70	72.80	70.00	64.40
KvLIF	77.10	75.90	74.40	72.10	69.00

Appendix

Table 10: Robustness test on CIFAR10-DVS under different noise levels.

Neuron	0.10	0.20	0.30	0.40	0.50
LIF-S	96.53	95.83	94.10	92.36	89.93
LIF-H	95.83	95.14	90.28	78.47	60.07
GLIF	95.83	96.18	96.18	94.79	93.06
CLIF	96.18	95.49	94.44	93.06	90.97
ILIF	96.86	96.53	94.44	92.01	88.54
KvLIF	97.22	96.53	96.53	95.83	94.10

Appendix

Table 11: Temporal Robustness test on DVS-Gesture under different noise levels.

Neuron	0.10	0.20	0.30	0.40	0.50
LIF-S	75.40	74.10	71.50	66.60	55.10
LIF-H	77.40	76.00	74.00	71.40	61.60
GLIF	74.10	69.70	63.80	55.50	43.70
CLIF	77.90	76.30	74.10	68.50	57.70
ILIF	78.10	76.20	74.60	71.60	63.80
KvLIF	79.00	77.90	74.70	71.10	65.00

Appendix

Table 12: Temporal Robustness test on CIFAR10-DVS under different noise levels.

Explore similar work

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
DVA-Neurons: Design and Verification of Adaptive LIF Neurons: From Single-Neuron Dynamics to Multi-Neuron Spiking Networks
Sep 19, 2026Thanh Pham, Riadul IslamSpiking Neural Networks Neuromorphic Computing
Department of Computer Science and Electrical Engineering, University of Maryland, Baltimore County, MD 21250, United States
ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization
May 3, 2026Kaiwen Tang, Di Yu, Jiaqi Zheng +4Spiking Neural Networks Neural Network Quantization
Zhejiang University · Shandong University

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

Explore similar work

Adaptive Spiking Neurons for Vision and Language Modeling

DVA-Neurons: Design and Verification of Adaptive LIF Neurons: From Single-Neuron Dynamics to Multi-Neuron Spiking Networks

ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization