[论文解读] Ternary Spiking Neural Networks Enhanced by Complemented Neurons and Membrane Potential Aggregation

Boxuan Zhang, Jiaxin Wang+2 more|arXiv (Cornell University)|Jan 22, 2026

Advanced Memory and Neural Computing被引用 0

一句话总结

本文提出 Complemented Ternary Spiking Neurons (CTSN) 与 Temporal Membrane Potential Regularization (TMPR)，解决三值脉冲神经网络中的迭代信息丢失、时间梯度消失与膜电位不规则性问题，在静态和类脑数据集上实现了最先进的结果。

ABSTRACT

Spiking Neural Networks (SNNs) are promising energy-efficient models and powerful framworks of modeling neuron dynamics. However, existing binary spiking neurons exhibit limited biological plausibilities and low information capacity. Recently developed ternary spiking neuron possesses higher consistency with biological principles (i.e. excitation-inhibition balance mechanism). Despite of this, the ternary spiking neuron suffers from defects including iterative information loss, temporal gradient vanishing and irregular distributions of membrane potentials. To address these issues, we propose Complemented Ternary Spiking Neuron (CTSN), a novel ternary spiking neuron model that incorporates an learnable complemental term to store information from historical inputs. CTSN effectively improves the deficiencies of ternary spiking neuron, while the embedded learnable factors enable CTSN to adaptively adjust neuron dynamics, providing strong neural heterogeneity. Furthermore, based on the temporal evolution features of ternary spiking neurons' membrane potential distributions, we propose the Temporal Membrane Potential Regularization (TMPR) training method. TMPR introduces time-varying regularization strategy utilizing membrane potentials, furhter enhancing the training process by creating extra backpropagation paths. We validate our methods through extensive experiments on various datasets, demonstrating remarkable performance advances.

研究动机与目标

• Motivate the limitations of binary/ternary spiking neurons in information capacity and biological plausibility.
• Propose CTSN to store historical input information via a learnable complemental term.
• Introduce TMPR to regularize membrane potential evolution and improve gradient flow.
• Demonstrate improved performance on CIFAR-10/100, ImageNet-100, and CIFAR10-DVS across backbones.
• Provide analysis linking method to reduced gradient vanishing and smoother membrane potentials.

提出的方法

• Introduce CTSN by adding a learnable complemental term h(t) in the integration process to preserve historical information.
• Define the complemented membrane potential tilde u(t) = h(t) + x(t) and update with a learnable function G to produce o(t).
• Learnable parameters (alpha, beta, gamma) control the complement dynamics, constrained via sigmoid to stay in (0,1).
• Propose TMPR as a spatio-temporal regularizer based on the aggregation of membrane potentials across layers and timesteps, added to the cross-entropy loss.
• Provide a gradient analysis showing CTSN adds temporal backpropagation paths and TMPR adds direct backprop paths to mitigate gradient vanishing.
• Ablate CTSN and TMPR to show their individual and joint contributions to performance.

实验结果

研究问题

• RQ1 Can a complemented ternary spiking neuron preserve historical information to mitigate iterative information loss?
• RQ2 Does a time-aware regularization of membrane potentials improve gradient flow and training stability in ternary SNNs?
• RQ3 How do CTSN and TMPR affect membrane potential distributions and neuron heterogeneity across datasets?
• RQ4 What performance gains do CTSN and TMPR yield on standard static datasets (CIFAR-10/100, ImageNet-100) and neuromorphic data (CIFAR10-DVS) compared to existing ternary SNNs?

主要发现

• CTSN with a learnable complemental term significantly improves temporal gradient propagation and smooths membrane potential distributions.
• TMPR provides additional backpropagation paths by regularizing the squared membrane potential, enhancing gradient flow.
• On CIFAR-10/100 with ResNet backbones, CTSN yields top-tier accuracy, e.g., CIFAR-10: 96.46% (4 timesteps); CIFAR-100: 81.19% (4 timesteps).
• On ImageNet-100, the method achieves 83.78% accuracy with 4 timesteps, approaching or surpassing competing SNN methods.
• On CIFAR10-DVS (neuromorphic), results include 79.06% (VGG16, 10 timesteps) and 81.23% (ResNet20, 10 timesteps), with VGGSNN achieving 83.20% under open-code replication.
• Ablation studies confirm CTSN outperforms plain ternary spiking neurons across timesteps, and TMPR further boosts performance (lambda sensitivity analyzed).