By name: circulate-firing-snn-direct-training description: Direct training algorithm for SNNs with circulate-firing neurons and learnable surrogate gradients. Three core innovations for membrane potential dynamics optimization. skill_type: research_methodology paper_id: arXiv:2605.27412 paper_title: Advancing Direct Training for Spiking Neural Networks with Circulate-Firing Neurons and Learnable Gradients paper_date: 2026-05-14 authors: Feifan Zhou, Xiang Wei, Yang Liu, Qiang Yu activation_keywords: - circulate-firing-snn - learnable-surrogate-gradient - direct-snn-training - membrane-potential-dynamics - spiking-transformer related_domains: - spiking neural networks - neuromorphic computing - machine learning - neural architecture
Circulate-Firing SNN Direct Training
Direct training algorithm for Spiking Neural Networks (SNNs) with three core innovations that leverage intrinsic membrane dynamics for performance improvement.
Three Core Innovations
1. Circulate-Firing Spiking Neuron Model
- Enhanced information capacity leveraging membrane potentials effectively
- Rich dynamics utilizing full membrane potential trajectory
- Better information encoding through circulate dynamics
2. Time-Step-Wise Learnable Surrogate Gradient
- Adaptive gradients not fixed across all time steps
- Accurate estimation enabling precise gradient propagation
- Learning optimization improving training convergence
3. Positive-Negative Balanced Loss Function
- Equilibrium between positive and negative membrane potentials
- Performance boost for SNN systems
- Stability preventing potential imbalance issues
Key Findings
- Competitive performance across multiple datasets
- Architecture generalization with Transformer architectures
- Consistent outperformance of existing direct training methods
- New pathway for advancing high-performance spiking architectures
Applications
- Training SNNs with direct methods
- Improving SNN performance on benchmarks
- Implementing SNN-Transformer architectures
- Neuromorphic computing optimization research
Related Skills
- snn-learning-survey
- surrogate-gradient-snn-training
- direct-to-event-snn-transfer
- spiking-transformer-unification
References
- Paper arXiv 2605.27412 (14 May 2026)
- Category cs.NE cs.AI cs.LG
- DOI 10.48550/arXiv.2605.27412