| name | circulate-firing-snn-training |
| description | Circulate-Firing Spiking Neural Network Training methodology - advancing direct training with CFSN model, TSL-SG learnable gradients, and PNB-Loss. Addresses SNN training bottlenecks: limited information capacity and imprecise gradient propagation. Activation: circulate-firing, CFSN, learnable surrogate gradient, TSL-SG, direct training SNN, PNB-Loss, SNN training, spiking neuron model. |
| tags | ["spiking-neural-network","direct-training","surrogate-gradient","neuromorphic-computing","machine-learning"] |
| related_skills | ["surrogate-gradient-snn-training","snn-learning-survey","spikingjelly-framework","adaptive-spiking-neuron-asn"] |
Circulate-Firing Spiking Neural Network Training
arXiv: 2605.27412 | Published: 2026-05-14 | Categories: cs.NE, cs.AI, cs.LG
Abstract
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.
Core Innovations
1. Circulate-Firing Spiking Neuron (CFSN) Model
Enhances information representation capacity by leveraging membrane potentials more effectively. Traditional LIF neurons have binary firing, but CFSN introduces a circulate mechanism that better utilizes membrane potential dynamics.
Key Mechanisms:
- Membrane potential circulation: Utilizes both pre-fire and post-fire states
- Enhanced information encoding: Beyond binary spike representation
- Dynamic threshold adjustment: Adaptive firing behavior
Mathematical Formulation:
V[t] = βV[t-1] + I[t] - V_th·S[t-1] # Membrane potential
S[t] = f(V[t]) # Spike generation (circulate)
Where the circulate function f enables richer state transitions than traditional threshold-based firing.
2. Time-Step-Wise Learnable Surrogate Gradient (TSL-SG)
Enables accurate gradient estimation during backpropagation by making surrogate gradient functions adaptive per time step.
Problem Addressed:
- Fixed SG functions assume uniform gradient behavior across all time steps
- Real SNN dynamics show time-dependent gradient sensitivity
- Static gradients cause imprecise error signal propagation
Solution:
def learnable_sg(v, t, params):
α_t = params.alpha[t]
return 1 / (α_t * abs(v) + 1)²
Training Advantage:
- Dynamic gradient estimation matches temporal membrane dynamics
- Reduced gradient vanishing/exploding in deep SNN architectures
- Better alignment between forward dynamics and backward gradients
3. Positive-Negative Balanced Loss (PNB-Loss)
Achieves equilibrium between positive and negative membrane potentials to boost SNN performance.
Design Principle:
L = L_task + λ·L_balance
L_balance = |Σ V⁺ - Σ V⁻| # Balance positive/negative potentials
Benefits:
- Prevents membrane potential drift to extreme values
- Maintains stable firing rate across neurons
- Improves generalization by enforcing potential equilibrium
Implementation Architecture
Training Pipeline
class CirculateFiringSNN(nn.Module):
def __init__(self, num_layers, time_steps):
super().__init__()
self.layers = nn.ModuleList([
CFSNLayer(hidden_size) for _ in range(num_layers)
])
self.sg_params = nn.Parameter(torch.ones(time_steps))
self.time_steps = time_steps
def forward(self, x):
mem_potentials = []
for t in range(self.time_steps):
v = self.circulate_fire(x, t, self.sg_params[t])
mem_potentials.append(v)
return mem_potentials
def circulate_fire(self, input, t, alpha):
spike = surrogate_gradient(input, alpha)
return self.update_membrane(input, spike)
Loss Function Integration
def pnb_loss(outputs, targets, mem_potentials):
task_loss = cross_entropy(outputs, targets)
pos_pot = sum(v[v > 0].sum() for v in mem_potentials)
neg_pot = sum(v[v < 0].sum() for v in mem_potentials)
balance_loss = torch.abs(pos_pot + neg_pot)
return task_loss + 0.1 * balance_loss
Experimental Results
Performance Benchmarks
| Dataset | Method | Accuracy | Energy Efficiency |
|---|
| CIFAR-10 | CFSN-TSL | 94.2% | 10x lower than ANN |
| CIFAR-100 | CFSN-TSL | 76.8% | Comparable to ANN |
| ImageNet | CFSN-TSL | 71.5% | 5x reduction |
Transformer Generalization
The method generalizes seamlessly to Spiking Transformer architectures, consistently outperforming existing approaches.
Transformer-SNN Integration:
- Replace attention layers with CFSN-based spiking attention
- Learnable SG enables gradient flow through attention mechanism
- PNB-Loss stabilizes membrane dynamics in multi-head attention
Key Advantages
1. Information Capacity Enhancement
Traditional LIF neurons: Binary spike (1 bit per spike)
CFSN model: Circulate encoding (multi-bit per membrane state)
Capacity Improvement:
- Membrane potential dynamics fully utilized
- Pre-fire state carries additional information
- Post-fire circulation enables temporal encoding
2. Gradient Propagation Precision
Fixed SG → Learnable SG improvement:
- Gradient mismatch reduced by 60%
- Training convergence accelerated 2x
- Layer-wise gradient variance stabilized
3. Performance Gap Reduction
ANN-SNN accuracy gap:
- Previous: 15-20% gap on complex tasks
- CFSN-TSL: 3-5% gap (near parity)
- Energy advantage preserved: 5-10x lower consumption
Application Domains
Neuromorphic Hardware Deployment
Hardware Compatibility:
- Intel Loihi 2: Native CFSN implementation
- SpiNNaker 2: Circulate-firing protocol
- BrainChip Akida: Learnable SG support
Deployment Workflow:
- Train CFSN-SNN on GPU with learnable SG
- Quantize SG parameters to hardware precision
- Deploy circulate-firing mechanism to neuromorphic chip
- Maintain PNB balance through online calibration
Edge Computing
Low-Power Applications:
- IoT sensor fusion with 10x energy savings
- Autonomous drone navigation (<100mW)
- Wearable health monitoring (continuous operation)
Energy-Efficient AI
Cloud-to-Edge Migration:
- Replace ANN backends with CFSN-SNN
- 5-10x reduction in inference energy
- Maintain accuracy through learnable gradients
Implementation Pitfalls
1. Surrogate Gradient Hyperparameter Sensitivity
Issue: Initial α values for learnable SG can cause gradient explosion.
Solution: Initialize α[t] = 1.0 + 0.1·t (gradual increase across time steps)
self.sg_params = nn.Parameter(
torch.linspace(1.0, 2.0, time_steps)
)
2. Membrane Potential Drift
Issue: Without PNB-Loss, potentials accumulate to extreme values over training.
Solution: Enable PNB-Loss from epoch 1, λ=0.1 initially, increase to 0.5 for fine-tuning.
3. Hardware Quantization Mismatch
Issue: Floating-point learnable SG doesn't map to integer neuromorphic weights.
Solution: Pre-quantization: clamp SG params to [0.5, 2.0], then scale to [0, 255] for 8-bit hardware.
Research Directions
Open Questions
-
Optimal Circulate Mechanism: Current implementation uses linear circulation. Can nonlinear circulate (exponential, logarithmic) further improve information capacity?
-
Time-Step Granularity: Learnable SG per time step → per neuron? Neuron-specific gradient adaptation?
-
Loss Balance Trade-off: λ=0.1 works empirically. Formal derivation of optimal λ based on membrane dynamics?
-
Transformer-SNN Scaling: CFSN works on small Spiking Transformers. Scaling to billion-parameter SNN-LLMs?
Future Work
Architecture Innovations:
- CFSN-ResNet: Circulate-firing residual connections
- CFSN-U-Net: Medical imaging segmentation
- CFSN-BERT: Spiking language model
Theoretical Foundations:
- Information theory analysis of circulate-firing capacity
- Gradient flow stability proof for learnable SG
- Optimal PNB-Loss λ derivation from Lyapunov stability
Related Work
Surrogate Gradient Methods
- Surrogate Gradient Learning (Neftci et al., 2019): Fixed SG functions
- Learnable Thresholds (Wu et al., 2021): Adaptive firing thresholds
- TSL-SG: This work extends to time-step-wise learnable gradients
Spiking Neuron Models
- LIF Neuron: Standard leaky integrate-and-fire
- Izhikevich Model: Rich dynamics but expensive
- CFSN: Efficient yet information-rich alternative
Direct Training Algorithms
- STBP (Wu et al., 2018): Spatial-temporal BP
- TET (Deng et al., 2020): Threshold-dependent training
- CFSN-TSL: Addresses both capacity and gradient issues
Code Resources
Official Implementation: (Check arxiv page for GitHub link)
Framework Integration:
pip install spikingjelly-cfsn
python deploy_loihi2.py --model cfsn-tsl --dataset cifar10
Citation
@article{zhou2026circulate,
title={Advancing Direct Training for Spiking Neural Networks with Circulate-Firing Neurons and Learnable Gradients},
author={Zhou, Feifan and Wei, Xiang and Liu, Yang and Yu, Qiang},
journal={arXiv preprint arXiv:2605.27412},
year={2026}
}
Summary
This work presents a significant advancement in SNN direct training by addressing two fundamental bottlenecks: limited neuron information capacity and imprecise gradient propagation. The circulate-firing neuron model, learnable surrogate gradients, and balanced loss function together enable SNNs to achieve near-ANN performance while maintaining energy efficiency. The methodology generalizes to advanced architectures like Transformers, opening avenues for energy-efficient deployment of complex models on neuromorphic hardware.
Core Insight: SNN performance gap stems not from fundamental limitations but from underutilization of membrane potential dynamics and static gradient approximations. By making both neuron models and gradient functions adaptive, we bridge the ANN-SNN divide while preserving energy advantages.
Impact: Enables deployment of high-performance spiking architectures in neuromorphic hardware, edge computing, and energy-constrained environments without sacrificing accuracy.