Research map / Architectures and efficiency / Spiking networks
Efficiently Training Time-to-First-Spike Spiking Neural Networks from Scratch
Kaiwei Che, Zhengyu Ma, Yifan Huang, Peng Xue, Li Yuan, Wei Fang, Yonghong Tian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ETTFS-init, average-pooling, weight normalization with affine and temporal weighting decoder together (ablation) | SCNN5 TTFS SNN · Fashion-MNIST | accuracy (%): 89.61 -> 92.90 | Improved |
| replace Kaiming init with ETTFS-init (ablation) | SCNN5 TTFS SNN · Fashion-MNIST | accuracy and convergence speed | Improved |
| add weight normalization on top of ETTFS-init (ablation) | SCNN5 TTFS SNN · Fashion-MNIST, CIFAR10, DVS Gesture | accuracy | Improved |
| use ETTFS-init instead of Kaiming init in deeper networks (5 to 11 layers) (ablation) | TTFS SNN · Fashion-MNIST | accuracy | Improved |
| use ETTFS-init instead of Kaiming init for gradient scale (ablation) | SCNN5 · weight gradients | gradient magnitude: Kaiming 1e-7 to 1e-10 -> ETTFS-init larger than 1e-3 | Improved |
| replace TQ-TTFS decoder with temporal weighting decoder (ablation) | TTFS SNN · CIFAR10 | average inference time-steps: 4.62 -> 2.93 | Improved |
| replace TQ-TTFS decoder with temporal weighting decoder (ablation) | TTFS SNN · CIFAR10 | estimated energy E1: 20.59uJ -> 12.19uJ | Improved |
| replace TQ-TTFS decoder with temporal weighting decoder | TTFS SNN · DVS Gesture | accuracy (%): 88.89 -> 95.83 | Improved |
| use average-pooling instead of max-pooling (ablation) | TTFS SNN · one-spike constraint | spike constraint preservation | Improved |
| increase decay factor gamma in temporal weighting decoder (ablation) | temporal weighting decoder · Fashion-MNIST, CIFAR10, DVS Gesture | accuracy: optimal gamma = 3 for Fashion-MNIST (exp) and CIFAR10 (linear) | Mixed |
| choose exponential vs linear decay in the decoder (ablation) | temporal weighting decoder · DVS Gesture | accuracy | No clear effect |
| initialize from pretrained ReLU weights | ETTFS · CIFAR10 | accuracy (%): 90.56 -> 93.89 | Improved |
Closest papers by meaning
- Parallel Training Time-to-First-Spike Spiking Neural Networks (AAAI 2026 · similarity 0.95)
- Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold: Learning with Event-Driven Backpropagation (ICCV 2023 · similarity 0.94)
- Going Deeper With Directly-Trained Larger Spiking Neural Networks (AAAI 2021 · similarity 0.93)
- Advancing Training Efficiency of Deep Spiking Neural Networks through Rate-based Backpropagation (NeurIPS 2024 · similarity 0.91)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.91)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
