Research map / Architectures and efficiency / Spiking networks
Enhancing Training of Spiking Neural Network with Stochastic Latency
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Stochastic Latency Training (SLT), which samples a random inference latency per batch during direct SNN training so a single model degrades gracefully at lower latencies while training faster.
Problem
Directly trained spiking neural networks are optimized for one fixed latency and lose accuracy when run at lower latencies, while higher latencies increase training and inference energy and time.
Approach
Sample the simulation latency uniformly from {1..T} for each batch during training (SLT), which in expectation down-weights losses at high latencies and yields a single model robust across latencies; combined with TET or tdBN losses, with a theoretical bound showing SLT-TET upper bounds SLT-SDT in expectation.
Key results
- SLT-TET reaches 95.26% on CIFAR-10 at T=6 with ResNet-19, versus 94.50% for TET at T=6.
- SLT-tdBN improves DVS-CIFAR10 accuracy at T=1 by 28.12 points over tdBN (Table 3), and SLT-tdBN gives 1.69x training speedup (82.9 vs 139.7 s/epoch) on CIFAR-10.
- SLT-TET achieves 81.46% on DVS-CIFAR10 at T=10, comparable to TET's 81.56%, with a 1.67x training speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add stochastic latency training to TET | TET · CIFAR-10 | accuracy: 94.50 -> 95.26 | Helped |
| add stochastic latency training to tdBN | tdBN · CIFAR-10 | accuracy: 93.16 -> 94.66 | Helped |
| add stochastic latency training to TET | TET · CIFAR-100 | accuracy: 74.72 -> 74.87 | Helped |
| add stochastic latency training to tdBN | tdBN · CIFAR-100 | accuracy: 71.12 -> 71.34 | Helped |
| add stochastic latency training to TET | TET · DVS-CIFAR-10 | accuracy: 81.56 -> 81.46 | Hurt |
| add stochastic latency training to tdBN | tdBN · DVS-CIFAR-10 | accuracy: 67.8 -> 79.27 | Helped |
| add stochastic latency training to tdBN | tdBN · CIFAR-100, T=1 inference | accuracy: 56.65 -> 64.89 | Helped |
| add stochastic latency training to tdBN | tdBN · DVS-CIFAR-10, T=1 inference | accuracy: 36.56 -> 64.69 | Helped |
| add stochastic latency training to TET | TET · DVS-Gesture | accuracy: 88.60 -> 87.50; 98.43 -> 98.43 | Mixed |
| add stochastic latency training to tdBN | tdBN · DVS-Gesture, T=1 inference | accuracy: 64.06 -> 78.13 | Helped |
Surprising findings
- SLT-tdBN is below TET-trained baselines on DVS-CIFAR10 at T=10 (79.27 vs 81.56), so gains are not uniform across datasets.
Limitations
- On DVS-CIFAR10 SLT-tdBN lags TET at high latency and DVS-Gesture SLT-TET is worse than TET at T=1.
- Theoretical justification is only partial (heuristic weighting argument plus one lemma).
- TET results on DVS-CIFAR10 were re-computed by the authors while other baselines were copied from publications.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 DVS-CIFAR10 DVS-Gesture
Benchmarks: CIFAR-10 CIFAR-100 DVS-CIFAR10 DVS-Gesture
Metrics: test accuracy training time per epoch energy per inference (SOP-based)
Baselines: TET tdBN GLIF DSpike Diet-SNN A2S-QCFS
Builds on: TET tdBN surrogate gradient STBP
Closest papers by meaning
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.91)
- LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model (NeurIPS 2024 · similarity 0.90)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.90)
- Enhancing the Robustness of Spiking Neural Networks with Stochastic Gating Mechanisms (AAAI 2024 · similarity 0.90)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.90)
- Efficient Parallel Training Methods for Spiking Neural Networks with Constant Time Complexity (ICML 2025 · similarity 0.89)
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.