Research map / Architectures and efficiency / Spiking networks
Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness
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 Mixed Time-step Training (MTT), which partitions an SNN into stages assigned random time steps per iteration with temporal transformation modules between them, yielding SNNs that generalize across time steps and deploy on fully event-driven chips.
Problem
Directly trained SNNs are tied to the specific time step used in training ("temporal inflexibility"), which hinders deployment on time-step-free fully event-driven neuromorphic chips and dynamic energy-performance time-step adjustment.
Approach
Randomly sample different time steps for different network stages in each iteration, connect stages with temporal transformation modules (pooling-like summing/replicating of frames), and sum losses over several sampled temporal configurations so one set of weights works across many time steps and on event-driven hardware.
Key results
- MTT-trained ResNet-19 reaches 96.84% on CIFAR10 (T=6) and 81.98% on CIFAR100 (T=6), exceeding the compared direct-training methods at the same time steps.
- On N-MNIST, the MTT model deployed on Speck2e achieves 98.57% versus 92.77% for SDT; CIFAR10-DVS simulator accuracy is 58.5% for MTT vs 48.4% for SDT.
- The spike difference between the event-driven simulator and the on-chip output is 3.92% on N-MNIST, comparable to the 2.81% between two identical on-chip runs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with MTT instead of SDT | SNN on N-MNIST · N-MNIST, Speck2e | accuracy: 92.77 -> 98.57 | Helped |
| Train with MTT instead of SDT | VGGSNN · CIFAR10-DVS, simulator | accuracy: 48.4 -> 58.5 | Helped |
| Train with MTT instead of SDT | SNN · DVS-Gesture, simulator | accuracy: 80.68 -> 81.82 | Helped |
| Train ResNet-19 with MTT instead of InfLoR-SNN | ResNet-19 · CIFAR10, T=6 | accuracy: 96.49±0.08 -> 96.84±0.03 | Helped |
| Train ResNet-19 with MTT instead of InfLoR-SNN | ResNet-19 · CIFAR100, T=6 | accuracy: 79.51±0.11 -> 81.98±0.03 | Helped |
| Increase partitioning granularity from g=8 toward g=1 (ablation) | MTT ResNet-18 · CIFAR100, T=2–6 | accuracy | Helped |
| Add BN calibration using 10 batches (ablation) | MTT ResNet-18 · CIFAR100 | accuracy | Helped |
| Train with MTT instead of SDT | SNN · Multiple inference time steps | accuracy | Helped |
| Train with MTT instead of SDT | SNN weights · CIFAR100, Gaussian weight-noise robustness | accuracy | Helped |
| Use T=Tmax BN statistics for other time steps instead of recalibrating each T (ablation) | MTT ResNet-19 · CIFAR100, other uniform time steps | accuracy | No clear effect |
| Use MTT instead of TET for dynamic inference | ResNet-19 with SEENN-I · CIFAR10 | accuracy | Helped |
| Train with MTT instead of SDT | SNN · Fully event-driven deployment | spike difference: 28.77% -> 3.92% | Helped |
Surprising findings
- Fully event-driven deployment of SDT-trained models loses substantial accuracy (N-MNIST 92.77% on Speck vs 98.09% in PyTorch), while MTT models stay nearly lossless.
Limitations
- Training cost is about 1.5x that of standard direct training because of multiple sampled temporal configurations per iteration.
- Event-driven experiments use bias-free networks and a simulator for larger models, since Speck chip capacity is limited.
- Standard BN layers degrade under MTT, requiring BN calibration.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet CIFAR10-DVS N-Caltech101 N-MNIST DVS-Gesture seqMNIST Spiking Heidelberg Digits Spiking Speech Commands
Benchmarks: CIFAR10 CIFAR100 ImageNet CIFAR10-DVS N-Caltech101 N-MNIST DVS-Gesture
Metrics: top-1 accuracy spike difference noise-injection robustness
Baselines: SDT TET InfLoR-SNN GLIF Dspike STBP-tdBN SEW ResNet TCJA-SNN NDA SALT
Builds on: spatiotemporal backpropagation tdBN TET SEENN
Closest papers by meaning
- TS-SNN: Temporal Shift Module for Spiking Neural Networks (ICML 2025 · similarity 0.92)
- High-Performance Temporal Reversible Spiking Neural Networks with $\mathcal{O}(L)$ Training Memory and $\mathcal{O}(1)$ Inference Cost (ICML 2024 · similarity 0.91)
- Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment (ICML 2025 · similarity 0.91)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.91)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.91)
- Enhancing Training of Spiking Neural Network with Stochastic Latency (AAAI 2024 · 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.