Research map / Architectures and efficiency / Spiking networks
TS-SNN: Temporal Shift Module for Spiking Neural Networks
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 TS-SNN, which inserts a parameter-light temporal shift module (zero-padded forward/backward shifts of channel groups with a residual penalty factor alpha) into SNN architectures to fuse past, present and future spike features, reporting SOTA accuracy with fewer timesteps on CIFAR-10/100, ImageNet and CIFAR10-DVS.
Problem
SNNs need to exploit temporal information from spike timing while keeping energy consumption low, but existing temporal modeling methods either under-use the temporal dimension or add substantial computation.
Approach
Adapt the temporal shift operation from TSM to SNNs: shift channel groups along the timestep axis (left, right, unchanged) with zero padding, then combine shifted and original spike features via a residual sum Z' = alpha*Z + X, so each timestep sees past, present and future spike features at near-zero extra compute.
Key results
- TS-SNN with ResNet-19 reaches 96.72% on CIFAR-10 and 80.28% on CIFAR-100 at 2 timesteps.
- ResNet-34 TS-SNN reaches 70.61% top-1 on ImageNet at 4 timesteps, versus 67.69% for Real Spike ResNet34 at 4 timesteps.
- On CIFAR10-DVS, TS-SNN reaches 83.80% at 10 timesteps and 83.90% at 16 timesteps; ResNet-20 TS-SNN energy on ImageNet is 5.857 mJ versus 13.10 mJ reported for SEW ResNet18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TS to ResNet-19 | SNN on CIFAR-10 · CIFAR-10 | accuracy: 96.52% -> 96.72% | Helped |
| Add TS to ResNet-19 | SNN on CIFAR-100 · CIFAR-100 | accuracy: 80.10% -> 80.28% | Helped |
| Add TS to ResNet-34 | SNN on ImageNet · ImageNet | accuracy: 67.69% -> 70.61% | Helped |
| Add TS to ResNet-20 | SNN on CIFAR10-DVS · CIFAR10-DVS | accuracy: 80.90% -> 83.90% | Helped |
| Add TS to Spikeformer-4-384 (ablation) | Spikeformer-4-384 · CIFAR-10 | accuracy: 95.19% -> 95.28% | Helped |
| Add TS to Spikeformer-4-384 (ablation) | Spikeformer-4-384 · CIFAR-100 | accuracy: 77.86% -> 78.24% | Helped |
| Use random transformation instead of fixed-length transformation (ablation) | TS-SNN · CIFAR-100 | accuracy: 71.24\% -> 71.63\% | Helped |
| Use left-right-no-shift direction combination (ablation) | TS-SNN · CIFAR-100 | accuracy: 72.52\% | Helped |
| Apply TS consistently during training and inference rather than only during training (ablation) | TS-SNN · CIFAR10-DVS | accuracy: 81.20\% -> 83.60\% | Helped |
| Vary the channel folding factor C_k (ablation) | TS-SNN with ResNet20 · CIFAR-100 | accuracy: up to 0.96\% | Mixed |
Surprising findings
- At T=1 the temporal shift has no temporal effect in principle, yet the authors still observe accuracy gains attributed to the residual scaling factor alpha.
- Spike firing rate rises only modestly, from 33.06% (vanilla SNN) to 35.61% (TS-SNN).
Limitations
- Energy figures are estimated from operation counts at 45nm rather than measured on neuromorphic hardware.
- Ablations are run on a single configuration (ResNet20, CIFAR-100, T=2), and CIFAR-100 at T=1 lags MPBN.
- Reported results on CIFAR-10/100 are mean and std over only 3 trials.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Benchmarks: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Metrics: top-1 accuracy energy consumption (mJ) ACs MACs FLOPs firing rate
Baselines: TET LSG PFA MPBN IM-LIF RecDis-SNN Diet-SNN IM-loss GLIF SEW ResNet Spikeformer
Builds on: Temporal Shift Module (TSM) spatial-temporal backpropagation (STBP) surrogate gradient Spikeformer
Closest papers by meaning
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.92)
- 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)
- Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment (ICML 2025 · similarity 0.89)
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting (ICLR 2025 · similarity 0.89)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (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.