Research map / Architectures and efficiency / Spiking networks
Shrinking Your TimeStep: Towards Low-Latency Neuromorphic Object Recognition with 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 Shrinking SNN (SSNN), which splits an SNN into stages with progressively fewer timesteps linked by a learnable temporal transformer, and adds early classifiers during training to achieve low-latency neuromorphic recognition without accuracy loss.
Problem
Existing SNNs need 10 to 40+ timesteps to recognize neuromorphic objects, and their accuracy degrades sharply at low latency. Surrogate-gradient training also suffers from gradient mismatch and vanishing/exploding gradients with binary spikes.
Approach
Divide the SNN into stages with progressively shrinking timesteps (T1>T2>...>Tn) and transform information between stages with a learnable temporal transformer that redistributes the total activity across the new timesteps; add cross-entropy early classifiers after each non-final stage during training only, so there is no inference overhead.
Key results
- At an average timestep of 5 without data augmentation, SSNN (VGG-9) reaches 73.63% on CIFAR10-DVS, 77.97% on N-Caltech101, and 90.74% on DVS-Gesture.
- SSNN improves over the baseline by 6.55% to 21.41% in accuracy across the three datasets and both architectures in the ablation.
- At average timestep 5, SSNN on CIFAR10-DVS beats Spikformer (68.55%) by about 5.08 points, and at timestep 8 it reaches 78.57%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add timestep shrinkage (ablation) | VGG-9 SNN · CIFAR10-DVS, N-Caltech101, DVS-Gesture, average timestep 5 | accuracy: 67.08 -> 71.53; 66.45 -> 75.42; 80.29 -> 85.42 | Helped |
| Add early classifiers (ablation) | VGG-9 SNN · CIFAR10-DVS, N-Caltech101, DVS-Gesture, average timestep 5 | accuracy: 67.08 -> 70.10; 66.45 -> 70.13; 80.29 -> 84.49 | Helped |
| Add timestep shrinkage and early classifiers (ablation) | VGG-9 SNN · CIFAR10-DVS, N-Caltech101, DVS-Gesture, average timestep 5 | accuracy: 67.08 -> 73.63; 66.45 -> 77.97; 80.29 -> 90.74 | Helped |
| Add timestep shrinkage (ablation) | ResNet-18 SNN · CIFAR10-DVS, N-Caltech101, DVS-Gesture, average timestep 4.94 | accuracy: 58.60 -> 61.97; 52.81 -> 62.62; 65.17 -> 72.68 | Helped |
| Add early classifiers (ablation) | ResNet-18 SNN · CIFAR10-DVS, N-Caltech101, DVS-Gesture, average timestep 4.94 | accuracy: 58.60 -> 64.30; 52.81 -> 57.29; 65.17 -> 77.43 | Helped |
| Add timestep shrinkage and early classifiers (ablation) | ResNet-18 SNN · CIFAR10-DVS, N-Caltech101, DVS-Gesture, average timestep 4.94 | accuracy: 58.60 -> 68.10; 52.81 -> 67.51; 65.17 -> 86.58 | Helped |
| Remove the temporal transformer and truncate to the first shrunken timesteps (ablation) | SSNN with VGG-9 · CIFAR10-DVS, average timestep 5 | accuracy | Hurt |
| Use only the final-stage loss (set auxiliary loss coefficients to zero) (ablation) | SSNN with VGG-9 · CIFAR10-DVS, timestep 5 | accuracy: 67.08 -> 68.70%–70.10% | Hurt |
| Vary nonzero early-classifier loss coefficients within the tested ranges (ablation) | SSNN with VGG-9 · CIFAR10-DVS, timestep 5 | accuracy: 68.70%–70.10%; 68.80%–70.36% | No clear effect |
| Vary stage division and stage-wise timesteps while keeping average timestep at 5 (ablation) | SSNN with VGG-9 · DVS-Gesture | accuracy: 84.38%–88.43% | No clear effect |
| Increase SSNN average timestep from 5 to 8 | SSNN with VGG-9 · CIFAR10-DVS | accuracy: 73.63 -> 78.57 | Helped |
| Increase SSNN average timestep from 5 to 8 | SSNN with VGG-9 · N-Caltech101 | accuracy: 77.97 -> 79.25 | Helped |
Surprising findings
- Using the temporal transformer is essential: shrinking timesteps without it lowers accuracy below the baseline, even for SSNN.
- A two-stage setting with timesteps 9 and 1 still reached 86.34% on DVS-Gesture, showing shrinkage works with very large temporal scale changes.
Limitations
- Baselines were self-reimplemented under a common low-latency setting for most methods, so comparisons depend on those reproductions.
- Experiments are limited to small networks (VGG-9, ResNet-18) and three neuromorphic datasets at 42x42 or 32x32 resolution.
Experimental setup
Datasets: CIFAR10-DVS N-Caltech101 DVS-Gesture
Benchmarks: CIFAR10-DVS N-Caltech101 DVS-Gesture
Metrics: test accuracy (%) average timestep
Baselines: STBP-tdBN PLIF BackEISNN MLF TEBN Spikformer TET SLTT GLIF DSR
Builds on: LIF neuron surrogate gradient BPTT TET MLF
Closest papers by meaning
- Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips (ICLR 2024 · similarity 0.88)
- Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.88)
- TS-SNN: Temporal Shift Module for Spiking Neural Networks (ICML 2025 · similarity 0.87)
- Rethinking Spiking Neural Networks from an Ensemble Learning Perspective (ICLR 2025 · similarity 0.87)
- Self-cross Feature based Spiking Neural Networks for Efficient Few-shot Learning (ICML 2025 · similarity 0.86)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.86)
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.