Research map / Architectures and efficiency / Spiking networks
SMixer: Rethinking Efficient-Training and Event-Driven SNNs
Yijie Lu, Xinhao Luo, Yixing Zhang, Zhiyan Wang, Wentao Li, YanhanWang, Zhi Liu, Zhaokun Zhou, Guoqi Li
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 DSTSP spatial-temporal pruning | STMixer-8-768 · ImageNet-1K | accuracy / energy: acc 76.68 -> 76.87; energy 4.45 -> 2.36 mJ | Improved |
| add DSTSP to Spikformer-V2 with Spiking-token Mixer (T=4) | SpikformerV2->M · ImageNet-1K | top-1 accuracy: 79.12 -> 78.03 | Hurt |
| add DSTSP to SDT-V3 with Spiking-token Mixer | SDT-V3->M · ImageNet-1K | top-1 accuracy: 79.25 -> 77.75 | Hurt |
| replace spiking self-attention with Spiking-token Mixer | Spiking Transformers · ImageNet-1K | top-1 accuracy: 79.05 -> 79.16 (Spikformer V2-8-512) | No clear effect |
| add DSTSP (spatial 0.3, one time step) | SMixer · CIFAR10/CIFAR100 | top-1 accuracy: 96.01 -> 95.67; 81.87 -> 81.03 | Hurt |
| add DSTSP | SMixer · CIFAR10-DVS and DVS128 | top-1 accuracy: 83.02 -> 82.34; 98.61 -> 98.26 | Hurt |
| add DSTSP | SMixer (SpikeYOLO) · COCO val2017 | mAP@50 and power: mAP 58.9 -> 57.4; power 36.8 -> 21.2 mJ | Mixed |
| add DSTSP | SMixer · time-series forecasting (Metr-la, Pems-bay, Solar, Electricity) | R2 / RSE | No clear effect |
| prune low-SIV tokens instead of random or high-SIV tokens (ablation) | DSTSP · CIFAR100 | top-1 accuracy: 81.03 vs 78.11 (high-SIV pruning) | Improved |
| apply DSTSP in both training and inference instead of inference only (ablation) | DSTSP · CIFAR100 | top-1 accuracy: 79.74 -> 81.03 | Improved |
| apply DSTSP in training only instead of training and inference (ablation) | DSTSP · CIFAR100 | top-1 accuracy: 81.03 -> 80.95 | No clear effect |
| increase spatial pruning ratio from 0.4 to 0.6 (ablation) | DSTSP · CIFAR100 | top-1 accuracy: 80.81 -> 79.13 | Hurt |
Closest papers by meaning
- Spiking Token Mixer: An event-driven friendly Former structure for spiking neural networks (NeurIPS 2024 · similarity 0.92)
- QP-SNN: Quantized and Pruned Spiking Neural Networks (ICLR 2025 · similarity 0.89)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.89)
- Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips (ICLR 2024 · similarity 0.89)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.89)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.88)
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.
