Research map / Architectures and efficiency / Spiking networks
Dynamic Token Masking in Spiking Neural Network
Yuetong Fang, Ziqing Wang, Deming Zhou, Hongwei Ren, Shibo Zhou, Renjing Xu
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 depth-wise convolution (spatial locality enhancement) at T=1 (ablation) | DynMask-T · ImageNet-1K | top-1 accuracy (%): 66.88 -> 70.47 | Improved |
| add depth-wise convolution at T=64 (ablation) | DynMask-T · ImageNet-1K | top-1 accuracy (%): 77.81 -> 80.23 | Improved |
| add depth-wise convolution (energy cost vs accuracy, T=8) (ablation) | DynMask-T · ImageNet-1K | accuracy (%) and energy (mJ): 76.82 -> 79.39 accuracy; 7.52 -> 8.55 mJ | Improved |
| replace quadratic spiking self-attention with DynMask linear token mixing (energy) | MST · ImageNet-1K | energy (mJ): 380.86 -> 8.55 | Improved |
| replace vanilla softmax self-attention (MST T=128) with DynMask at T=8 | Swin-T (BN) converted SNN · ImageNet-1K | top-1 accuracy (%): 77.88 -> 79.39 | Improved |
| use DynMask under similar membrane-potential-init-free setting (T=48) vs MST with init (T=128) | DynMask-T · ImageNet-1K | accuracy (%) and energy (mJ): 77.79 -> 78.00 accuracy; 99.3 -> 50.61 mJ | Improved |
| increase simulation timesteps from 1 to 64 | DynMask-T · ImageNet-1K | top-1 accuracy (%): 70.47 -> 80.23 | Improved |
| scale model from DynMask-T to DynMask-B at T=64 | DynMask · ImageNet-1K | top-1 accuracy (%): 80.23 -> 82.00 | Improved |
| replace MST (T=256) with DynMask at T=1 | Swin-T (BN) converted SNN · CIFAR-10 | accuracy (%): 97.27 -> 97.75 | Improved |
| increase timesteps from 4 to 16 | DynMask-T · COCO 2017 detection | mAP@0.5: 52.6 -> 60.2; 241.4 -> 405.0 mJ | Improved |
| increase timesteps from 4 to 16 | DynMask-T · ADE20K segmentation | MIoU (%): 32.5 -> 38.6 | Improved |
| random spike masking of 75% of inputs in spiking self-attention (ablation) | Spiking Transformer · CIFAR-10 | accuracy (%): 97.27 -> 97.29 | No clear effect |
Closest papers by meaning
- Neural Dynamics Self-Attention for Spiking Transformers (ICLR 2026 · similarity 0.89)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.88)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.88)
- SpikingResformer: Bridging ResNet and Vision Transformer in Spiking Neural Networks (CVPR 2024 · similarity 0.88)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.87)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.87)
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.
