Efficient spiking models for neuromorphic touch面向边缘设备的脉冲神经网络压缩
Compressing spiking networks for low-resource deployment.这项研究压缩训练好的脉冲神经网络，使其能部署在资源有限的设备上。
Develops one-shot post-training pruning and quantisation methods that preserve useful spike dynamics while reducing model cost.我参与训练后单步剪枝与量化研究：模型训练完成后直接压缩，无需重新训练。
- Role项目角色
- Researcher研究成员
- Period项目时间
- 2023–present
01 / problem问题
Deployment tension部署限制
Spiking models promise efficient temporal processing, but large networks still need practical compression before low-resource deployment.脉冲模型可以高效处理时序信号，但大型网络在资源有限的设备上仍需要压缩。
02 / contribution工作
My contribution我负责的工作
I contributed to the research and evaluation of one-shot post-training pruning and quantisation for spiking neural networks.我参与研究并评估训练后单步剪枝与量化方法，重点验证其在不同脉冲网络上的表现。
03 / approach方法
Compression path压缩方法
The method compresses a trained model in one post-training step and evaluates pruning and quantisation across event and static datasets, including ImageNet-scale models.训练后单步剪枝与量化直接作用于已经训练好的模型，无需重新训练。实验覆盖事件数据、静态图像和 ImageNet 规模的模型。
04 / outcome结果
Research outcome验证规模
The study reports state-of-the-art one-shot results across architectures up to SEW-ResNet152 and spike-driven Transformers. Preliminary work was accepted at the NeurIPS 2025 OPT-ML workshop.研究在 SEW-ResNet152 和脉冲驱动 Transformer 等架构上验证了方法，并报告了当前最优的训练后单步压缩结果。前期工作已被 NeurIPS 2025 OPT-ML 研讨会接收。
Evidence论文与材料