This is the official code repository for the project: SAFA-SNN: Sparsity-Aware On-Device Few-Shot Class-Incremental Learning with Fast-Adaptive Structure of Spiking Neural Network (ICLR 2026) [paper]
The overall framework of our proposed method SAFA-SNN.
SAFA-SNN framework includes three main components: (a) Training abundant data in the base session and selecting active and stable neurons by masks. In incremental learning sessions, activate neurons will substantially adjust their firing thresholds based on sparsity to incorporate new classes, while stable neurons maintain their prior firing behavior through only slight threshold variations. (b) Top: forward propagation through FSCIL process; Bottom: backpropagation using zeroth-order optimization only in the base class training. (c) Freezing backbones and updating the prototypes by orthogonal subspace projection in the incremental learning.
python==3.10.9
torch==2.5.1
tqdm==4.64.1
numpy==1.23.5
cuda==12.2
torchvision==0.20.1
spikingjelly==0.0.0.0.14
You can download the datasets used in this project from the following links:
python main.py --config jsons/safa.json
- The results on static datasets:
- The results on neuromorphic datasets:
If you have any questions about our work or this repository, please submit a GitHub issue. For further discussions, feel free to contact us by email huijingzhang@zju.edu.cn.
We would like to thank the developers of the following open-source projects for their invaluable contributions:
If you find this repository useful in your research, please cite our paper:
@inproceedings{
zhang2026safa,
title={SAFA-SNN: Sparsity-Aware On-Device Few-Shot Class-Incremental Learning with Fast-Adaptive Structure of Spiking Neural Network},
author={Zhang, Huijing and Cao, Muyang and Jiang, Linshan and Du, Xin and Yu, Di and Lv, Changze and Deng, Shuiguang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/pdf?id=9jcB40wjk3}
}