Inherent Redundancy in Spiking Neural Networks ICCV2023

Spiking Neural Networks (SNNs) are well known as a promising energy-efficient alternative to conventional artificial neural networks. Subject to the preconceived impression that SNNs are sparse firing, the analysis and optimization of inherent redundancy in SNNs have been largely overlooked, thus the potential advantages of spike-based neuromorphic computing in accuracy and energy efficiency are interfered. In this work, we pose and focus on three key questions regarding the inherent redundancy in SNNs. We argue that the redundancy is induced by the spatio-temporal invariance of SNNs, which enhances the efficiency of parameter utilization but also invites lots of noise spikes. Further, we analyze the effect of spatio-temporal invariance on the spatio-temporal dynamics and spike firing of SNNs. Then, motivated by these analyses, we propose an Advance Spatial Attention (ASA) module to harness SNNs' redundancy, which can adaptively optimize their membrane potential distribution by a pair of individual spatial attention sub-modules. In this way, noise spike features are accurately regulated. Experimental results demonstrate that the proposed method can significantly drop the spike firing with better performance than state-of-the-art baselines.

Our module's details are in ./models/module/Attn.py.

train & test

requirements

python >= 3.7
torch >= 1.10
torchvision >= 0.11
spikingjelly == 0.0.0.0.12
h5py
pandas
einops
seaborn: only for plot features
torchinfo: only for see the details of networks

quick-start

sh final.sh

train on your own dataset

CUDA_VISIBLE_DEVICES=0 python3 train_test.py                   \
    -arch A                     \
    -T 60                       \
    -attention ASA              \
    -dataset yourdata           \
    -data_path DATA_PATH        \
    -batch_size_train 32        \
    -batch_size_test 4          \
    -clip 10                    \
    -num_epochs 200             \
    -seed 0                     \
    -ps 00011

args' helper

python3 train_test.py -h

arch: A for three-layer SNN; B for five-layer SNN
T: the time steps you want to set
attention: the module you want to use, "no" or "ASA" in this project
dataset: the dataset you want to use
data_path: the path of your dataset
batch_size_train / test: the batch-size used in training and inference
clip: the times you want to set for data augments algorithm "RCS"
num_epochs: the epochs you want to set
seed: the random seed you want to set
ps: the dropput is on / off on different layers

configs' details could be seen in ./config.py

performance

Dataset	Model	Acc.(%)	NASFR
DVS128 Gesture	LIF-SNN	91.3	0.176
	. + ASA	95.2(+3.9)	0.038(-78.4%)
DVS128 Gesture by spikingjelly	LIF-SNN-Five	95.5	0.023
	. + ASA	97.7(+2.2)	0.018(-21.7%)
DVS128 Gait-day	LIF-SNN	88.6	0.214
	. + ASA	93.6(+5.0)	0.045(-78.9%)
DVS128 Gait-night	LIF-SNN	96.4	0.197
	. + ASA	98.6(+2.2)	0.126(-36.0%)
DailyAction-DVS	LIF-SNN	92.5	0.017
	. + ASA	94.6(+2.1)	0.013(-23.5%)
HAR-DVS	SEW-Res-SNN-18	45.5	0.206
	. + ASA	47.1(+1.6)	0.183(-11.2%)

Citation

@inproceedings{yao2023inherent,
  title={Inherent Redundancy in Spiking Neural Networks},
  author={Yao, Man and Hu, Jiakui and Zhao, Guangshe and Wang, Yaoyuan and Zhang, Ziyang and Xu, Bo and Li, Guoqi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={16924--16934},
  year={2023}
}

Contact Information

For help or issues using this git, please submit a GitHub issue.

For other communications related to this git, please contact manyao@ia.ac.cn and jkhu29@stu.pku.edu.cn.

BICLab/ASA-SNN