Related Experiment Video
Updated: Jul 18, 2025
10:45
Time-dependent Increase in the Network Response to the Stimulation of Neuronal Cell Cultures on Micro-electrode Arrays
Published on: May 29, 2017
9.9K
IDSNN: Towards High-Performance and Low-Latency SNN Training via Initialization and Distillation
Xiongfei Fan1, Hong Zhang1, Yu Zhang1,2
1State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, China.
Biomimetics (Basel, Switzerland)
|August 25, 2023
Summary
This study introduces IDSNN, a new method for training spiking neural networks (SNNs) using artificial neural networks (ANNs). IDSNN significantly improves SNN accuracy and training speed, making them more practical for real-world applications.
Area of Science:
- Artificial Intelligence
- Computational Neuroscience
- Machine Learning
Background:
- Spiking neural networks (SNNs) offer biomimetic and efficient computation using spikes.
- Current SNN training methods lead to low accuracy and high inference latency.
- Existing approaches often involve direct training or conversion from artificial neural networks (ANNs).
Purpose of the Study:
- To address the limitations of low accuracy and high latency in SNNs.
- To propose a novel training pipeline, IDSNN, for enhanced SNN performance.
- To leverage ANNs as a source for parameter initialization and knowledge distillation.
Main Methods:
- Developed IDSNN, a training pipeline utilizing parameter initialization and knowledge distillation.
- Employed ANNs as both parameter sources and knowledge teachers.
- Evaluated IDSNN on CIFAR10 and CIFAR100 datasets.
Main Results:
- Achieved competitive top-1 accuracy: 94.22% on CIFAR10 and 75.41% on CIFAR100.
- Demonstrated low inference latency.
- Showcased a 14× faster convergence speed compared to direct SNN training.
Conclusions:
- IDSNN effectively overcomes accuracy and latency issues in SNNs.
- The proposed method maximizes knowledge transfer from ANNs.
- IDSNN exhibits significant practical value for SNN applications due to improved efficiency and faster convergence.
