基于知识蒸馏的多阶段脉冲神经网络训练方法及装置
By mapping spiking neural networks and artificial neural networks to a unified logits space in stages, calculating the joint loss, and updating the parameters of spiking neural networks, the problems of low training efficiency and poor accuracy of spiking neural networks are solved, achieving efficient and low-cost training results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2024-09-12
- Publication Date
- 2026-07-17
AI Technical Summary
Existing spiking neural network training methods are inefficient in deep structures and complex tasks, and their accuracy is significantly lower than that of artificial neural networks, while also suffering from high power consumption.
A multi-stage spiking neural network training method based on knowledge distillation is adopted, which divides the spiking neural network and artificial neural network into multiple stages, maps them to a unified logits space through fully connected layers, calculates the joint loss based on confidence and features, and uses the backpropagation method to update the network parameters.
It improves the training efficiency and accuracy of spiking neural networks, narrows the accuracy gap with artificial neural networks, and reduces the power consumption during the training process.
Smart Images
Figure CN119129702B_ABST