Graph convolutional network training method and device for asynchronous event stream, and storage medium

By sampling and converting the asynchronous event stream into graphs of different scales, and performing progressive training in graph convolution network, the problems of large amount of computation, slow learning speed and overfitting in asynchronous event stream processing are solved. , achieving higher model accuracy and performance.

CN119494368APending Publication Date: 2025-02-21TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311028162.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-08-15
Publication Date
2025-02-21

AI Technical Summary

Technical Problem

When using graph convolution networks to process asynchronous event streams output by neuromorphic vision sensors, the existing technology is large in computing, slow learning speed, time-consuming training, and easy to cause overfitting problems. The downsampling method destroys the high resolution of the data. simplifying the network structure will reduce the model accuracy.

Method used

A progressive graph convolution network training method for asynchronous event streams is proposed. The asynchronous event stream is obtained by sampling multiple event streams in segments and intervals, and converted them into graphs of different scales, and gradually input them into graph convolution networks. In this paper, we perform progressive training, extract features and fully connected layer output, and finally retrain the graph convolutional network to obtain the trained target model.

Benefits of technology

While reducing the amount of calculation, this method effectively overcomes the overfitting problem, improves the accuracy and performance of the model, extends the learning time of the model, and enables it to learn more complex information for a longer time.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119494368A_ABST
    Figure CN119494368A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of computer visual recognition, in particular to a graph convolutional network training method and device for asynchronous event streams and a storage medium. The method comprises the following steps: carrying out segmented interval sampling on a plurality of event streams to obtain asynchronous event streams, wherein the event streams are data output by sensing light intensity change through an event camera; converting the asynchronous event flow into graphs of different scales; and inputting the graphs of different scales into a graph convolutional network for progressive training to obtain a trained target graph convolutional network. Compared with a conventional training method, the method provided by the embodiment of the invention has the advantages that the node scale of the graph is updated and enriched from simple to complex along with the iterative increase of the model training period, the overfitting problem is effectively overcome while the calculation amount is reduced, and the model precision is improved.
Need to check novelty before this filing date? Find Prior Art