A Distributed Neural Network Training System Based on Edge Devices

Abstract

The present invention relates to the field of artificial intelligence technology, specifically a distributed neural network training system based on edge devices, which implements the function of effectively utilizing massive data generated at the edge end by means of distributed joint training through IoT edge devices; By optimizing the network configuration of the equipment for machine learning distributed training under the Parameter Server distributed architecture, the optimal configuration is achieved by comprehensively matching the performance and bandwidth of each equipment; through the model size of the training task, the performance and bandwidth of each equipment Optimize the allocation of time and determine the configuration that can maximize the performance of the equipment group at present. The method provided by the present invention not only realizes the deployment of neural network training to the IoT field, but also optimizes the configuration of distributed networking tasks, effectively avoids unnecessary performance loss between clusters, and improves the computing power of the distributed training system. performance.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a distributed neural network training system based on edge devices. Background technique [0002] The Internet of Things (IoT) and Cyber ​​Physical Systems (CPSs) rely on a large number of devices from the cloud to the edge to transmit a large amount of collected physical information to the Internet. In the future, the number of electronic devices will increase, and the amount of data generated by the devices will show an explosive growth. May cause the cloud computing server to be overwhelmed. The performance gap between these exploding data and cloud computing servers is an urgent problem that needs to be solved. At present, with the rapid growth of the performance of electronic devices at the edge of the Internet of Things, the concept of edge computing has gradually become the key to solving such problems. It can realize localized processing and rapid respon...

AI Technical Summary

Problems solved by technology

Under the limitation of existing equipment, an excessively high PS / worker ratio will lead to excess bandwidth and insufficient cluster computing performance, and an excessively low PS / worker ratio will cause large-scale network congestion in the cluster and cause distributed training problems bandwidth bottleneck

Images