Edge cloud collaborative deep learning model training method with classification precision maintenance and bandwidth protection

Abstract

The invention discloses an edge cloud collaborative deep learning model training method with classification precision maintenance and bandwidth protection, and solves the problems of network bandwidthconsumption and classification precision in edge cloud collaborative classification model training. A high-classification effectiveness image compression scheme created by using a compression auto-encoder is realized, and the coupling relationship between image compression and an image classification target component is increased. Through two stages of processing, in the first stage, a simple classifier model is used for back propagation gradient optimization of an auto-encoder, a loss target of lossy compression is controlled to a certain extent, in the second stage, a compression result isclouded, and powerful computing resources are used for training a high-precision classification model. Meanwhile, the edge cloud communication bandwidth consumption and the cloud classification precision of the system are balanced through the control compression ratio parameter of the compression auto-encoder on the edge, so that a set of perfect edge cloud collaborative training scheme is realized while the data classification characteristic is effectively ensured to be retained.

Description

Technical field [0001] The invention belongs to the field of artificial intelligence, and specifically relates to a side-cloud collaborative deep learning model training method with classification accuracy maintenance and bandwidth protection. Background technique [0002] The growing edge computing paradigm has become one of the most promising methods for applying deep learning to a wider range of practical fields, and it has shown great advantages in network bandwidth saving, latency reduction and privacy protection. Traditional centralized cloud computing is often used to train high-precision deep learning models, such as deep neural networks. However, the bandwidth pressure caused by data uploading to the cloud makes it difficult to aggregate global data on the cloud. Using the distributed edge computing paradigm, the edge server obtains raw data such as images and videos from nearby terminal nodes, and performs local learning without large-scale raw data upload to reduce ban...

AI Technical Summary

Problems solved by technology

The lossless compression method completely preserves the data information, but has limited ability to reduce the data size and cannot alleviate the bandwidth pressure brought by collaborative learning

Lossy compression methods can significantly reduce data size at the cost of losing data details, but have an impact on the accuracy of models trained on such compressed data

Images