Method and system for hybrid deployment of depth learning neural networks on terminals and clouds

Abstract

The invention provides a method and a system for hybrid deployment of a depth learning neural network on a terminal and a cloud end. The method adopts a distributed system structure, utilizes the cloud end and the terminal to deploy a neural network, and cooperates with each other to jointly complete reasoning operation tasks. the trained depth neural network compression model is deployed on the terminal equipment, and the depth neural network is deployed on the cloud end. The invention enhances sensor fusion, data privacy and system fault tolerance through distributed computing, allowing early exit points to be placed in terminal devices. The image classification can be completed on the local terminal and exited, and the fast local inference can be carried out. In order to improve the performance of the system, we can further use the deep depth of the cloud neural network to process.

Description

technical field [0001] The invention relates to software technology, in particular to a method for hybrid deployment of a deep learning neural network on a terminal and a cloud. Background technique [0002] Among many artificial intelligence technologies, deep neural network (DNN) is a machine learning technology that realizes artificial intelligence by simulating the neural network of the human brain. In view of its efficient data feature extraction and analysis capabilities, it has been widely used in computer vision, natural language processing, unmanned driving, smart home and other related fields or industries, affecting people's daily life. [0003] The essence of deep neural network is to simulate the human brain nerves and combine low-level features to form more abstract high-level features, so as to analyze the information expressed by the data. Building a DNN is mainly divided into two phases: the training phase and the speculation phase. In the training phase, ...

AI Technical Summary

Problems solved by technology

Either the terminal device collects data and transmits it to the large-scale deep neural network model on the cloud for processing, which increases the communication cost and brings delay and privacy issues; or the deep neural network model executed directly on the terminal must be greatly compressed and deleted. minus, thus losing the accuracy of the system

Images