An optimization method for deep learning of edge computing device

Abstract

The invention discloses an optimization method for deep learning of edge computing equipment, which comprises the following steps: obtaining the computing power of hardware of the edge computing equipment, determining a model quantization scheme according to the computing power of the hardware of the edge computing equipment, and reducing the model computing precision; carrying model light weightaccording to the model structure, reducing model parameter quantity and calculation quantity, and reducing memory access requirements of the model; performing model branch reduction and matrix decomposition operation according to the model structure and parameters; and deploying the deep neural network application to an edge computing device, carrying out deep reinforcement learning, and carryingout dynamic frequency modulation on a system layer through a DVFS frequency modulation strategy. Deep learning can be deployed on low-power-consumption edge computing equipment, basic availability isachieved, and meanwhile system energy efficiency is optimized.

Description

Technical field [0001] The present invention relates to the technical field of embedded edge computing, in particular to an optimization method for deep learning of edge computing equipment of a general embedded system. Background technique [0002] Nowadays, deep learning technology has become one of the current hot topics with its remarkable effects. Deep learning is a field of machine learning, which enables computers to train and learn through architectures such as convolutional neural networks (CNN). It mimics the way the human brain works by processing data and creating patterns for decision-making. The explosion of deep learning has brought new cognitive capabilities to computers, especially in computer vision perception. In some areas, the detection and recognition capabilities of computers have surpassed that of humans. The good effect of deep learning makes it possible to replace human work in many areas of daily life, especially in the field of computer vision. Large...

AI Technical Summary

Problems solved by technology

[0012] (1) The deep neural network model requires a huge amount of calculation for real-time calculation, but most embedded devices cannot provide such a huge amount of calculation;

[0013] (2) It is difficult and costly to develop using ASIC and FPGA dedicated hardware platforms;

[0014] (3) There are a large number of general-purpose computing devices in the market and in real life. If special hardware is used, the project volume and cost will be too large and impractical

Images