Optimization and Implementation of a Lightweight Neural Network Architectures for Edge Computing

Abstract

For the resource constraints and real-time requirements of edge computing scenarios, this paper systematically studies the optimization and implementation method of a lightweight neural network architecture. From the constraints of edge device computing power, power consumption constraints and the diversity of deployment environment, this paper discusses the architecture design strategy based on convolution decomposition, efficient embedding of attention mechanism and width depth collaborative optimization, and then proposes a comprehensive optimization scheme of training perception quantization, pruning and neural architecture search combined optimization, reasoning engine collaborative adaptation and end-to-end dynamic optimization. Through the deep integration of algorithm innovation and hardware characteristics, a complete technical framework from model design to edge deployment is constructed, aiming to achieve a good balance between accuracy and efficiency under resource constraints, and provide theoretical support and practical guidance for the model selection and deployment optimization of edge intelligent system.