Analysis on Lightweight Network Methods and Technologies

Abstract

There are currently two main schools of deep learning. One is academic. They pursue stronger performance through powerful, complex models. The other is the engineering school. Their purpose is to efficiently deploy models to various hardware platforms. Complex models have better performance. However, it also brings unavoidable consumption. With the increasing depth of convolutional neural networks, lightweighting has become a key research direction. There are currently four main methods for designing lightweight networks. This article will first introduce CNN model compression and basic convolution operations.This paper also introduces the model compression based on AutoML and the automatic animation design based on NAS. Finally, according to the above three points, this paper introduces the application of the above methods in artificially designed neural networks.This paper mainly introduces the step-by-step evolution of the existing methods. This paper analyzes aspects of current neural network improvements and emerging problems. The significance of this paper is to summarize and deepen the solved problems and key problems in the lightweight process through past experience.