神经网络结构自适应研究综述

A Review on Architecture Adaptation of Neural Networks

网络结构自适应旨在根据特定学习任务和数据对神经网络结构进行自动设计和模型优化,以适应开放环境智能感知学习任务的综合需求.文中旨在全面综述网络结构自适应方法.首先,阐述并分析神经架构搜索的主要方法.然后,分别从轻量化神经架构搜索、智能感知任务、连续学习三个方面呈现网络结构自适应的研究进展.在此基础上,建立一套面向开放环境应用的深度神经网络组件与结构的自适应评价指标体系,提出一种网络结构自适应方法,通过注意力引导的微观架构自适应机制和渐进式离散策略,在优化过程中实现网络结构的自适应调整优化和逐步离散化,并与现有方法进行对比分析.最后,探讨当前方法存在的问题与挑战,展望未来的研究方向.

Network architecture adaptation aims to automatically design and optimize the neural network architectures based on specific learning tasks and data to meet the comprehensive needs of intelligent perception learning tasks in open environment.This paper is intended to provide a comprehensive review of network architecture adaptation methods.Firstly,the main methods of neural architecture search are elucidated and analyzed.Then,the research progress of network architecture adaptation is presented from three aspects:lightweight neural architecture search,intelligent perception tasks and continuous learning.On this basis,an adaptive evaluation index system of deep neural network components and architectures for open environment applications is established,and a network architecture adaptive method is proposed.Through the attention-guided micro-architecture adaptive mechanism and progressive discretization strategy,adaptive adjustment,optimization and gradual discretization of network structures are realized in the optimization process.The proposed method is compared with the existing methods.Finally,problems and challenges of current methods are discussed,and the future research directions are prospected.