卷积神经网络研究综述

发布时间：2018-03-02 02:03

本文关键词： 卷积神经网络深度学习网络结构训练方法领域数据　出处：《计算机学报》2017年06期 　论文类型：期刊论文

【摘要】：作为一个十余年来快速发展的崭新领域,深度学习受到了越来越多研究者的关注,它在特征提取和建模上都有着相较于浅层模型显然的优势.深度学习善于从原始输入数据中挖掘越来越抽象的特征表示,而这些表示具有良好的泛化能力.它克服了过去人工智能中被认为难以解决的一些问题.且随着训练数据集数量的显著增长以及芯片处理能力的剧增,它在目标检测和计算机视觉、自然语言处理、语音识别和语义分析等领域成效卓然,因此也促进了人工智能的发展.深度学习是包含多级非线性变换的层级机器学习方法,深层神经网络是目前的主要形式,其神经元间的连接模式受启发于动物视觉皮层组织,而卷积神经网络则是其中一种经典而广泛应用的结构.卷积神经网络的局部连接、权值共享及池化操作等特性使之可以有效地降低网络的复杂度,减少训练参数的数目,使模型对平移、扭曲、缩放具有一定程度的不变性,并具有强鲁棒性和容错能力,且也易于训练和优化.基于这些优越的特性,它在各种信号和信息处理任务中的性能优于标准的全连接神经网络.该文首先概述了卷积神经网络的发展历史,然后分别描述了神经元模型、多层感知器的结构.接着,详细分析了卷积神经网络的结构,包括卷积层、池化层、全连接层,它们发挥着不同的作用.然后,讨论了网中网模型、空间变换网络等改进的卷积神经网络.同时,还分别介绍了卷积神经网络的监督学习、无监督学习训练方法以及一些常用的开源工具.此外,该文以图像分类、人脸识别、音频检索、心电图分类及目标检测等为例,对卷积神经网络的应用作了归纳.卷积神经网络与递归神经网络的集成是一个途径.为了给读者以尽可能多的借鉴,该文还设计并试验了不同参数及不同深度的卷积神经网络来分析各参数间的相互关系及不同参数设置对结果的影响.最后,给出了卷积神经网络及其应用中待解决的若干问题.
[Abstract]:As a new field of rapid development in the past ten years, in-depth learning has attracted more and more researchers' attention. It has obvious advantages over shallow model in feature extraction and modeling. Depth learning is good at mining more and more abstract feature representations from raw input data. These representations have a good generalization ability. They overcome some of the problems that used to be difficult to solve in artificial intelligence. And with the significant increase in the number of training data sets and the rapid increase in chip processing power, It has been effective in target detection and computer vision, natural language processing, speech recognition and semantic analysis, thus promoting the development of artificial intelligence. Deep neural network is the main form at present. The connection pattern between neurons is inspired by animal visual cortical tissue, and convolutional neural network is one of the classical and widely used structures. The characteristics of weight sharing and pool operation can effectively reduce the complexity of the network, reduce the number of training parameters, make the model invariant to a certain extent for translation, distortion and scaling, and have strong robustness and fault-tolerant ability. It is also easy to train and optimize. Based on these advantages, it performs better than standard fully connected neural networks in various signal and information processing tasks. In this paper, the development history of convolutional neural networks is first summarized. Then we describe the neuron model and the structure of multilayer perceptron respectively. Then, we analyze the structure of convolutional neural network in detail, including convolution layer, pool layer, full junction layer, which play different roles. In this paper, the improved convolution neural networks such as network model and spatial transformation network are discussed. At the same time, the supervised learning, unsupervised learning training methods and some common open source tools of convolutional neural networks are also introduced. This paper takes image classification, face recognition, audio retrieval, ECG classification and target detection as examples. The application of convolution neural network is summarized. The integration of convolutional neural network and recurrent neural network is a way. In this paper, convolution neural networks with different parameters and different depths are designed and tested to analyze the relationship between the parameters and the effects of different parameter settings on the results. Finally, some problems to be solved in the convolution neural network and its application are given.
【作者单位】：中国科学院苏州纳米技术与纳米仿生研究所;中国科学院大学;
【分类号】：TP183

【参考文献】