摘要
卷积神经网络作为深度学习的一种经典而广泛应用的结构,克服了过去人工智能中被认为难以解决的一些问题。卷积神经网络的局部连接、权值共享及下采样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,也易于训练和优化。文章介绍了卷积神经网络的训练方法,开源工具,及在图像分类领域中的一些应用,给出了卷积神经待解决的问题及展望。
As a classical kind of widely used network structure in deep learning,convolutional neural networks has successfullysolved some problems which were considered difficult to solve in artificial intelligence in the past. The characteristics such as localconnections, shared weights,under-sampling etc. can reduce the complexity of networks and the number of training parameters, andcan also make the model creating invariance to translation, distortion and zoom and having strong robustness and fault tolerance.So it is easy to train and optimize its network structure. This paper introduces the training methods, open source tools ofconvolutional neural networks and its applications in the field of image classification, the problems and prospects of theconvolutional neural network to be solved are given also.
作者
刘健
袁谦
吴广
喻晓
Liu Jian;Yuan Qian;Wu Guang;Yu Xiao(Zhejiang Provincial Testing Institute of Electronic Information Products,Hangzhou,Zhejiang 310007,China)
出处
《计算机时代》
2018年第11期19-23,共5页
Computer Era
基金
浙江省科技计划项目"移动应用软件自动化检测平台"(2017F10031)
关键词
深度学习
卷积神经网络
网络结构
训练方法
deep learning
convolutional neural networks
network structure
training method
