Abstract: Deep learning is a broader class of machine learning method based on data representation. Its emergence has not only promoted the development of machine learning, but also accelerated the innovation of artificial intelligence. This paper discusses and compares several typical models of deep learning. It first investigates restricted Boltzmann machine, deep belief network and auto-encoder, and explores their structure, principle, advantages and disadvantages of these unsupervised learning models in detail. Secondly, it discusses several supervised learning models including convolutional neural network, recurrent neural network and deep stacked network. And it also evaluates and analyzes the model structure and working principle. Then it makes a contrastive analysis of typical deep learning models and performs comparative experiments. Deep belief network and convolutional neural network are applied to the handwriting digits recognition task and experimental results show that deep learning models have better recognition performance than traditional neural network. Finally, it discusses the developments and challenges of deep learning in the future.

Key words: deep learning, convolutional neural network, deep belief network, auto-encoder, recurrent neural network, deep stacked network

摘要： 深度学习是基于数据表示的一类更广的机器学习方法，它的出现不仅推动了机器学习的发展，而且促进了人工智能的革新。对深度学习的几种典型模型进行研究与对比。首先介绍受限玻尔兹曼机、深度置信网络、自编码器等无监督学习模型，对其结构、原理和优缺点进行了详细探讨。讨论卷积神经网络、循环神经网络和深度堆叠网络等监督学习模型，分别从模型架构和工作原理来评价与分析。对深度学习的典型模型进行对比分析，将深度置信网络和卷积神经网络应用在手写体数字识别任务中，结果证实深度学习比传统的神经网络具有更好的识别性能。最后探讨深度学习未来的发展与挑战。

关键词: 深度学习, 卷积神经网络, 深度置信网络, 自编码器, 循环神经网络, 深度堆叠网络