Research and Development of Deep Learning
Download as PDF
DOI: 10.25236/fetms.2017.153
Corresponding Author
Ruihua Zhao
Abstract
In view of the shortcomings of shallow learning in the past, such as lack of feature expression and excessive dimensionality, the depth learning solves these problems well by the unique hierarchical structure and the ability to extract high-level features from low-level features, and brought new hope for artificial intelligence. The development of deep learning during different periods was introduced. The basic models of restricted boltzmann machines (RBM),auto encoder (AE) and convolutional neural networks (CNN) were analyzed to present the deep hierarchical structures of deep belief networks (DBN), deep boltzmann machine(DBM) and stacked auto encoders (SAE).The applications of deep learning in the fields of speech recognition,computer vision,natural language processing and information retrieval in recent years were introduced to illustrate the superiority and flexibility of deep learning compared with other shallow learning algorithms. Some future research directions were predicted based on the analysis,and some conclusions were made according to the improvement of deep learning on algorithm generalization, adaptation of big data and modifying on deep structure.
Keywords
Shallow Learning, Deep Learning, Hierarchical Structure, Artificial Intelligence, Machine Learning.