Dimension Reduction Of High Dimensional Data Based On The Autoencoder

With the development of the technology during the past decades, especially advances in data collection and storage capabilities, there comes an information overload in most application fields. Researchers working in domains such as engineering, astronomy, biology, remote sensing, economics, and consumer transactions, face larger and larger observations and simulations on a daily basis. Such datasets are always high-dimensional, and the methods which used to process the smaller datasets are not useful for them. To deal with the problem, people must propose new algorithms, one of which is dimension reduction algorithm.The methods of dimension reduction developed fast in the past decades, of which are linear like PCA and of which are nonlinear like LLE, SNE, Autoencoder and so on. These methods play different role in different domains.This article focus on the methods of one model of an Autoencoder that it is one of an Artificial Neural Network. The Artificial Neural Network (ANN) is a multi-interdisciplinary methods that driven by people’s purpose of imitating the animals’ brain function. The architecture of ANN consists of the input layer, hidden layers, and the output layer. Once an ANN is trained by the certain data, it will have the ability of identifying the certain features of the data. So this ability of ANN can be used for face recognition and speech recognition.The Autoencoder can be used for the problem of dimension reduction if the units of the hidden layer are less than the input layer (or output layer), and the hidden layer gives a low-dimensional representation of the original data. In 2006, G. E. Hinton and R. R. Salakhutdinov present a model of ANN which is called the Unfolded Autoencoder. In 2012, Jing Wang et al. present an improved folded model which is based on the unfolded model. This paper presents an algorithm to improve the architecture of the Folded Autoencoder based on the previous studies.