Reducing Dimensionality of Data Using Autoencoders

Data dimensionality reduction is a long-standing pre-processing step in machine learning algorithms. Many of the efficient methods are devised to reduce the dimensionality of data very effectively; however, they are incapable of recovering the original data. But the autoencoders are effective not only in reducing the dimensionality but also reconstructing the original data. In this paper, we are attempting to explore the dimensionality reduction capability of autoencoders, and try to comprehend the difference between autoencoder and PCA dimensionality reduction methods. Experiments are conducted on both the methods using MNIST datasets. The results show that the autoencoder can indeed learn somewhat dissimilar from PCA method.