A STUDY ON STACKED AUTOENCODERS AND ITS FINE-TUNING

Abstract

A stacked autoencoder is a kind of unsupervised deep learning algorithm which looks like the automatically learning features of input data, such as edges and objects in images. Stacked autoencoders have been used as building blocks to build and initialize multi-layer neural networks. Neural networks based on them have shown outstanding performance in natural images and speeches classification tasks. In this paper, we especially focus on the image analysis. We first introduce neural networks and autoencoders, and provide an explanation of what autoencoders actually learn. Then, we explain how to stack up autoencoders and how to use the fine-tuning method for the purpose of making a high-performance image classifier. Finally, we carry out a numerical study with MNIST handwritten digit database.