Link: Extracting and Composing Robust Features with Denoising Autoencoders

This paper introduces a novel method of unsupervised training method: denoising autoencoders. The network is trained layerwise by adding noise to the output of each previously trained layer. The author also studied the similarity of this method and other training methods, such as data augmentation(image transformation), the equivalence of regularization and adding noise, etc.