We propose a novel regularizer when training an autoencoder for unsupervised feature extraction. We explicitly encourage the latent representation to contract the input space by regularizing the norm
