This repository contains an implementation of the denoising autoencoder, as defined in Stacked Denoising Autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion (P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio and P.A. Manzagol).
Training the encoder on handwritten digit images from the MNIST data set yields reasonable reconstruction performance and generates filters that look like edge and blob detectors:
