This page recommends the use of sparsified overcomplete autoencoders (i.e. the learned representation has a higher dimensionality than the input, but many of the values are zeros).
What is the advantage of this over a normal autoencoder?
[link][4 comments]