Understanding the limitations of autoencoders