Stacked Denoising Autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion