A New Training Principle for Stacked Denoising Autoencoders | IEEE Conference Publication | IEEE Xplore