Training Products of Experts by Minimizing Contrastive Divergence