Training Products of Experts by Maximizing Contrastive Likelihood