Learning in Markov Random Fields using Tempered Transitions