Inconsistent parameter estimation in Markov random fields: Benefits in the computation-limited setting