Learning Hidden Markov Models from Aggregate Observations