Expectation Maximization and Spectral Learning for Hidden Markov Model

BaumWelch (Forward-backward) learning algorithm for fitting the data.

LearnHMM utilizes only upto first three order moments and take advantage of the moment matching idea. Please refer to Daniel Hsu's paper (A spectral learning algorithm for Hidden Markov Models) for more details. Note that there is a well-known issue, called the negative probability problem associated spectral learning. In this implement we leave the problem as is.