6.891 HW #4 Clarifications

  1. Turn in solutions to part I or part II, not both.
    1. The code will not necessarily "work" without the modifications indicated in the homework. Updated versions of inimix.m and runmix.m have been posted. runmix.m has been changed into a function so that less top-level code will be needed.
      1. Note that there is no need to turn in anything for this part. It would be good to check that the likelihood of your model is increasing to make sure that it is working.
  2. Turn in solutions to part I and part II.
      1. Turn in a minimum of three (3) "outliers." Note that an "outlier" is an entire sequence (not just a single character).
    1. The `eps' parameter in newhmm.m dictates the amount of randomness that is injected into the initial HMM. A value of `1' will work fine.
      1. A bug has been found in viterbi.m. A new version has been posted.
    1. You can treat an "operation" as "updating the probabilities in this node."
    2. Show that the probabilities explicit in the junction tree are enough to do active learning. i.e. show that you will not need to invoke Bayes Law.