Week7-4HMM2

Observation likelihood

• Given multiple HMMs
  
  • which one is most likely to generate the observation sequence
• Naiive solution
  
  • try all possible state sequences

Forward algorithm

• compute a forward trellis that encodes all possible state path

HMM learning

-supervised
- training sequences are labeled
-unsupervised
- known number of states
- semi-supervised

Supervised HMM training

• estimate the static transition probabilities using MLE
  $$a_{ij} = \frac{Count(q_t = s_i, q_{t+1} = s_j)}{Count(q_t = s_i)}$$
• estimate the observation probabilities using MLE
  $$b_j(k) = \frac{Count(q_i = s_j, o_i = v_k)}{Count(q_i = s_j)}$$
• using smoothing

Unsupervised HMM training

• Given
  
  • observation sequences
• Goal
  
  • build the HMM
• Use EM methods
  
  • forward-backward(Baum-Welch) algorithm, an approximate solution for $P(O \mid \mu)$

Outline of Baum-Welch

• Algorithm
  
  • randomly set the parameters of HMM
  • Until the parameters converge repeat:
    
    • E step: determine the probability of the various state sequences for generating the observations
    • M step: reestimate the parameters based on these probabilities

The algorithm guarantees that at each iteration the likelihood of the data $P(O \mid \mu)$ increases
It converges to a local maximum