• apply chain rule for partial derivatives
  
  
• derivative of loss wrt misclassification measure
  
  
• derivative of misclassification measure wrt the discriminant function
  
  
• derivative of discriminant function wrt observation probability
  
  
  • note that the discriminant is probability of all possible state sequences
  • path with highest probability can be made to dominate this measure -- Viterbi approximation
  
  
  
• derivative of observation probability wrt mean