• optimize for minimum classification error on training data
  • classification could be at various levels
  • sentence, word and phone classification based loss functions
  • gradient descent on the loss function
  
  
• Generalized Probabilistic Descent
  • core of the MCE algorithm
  • guarantees convergence when certain forms of loss functions used
  
  
• computationally expensive
  • problems similar to MMI estimation with large data
  • simplification by using Viterbi alignments and N-best lists
  • online and batch modes available to update parameters