Seminar on Computational Learning and Adaptation


  Theory and Application of Discriminative Hidden Markov Model Training

Sadik Kapadia
Motorola
Human Interface Lab
Palo Alto, CA
Sadik.Kapadia@motorola.com

Although discriminative training of hidden Markov models has been attempted for many years, success has been isolated and intermittent. The Baum-Welch "maximum likelihood" algorithm remains the training method of choice. In this talk I shall show that this puzzling state of affairs is due to some fundamental misunderstandings about the training problem. We show that "maximum likelihood" is well founded, and there exist discriminative formulations (e.g., maximum mutual information and frame discrimination) that enjoy some advantages. Experimental results using these techniques will be presented.


Date: Thurs., Jan 11

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule