An Iterative Approach to Discriminative Structure Learning Peng Xu WS’2001
Discriminative Structure Learning Procedure Baseline HMM training Viterbi alignment of training data Bivariate MI computation Discriminate structure selection Parameter re-estimation of new model WS’2001
Problems With the Procedure Viterbi alignment may change after training the new model MLE for parameter estimation Viterbi approximation restricts the MI computation WS’2001
Discriminative Model Learning Goal: minimize D(P(Q|O)||P’(Q|O)) the divergence between the desired posterior probability distribution and the posterior probability distribution according to the model WS’2001
Geometric Illustration of the Iterative Approach Desired Posterior Distribution Model Family WS’2001
BMM Structures Iteration n Iteration n+1 WS’2001
EM Type Iterative Structure and Parameter Learning E-step: Viterbi alignment of training data, MI computation M-step: discriminative conditional mutual information based BMM edge detection, model parameter learning (MMI) WS’2001
Improving MI Computation Viterbi alignment: Label for each frame is deterministic Posterior probability P(Qt=q|xt) is a function Soft alignment: Compute P(Qt=q|xt) using forward-backward algorithm Each data frame contributes to all labels WS’2001
Proposal for the Next Year Formal formulation of the iterative model structure and parameter learning Theoretical study of the EM type learning procedure Implementation of the improved MI computation Application to different data sets: Aurora, Audio-visual, etc. WS’2001