Using a ring parallel processor for hidden Markov model training

摘要

The authors present a novel solution to the computationally intensive problem of training HMMs (hidden Markov models) by showing how a bidirectional ring multiprocessor can achieve potentially optimal speed in the training of left-to-right HMMs. The solution presented avoids interprocessor communications problems in the HMM training algorithm. This is achieved by having the ring multiprocessor calculate the alpha 's (from the forward-backward training algorithm) in a clockwise direction around the ring, and the beta 's in a counterclockwise direction at the same time. The two sets of calculations are designed so that when this stage of the iteration is completed, each processor will have all of the data needed for the next stage of the iteration already stored locally.