Temporal Parallelization of Inference in Hidden Markov Models

This paper presents algorithms for parallelization of inference in hidden Markov models (HMMs). In particular, we propose parallel backward-forward type filtering and smoothing algorithm as well Viterbi-type maximum-a-posteriori (MAP) algorithm. We define associative elements operators to pose these problems parallel-prefix-sum computations sum-product max-product parallelize them using parallel-scan algorithms. The advantage the proposed is that they are computationally efficient HMM with long time horizons. empirically compare performance methods classical on a highly graphical processing unit (GPU).