Bayesian Methods for Hidden Markov Models

By Steven L. Scott. 2002

610 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Statistics and Probability, Probabilistic & Bayesian Methods

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Statistics and Probability

In Probabilistic & Bayesian Methods

Citations in the corpus

Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates (2018, methodology): “In the ﬁrst pass, FFBS computes a set of forward probabilities for all subwords in the lattice, which provide the probability of ending up in any particular subword w .”
Effective Bayesian Modeling of Groups of Related Count Time Series (2014, methodology): “The hyperpriors that are used for the global parameters are given in the appendix.”
A segmental framework for fully-unsupervised large-vocabulary speech recognition (2016, methodology): “…Bayesian GMM; a new segmentation is sampled; and ﬁnally the embeddings from this new segmentation are added back into the Bayesian GMM. Line 5 uses the forward ﬁltering backward sampling dynamic programming al-gorithm [46] to sample the new embeddings; details of this step are given in Appendix A.”
Inference in Hidden Markov Models (2006, methodology): “For such models, the remark that the gradient of the log-likelihood may also be obtained using Fisher’s identity was made by Segal and Weinstein (1989); see also Weinstein et al. (1994). The sensitivity equations approach is in no way limited to Gaussian linear state-space models but may be applied ”
Fully unsupervised small-vocabulary speech recognition using a segmental Bayesian model (2015, methodology): “To sample a new segmentation for si (line 6), the forward filtering backward sampling algorithm is used [27].”
Exact and efficient Bayesian inference for multiple changepoint problems (2006, methodology): “The recursions are similar to those of the Forward-Backward algorithm (see Scott, 2002, for a review).”
Hidden Markov models for circular and linear-circular time series (2006, background): “(For key references to this literature see, e.g., Scott 2002 .) From a Bayesian point m is one of the unknown parameters to which one should assign a prior and then compute it’s posterior distribution.”
Likelihood inference for a class of latent Markov models under linear hypotheses on the transition probabilities (2006, background): “It has not been explicitly considered either within other inferential approaches or within the hidden Markov literature, which is strongly related to the literature concerning LM models (for a review see Archer and Titterington (2002) and Scott (2002)).”
Reinforcement learning with limited reinforcement: using Bayes risk for active learning in POMDPs (2008, positive): “We use a Markov chain Monte Carlo approach, forward-filtering backward-sampling (FFBS) [31] to sample dynamics models from this proposal distribution K(m).”
Reinforcement Learning with Limited Reinforcement : Using Bayes Risk for Active Learning in POMDPs Finale (2012, methodology): “To sample the state sequence, FFBS first computes the standard forward pass of the Viterbi algorithm [32]: αt+1(s) = Ω(s, at, ot) ∑ s′ αt(s ′)T (s, at, s ′) whereα0(s) is initialized to the initial belief state of the POMDP.”
The OS* Algorithm: a Joint Approach to Exact Optimization and Sampling (2012, methodology): “ond only to unigram max-backos, and thus each state corresponds to a NULLcontext. Over this WFSA, both optimization and sampling can be done biner, 1989]and \backward ltering-forward sampling" [Scott, 2002]), where the forward weights associated to states are computed similarly, either in the m”
Bayesian Nonparametrics: Hierarchical Bayesian nonparametric models with applications (2010, background): “This simple Gibbs sampler can converge very slowly due to strong dependencies among the latent states (Scott, 2002).”
Inference in hidden Markov models (2010, methodology): “For such models, the remark that the gradient of the log-likelihood may also be obtained using Fisher’s identity was made by Segal and Weinstein (1989); see also Weinstein et al. (1994). The sensitivity equations approach is in no way limited to Gaussian linear state-space models but may be applied ”
A Sticky HDP-HMM With Application to Speaker Diarization (2009, methodology): “An alternative changepoint detection technique, first proposed in Siegler et al. (1997), uses fixed length windows and computes the symmetric Kullback–Leibler (KL) divergence between a pair of Gaussians each fit by the data in their respective windows. A post-processing step then sets the changepoin”