Variational Bayesian Inference for Hidden Markov Models With Multivariate Gaussian Output Distributions

Christian Gruhl, Bernhard SickarXiv 2016 · arXiv

Abstract

This paper presents a robust framework for training Hidden Markov Models (HMMs) using Variational Bayesian Inference (VI) tailored for multivariate Gaussian output distributions. By treating model parameters as random variables rather than fixed point estimates, the proposed approach mitigates numerical singularities common in standard Expectation-Maximization (EM) and provides a principled mechanism for handling uncertainty in time-series modeling.

Introduction

Hidden Markov Models (HMMs) are fundamental tools in time-series analysis, widely applied in fields such as speech recognition and gesture tracking. Traditionally, these models are trained using the Baum-Welch algorithm, a specific instance of the Expectation-Maximization (EM) algorithm. However, standard EM relies on point estimates, which frequently leads to "singularities"—scenarios where a model component collapses onto a single data point, causing variance to vanish and likelihood to diverge.

Furthermore, standard EM lacks a natural way to incorporate prior knowledge or quantify parameter uncertainty. While Variational Inference has been applied to HMMs previously, existing literature predominantly focuses on univariate (scalar) outputs. This work addresses this gap by deriving the mathematical framework necessary for multivariate data (vectors) using Gaussian-Wishart distributions.

Methodology

The core contribution of this work is the transition from "first-order" training (finding fixed parameters) to "second-order" training (finding distributions over parameters).

Conjugate Priors and the Variational Framework

To maintain mathematical tractability, the authors utilize conjugate priors. Transition probabilities are modeled via Dirichlet distributions. For multivariate Gaussian outputs, the Gaussian-Wishart distribution serves as a prior for both the mean vector ($\mu$) and the precision matrix ($\Lambda$).

Instead of direct likelihood maximization, the framework maximizes the Variational Lower Bound (ELBO). This involves minimizing the Kullback-Leibler (KL) divergence between an approximate distribution and the true posterior.

Graphical representation of the Variational HMM framework, illustrating the relationship between latent variables and parameters during the E-step and M-step.

The Modified Baum-Welch Process

The training follows an iterative process similar to EM:

  • E-Step: A modified forward-backward algorithm estimates latent states ($z_n$) using expectations derived from parameter distributions rather than fixed probabilities.
  • M-Step: The hyperparameters of the Dirichlet and Gaussian-Wishart distributions are updated based on the evidence gathered in the E-step.

Key Insights

  • Singularity Protection: The use of a Wishart prior over the covariance matrix prevents the model from collapsing into singularities. The prior ensures that variance remains non-zero even when a state is associated with minimal data.
  • Automatic Model Selection: The VI framework can effectively "shut down" redundant states by driving their mixing coefficients toward zero, facilitating the discovery of the optimal number of hidden states.
  • Uncertainty Quantification: By outputting distributions, the model provides a numerical measure of parameter uncertainty, which is valuable for downstream tasks like anomaly detection.

Conclusion and Future Work

The primary achievement of this research is the rigorous mathematical derivation of Variational HMMs for multivariate outputs, including the Expected Log Likelihood and the Variational Lower Bound equations. While the current work focuses on theoretical foundations, future research will involve empirical validation against standard Baum-Welch algorithms on real-world datasets and extending the framework to accommodate diverse distribution types within the same model.

Summary updated

BibTeX
@misc{gruhl2016variational,
  author = {Christian Gruhl and Bernhard Sick},
  title = {{Variational Bayesian Inference for Hidden Markov Models With Multivariate Gaussian Output Distributions}},
  year = {2016},
  eprint = {1605.08618},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1605.08618},
}
Figure