Maximum A Posteriori Approximation of Hidden Markov Models for Proportional Sequential Data Modeling With...

Area of Science:

• Machine Learning
• Time Series Analysis
• Pattern Recognition

Background:

• Hidden Markov Models (HMMs) are foundational for generative machine learning in time series.
• HMMs have applications spanning speech recognition, video classification, and text translation.
• Generalized Dirichlet HMMs offer efficiency in modeling proportional sequential data.

Purpose of the Study:

• To investigate a maximum a posteriori (MAP) framework for inferring parameters of generalized Dirichlet HMMs.
• To introduce a novel approach that differs from the Baum-Welch algorithm by incorporating regularizing priors.
• To integrate a feature selection paradigm within the parameter inference algorithm.

Main Methods:

• Development of a maximum a posteriori (MAP) inference framework.
• Incorporation of priors for regularization of parameter estimation.
• Simultaneous integration of a feature selection paradigm.

Main Results:

• The proposed MAP framework with priors and feature selection was validated.
• Successful application in the classification of dynamic textures.
• Effective application in the recognition of infrared actions.

Conclusions:

• The developed MAP framework provides an effective method for parameter inference in generalized Dirichlet HMMs.
• The integrated feature selection enhances performance in complex recognition tasks.
• This approach advances the application of HMMs in dynamic texture and action recognition.