Entropy-Based Variational Scheme with Component Splitting for the Efficient Learning of Gamma Mixtures

Area of Science:

• Machine Learning and Data Mining
• Statistical Modeling

Background:

• Finite Gamma mixture models offer flexibility and incorporate prior information for improved generalization.
• These models are valuable for various machine learning and data mining tasks.

Purpose of the Study:

• To propose an efficient Gamma mixture model-based approach for proportional vector clustering.
• To develop an entropy-based variational algorithm for simultaneous model learning and complexity optimization.

Main Methods:

• Developed a sophisticated entropy-based variational algorithm for model learning.
• Investigated a component-splitting principle for model selection and overfitting prevention within the variational framework.

Main Results:

• An efficient Gamma mixture model-based framework for proportional vector clustering was successfully developed.
• The method simultaneously learns the model and optimizes its complexity.

Conclusions:

• The proposed framework effectively handles model selection and prevents overfitting.
• Demonstrated performance on challenging applications like dynamic texture clustering, object categorization, and human gesture recognition.