Generalized competitive learning of gaussian mixture models.

Area of Science:

• Machine Learning
• Statistical Modeling
• Computer Vision

Background:

• Selecting the optimal number of components (model selection) is critical for Gaussian mixture models.
• Existing methods like Expectation-Maximization (EM) can be sensitive to initial parameter choices.

Purpose of the Study:

• To address the model selection problem in Gaussian mixtures using regularization theory.
• To develop an Entropy Regularized Likelihood (ERL) learning algorithm for Gaussian mixtures.

Main Methods:

• Developed an Entropy Regularized Likelihood (ERL) learning framework for Gaussian mixtures.
• Proposed a gradient-based algorithm to implement the ERL learning.
• Analyzed the inherent generalized competitive learning mechanism within ERL.

Main Results:

• ERL learning enables automatic model selection for Gaussian mixtures.
• The ERL algorithm demonstrates reduced sensitivity to initialization compared to standard EM.
• Experimental results on simulated data validate the theoretical analysis.
• ERL outperformed other competitive learning algorithms in unsupervised image segmentation tasks.

Conclusions:

• Entropy Regularized Likelihood (ERL) learning provides an effective solution for model selection in Gaussian mixtures.
• The proposed gradient algorithm offers a robust and efficient approach for Gaussian mixture model fitting.
• ERL learning shows significant potential for applications such as unsupervised image segmentation.