Learning Log-Determinant Divergences for Positive Definite Matrices.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Symmetric Positive Definite (SPD) matrices are vital for visual learning, capturing second-order statistics.
• Existing SPD matrix similarity measures offer benefits but lack a systematic selection method, often relying on trial-and-error.
• The challenge lies in choosing the optimal similarity measure for diverse machine learning applications.

Purpose of the Study:

• To propose a data-driven approach for learning similarity measures for SPD matrices.
• To extend the αβ-log-det divergence by learning its continuous parameters from data.
• To enhance the expressiveness of similarity measures by learning vector-valued parameters.

Main Methods:

• Capitalizing on the αβ-log-det divergence, a meta-divergence with scalar parameters α and β.
• Treating the αβ-log-det divergence parameters as a continuum to be learned from data.
• Extending parameter learning to vector-valued parameters for increased measure expressiveness.
• Integrating divergence learning with discriminative dictionary learning and SPD matrix clustering.
• Employing Riemannian gradient descent for efficient optimization.

Main Results:

• Demonstrated the effectiveness of learned similarity measures across eight standard computer vision tasks.
• Successfully integrated divergence learning with supervised and unsupervised machine learning problems.
• Showcased improved performance in visual learning applications through data-driven similarity measure optimization.

Conclusions:

• The proposed data-driven learning of similarity measures for SPD matrices offers a significant advancement over traditional methods.
• This approach enhances the adaptability and performance of visual learning algorithms.
• The method provides a systematic and efficient way to optimize similarity measures for specific machine learning tasks.