Learning Discriminative Stein Kernel for SPD Matrices and Its Applications.

Area of Science:

• Machine Learning
• Computer Vision
• Linear Algebra

Background:

• Stein Kernel (SK) is effective for classifying images represented by symmetric positive definite (SPD) matrices by analyzing eigenvalues.
• Directly using eigenvalues can be problematic due to estimation bias with limited samples and suboptimal discrimination.
• Eigenvalues represent individual matrix properties, not necessarily optimal for class differentiation in kernel methods.

Purpose of the Study:

• To propose a Discriminative Stein Kernel (DSK) that enhances classification performance for SPD matrices.
• To address limitations of standard SK by introducing adjustable eigenvalue parameters.
• To improve the discriminative power of kernel evaluations for SPD matrix data.

Main Methods:

• Introduced Discriminative Stein Kernel (DSK) with an adjustable parameter vector to modify eigenvalues.
• Optimized parameter values by maximizing a proxy of classification performance.
• Employed three common kernel learning criteria as performance proxies.
• Conducted comprehensive experiments on diverse image classification tasks.

Main Results:

• DSK demonstrated superior discrimination compared to standard SK and other similarity evaluation methods.
• Altering eigenvalues using DSK significantly improved alignment with classification objectives.
• DSK achieved higher classification performance across various image datasets.

Conclusions:

• The proposed DSK effectively enhances the classification of SPD matrices by optimizing eigenvalue adjustments.
• DSK offers a more discriminative and classification-aligned approach than traditional SK.
• This method shows significant potential for improving image classification tasks involving SPD matrix representations.