Information-theoretic semi-supervised metric learning via entropy regularization.

Area of Science:

• Machine Learning
• Information Theory
• Computer Vision

Background:

• Semi-supervised learning leverages limited labeled data with abundant unlabeled data.
• Metric learning aims to learn a distance function for improved data representation.
• Existing methods often rely on manifold assumptions, limiting their applicability.

Purpose of the Study:

• To propose a general information-theoretic approach for semi-supervised metric learning.
• To develop a method, SERAPH, that does not rely on the manifold assumption.
• To integrate supervised and unsupervised learning components effectively.

Main Methods:

• SERAPH utilizes entropy maximization on labeled data and minimization on unlabeled data.
• Entropy regularization enhances manifold regularization by incorporating dissimilarity information.
• Trace-norm regularization is applied to encourage low-dimensional projections.
• Non-convex optimization is addressed using gradient projection or EM-like algorithms.

Main Results:

• SERAPH demonstrates superior performance compared to established metric learning methods.
• The learned Mahalanobis distance exhibits high discriminative power.
• The method is robust in noisy environments.

Conclusions:

• SERAPH offers a flexible and effective framework for semi-supervised metric learning.
• The information-theoretic approach provides a principled way to integrate labeled and unlabeled data.
• The method's independence from manifold assumptions broadens its applicability.