Broad Metric Learning: A Fast and Efficient Discriminative Metric Learning Model.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Classical metric learning methods using linear transformations have limited representation capabilities.
• Deep metric learning methods can suffer from unstable training and convergence issues.
• Traditional algorithms often require extensive computational time for optimization, especially with high-dimensional data.

Purpose of the Study:

• To propose a novel Broad Metric Learning (BML) model for efficient and effective metric space learning.
• To address the limitations of existing linear and deep metric learning approaches.
• To enhance intraclass compactness and interclass separation in learned feature spaces.

Main Methods:

• BML utilizes a broad network for nonlinear feature mapping with random weights to a broad feature space.
• A linear transformation is learned to project data into a discriminative output space.
• Intraclass distances are minimized by referencing class-specific points, and Hard-Triplet Distance Learning (HDL) is employed for sample pair optimization.
• Closed-form solutions are used for efficient optimization of the linear transformation.

Main Results:

• BML demonstrates fast learning capabilities.
• The model achieves high classification and clustering accuracies across nine experimental datasets.
• Experimental results validate the efficiency and effectiveness of the BML model.

Conclusions:

• Broad Metric Learning (BML) offers an efficient and effective approach to metric space learning.
• BML overcomes the limitations of classical and deep metric learning methods.
• The proposed method shows significant improvements in classification and clustering performance.