Pairwise Identity Verification via Linear Concentrative Metric Learning

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Pairwise identity verification tasks, such as face and speaker recognition, face challenges due to exclusive training and testing individuals and limited training data.
• Existing metric learning systems require robust optimization strategies to handle these constraints.

Purpose of the Study:

• To present a general framework for metric learning systems applied to pairwise identity verification.
• To investigate the effectiveness of different metric learning models (linear, nonlinear, similarity, distance) under restricted and unrestricted training settings.

Main Methods:

• Developed a general framework for metric learning systems.
• Employed the stochastic gradient descent algorithm for optimization.
• Evaluated linear and shallow nonlinear models for both similarity and distance metric learning.

Main Results:

• Learning a linear system on similar pairs only is preferable with limited training data, offering simplicity and competitive performance.
• The proposed approach achieved competitive results on the Labeled Faces in the Wild (LFW) face dataset and the NIST speaker dataset.
• A pretrained deep nonlinear model significantly improved face verification results.

Conclusions:

• For pairwise verification with limited data, a linear metric learning system trained on similar pairs is a practical and effective solution.
• Deep nonlinear models can enhance performance, particularly in face verification tasks.
• The study provides valuable insights into optimizing metric learning for identity verification systems.