Robust Distance Metric Learning via Bayesian Inference.

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Inference

Background:

• Existing distance metric learning (DML) algorithms are susceptible to label noise and overfitting due to point estimation.
• Robustness against noisy labels is crucial for reliable performance in real-world computer vision applications.

Purpose of the Study:

• To develop a robust DML algorithm using Bayesian inference to address the limitations of existing methods.
• To enhance DML's resilience to label noise and improve generalization performance.

Main Methods:

• Proposed a Bayesian extension to the large margin nearest neighbor classification method.
• Employed stochastic variational inference for estimating the posterior distribution of the transformation matrix.
• Theoretically analyzed the algorithm's robustness against label noise and derived generalization error bounds.

Main Results:

• Demonstrated bounded influence of noisy data points on the learned model.
• Derived generalization error bounds for the algorithm in the presence of label noise.
• Achieved state-of-the-art performance on three diverse datasets with varying label noise types.

Conclusions:

• The proposed Bayesian DML algorithm is robust to label noise and offers improved performance.
• The method is theoretically sound, providing generalization error bounds and demonstrating probably approximately correct-learnability.
• The algorithm represents a significant advancement for DML in noisy environments.