Enhancing Sample Utilization in Noise-Robust Deep Metric Learning With Subgroup-Based Positive-Pair Selection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Noisy labels in real-world data degrade deep learning model performance.
• Robustness to noisy labels is well-studied in classification but underexplored in deep metric learning (DML).
• Current DML methods for noisy labels often discard potentially useful data, reducing sample utilization.

Purpose of the Study:

• To propose a novel noise-robust DML framework, SubGroup-based Positive-pair Selection (SGPS).
• To enhance sample utilization by constructing reliable positive pairs for noisy samples in DML.
• To improve the performance of DML models in the presence of label noise.

Main Methods:

• SGPS employs a probability-based strategy to identify clean and noisy samples.
• A subgroup generation module discovers similar samples within subgroups for noisy data.
• Positive prototypes are aggregated from similar samples, and a tailored contrastive loss is applied.
• The framework integrates easily with existing pair-wise DML tasks.

Main Results:

• SGPS effectively identifies clean and noisy samples.
• It constructs informative positive prototypes for noisy samples, enhancing their utility.
• Experiments on synthetic and real-world datasets demonstrate SGPS's effectiveness.
• The proposed method outperforms state-of-the-art noisy label DML techniques.

Conclusions:

• SGPS offers a robust and effective solution for deep metric learning with noisy labels.
• The framework significantly improves sample utilization compared to existing methods.
• SGPS provides a valuable contribution to the field of robust machine learning.