Decouple then Converge: Handling Unknown Unlabeled Distributions in Long-Tailed Semi-Supervised Learning.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Long-tailed semi-supervised learning (LTSSL) faces challenges when labeled and unlabeled data distributions differ.
• Existing LTSSL methods often fail due to biased pseudo-labels under distribution mismatch.
• This necessitates robust LTSSL algorithms that handle unknown unlabeled class distributions.

Purpose of the Study:

• To propose DeCon, a novel approach for LTSSL that addresses unknown unlabeled class distributions.
• To develop a method that maintains high performance across all classes despite distribution shifts.
• To provide a simple yet effective solution for real-world LTSSL tasks.

Main Methods:

• DeCon employs a dual-branch architecture: a standard branch for head classes and a balanced branch for tail classes.
• The two branches interact and converge during training, complementing each other.
• This decoupled learning strategy aims to mitigate issues arising from distribution mismatch.

Main Results:

• DeCon achieves state-of-the-art performance on standard LTSSL benchmarks.
• Demonstrates an average 2.7% absolute increase in test accuracy over existing algorithms with mismatched distributions.
• Consistently outperforms sophisticated LTSSL algorithms even with identical class distributions.

Conclusions:

• DeCon offers a simple and effective solution for LTSSL with unknown unlabeled class distributions.
• The decoupled learning approach successfully handles distribution mismatches, improving overall classification.
• The method shows significant performance gains and robustness across various LTSSL scenarios.