Diversify and Conquer: Open-Set Disagreement for Robust Semi-Supervised Learning With Outliers.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) assumes identical class distributions in labeled and unlabeled data, which is often violated by outliers in real-world scenarios.
• Outliers in unlabeled data are typically treated as noise, significantly degrading SSL model performance.
• Existing open-set SSL (OSSL) methods struggle with insufficient labeled data, leading to performance degradation.

Purpose of the Study:

• To propose a novel framework, Diversify and Conquer (DAC), to enhance SSL robustness in open-set scenarios.
• To develop a method for robust outlier detection that is effective even with underspecified labeled data.
• To enable the identification of unknown concepts by exploiting prediction discrepancies among multiple models.

Main Methods:

• The DAC framework constructs multiple, differently biased models within a single training process.
• It encourages divergent model heads to exhibit varied biases towards outliers while maintaining consistent predictions for inliers.
• Prediction disagreements among these differently biased models are leveraged to detect unknown concepts.

Main Results:

• The proposed DAC method demonstrates robust outlier detection capabilities, even with limited labeled data.
• DAC significantly outperforms existing state-of-the-art OSSL methods across various experimental protocols.
• The framework effectively mitigates performance degradation caused by outliers in semi-supervised learning.

Conclusions:

• The DAC framework offers a robust solution for open-set semi-supervised learning by effectively handling outliers.
• Leveraging prediction disagreements among diverse models is a promising strategy for outlier detection and unknown concept identification.
• DAC enhances the practical applicability of SSL in real-world datasets containing unknown classes.