Robust Semi-Supervised Learning by Wisely Leveraging Open-Set Data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Open-set Semi-supervised Learning (OSSL) faces challenges with out-of-distribution (OOD) data, which can degrade performance in standard models.
• Existing OSSL methods often use all open-set data, including potentially detrimental samples, impacting model robustness.
• There is a need for strategies that effectively handle OOD data in OSSL settings.

Purpose of the Study:

• To develop a robust data selection strategy for OSSL that improves in-distribution (ID) classification.
• To propose a generic OSSL framework, WiseOpen, that selectively leverages open-set data.
• To enhance the performance and robustness of OSSL models in realistic scenarios.

Main Methods:

• Proposing Wise Open-set Semi-supervised Learning (WiseOpen), a framework utilizing a gradient-variance-based selection mechanism.
• Selectively training the model with a curated subset of open-set data, excluding unfriendly samples.
• Introducing two practical variants: low-frequency update and loss-based selection for computational efficiency.

Main Results:

• WiseOpen demonstrates superior performance compared to state-of-the-art OSSL methods.
• The gradient-variance-based selection effectively identifies and utilizes beneficial open-set data.
• Experimental results validate the effectiveness of WiseOpen and its variants in enhancing ID classification.

Conclusions:

• Selective data utilization in OSSL is crucial for mitigating the negative impact of OOD data.
• WiseOpen provides a theoretically grounded and practically effective approach to robust OSSL.
• The proposed method enhances model performance and offers computational advantages through its variants.