Dual-Space Contrastive Learning for Open-World Semi-Supervised Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) faces scalability challenges with unseen classes in unlabeled data.
• Open-world SSL (OWSSL) addresses this, but enhancing novel class representation remains difficult.
• Existing contrastive learning methods in OWSSL often focus on single spaces (feature or prediction).

Purpose of the Study:

• To propose a novel dual-space contrastive learning (DSCL) method for Open-World Semi-Supervised Learning (OWSSL).
• To enhance the representative ability of unlabeled samples, particularly those from novel classes.
• To explore and leverage information potentials across both feature and prediction spaces.

Main Methods:

• DSCL employs two modules: intraspace and interspace contrastive learning.
• The intraspace module bridges feature and prediction spaces using a learnable classifier for contrastive learning.
• The interspace module integrates neighborhood features and prediction space clustering to improve representation.

Main Results:

• DSCL significantly outperforms state-of-the-art methods on various benchmarks.
• Demonstrated superior performance on CIFAR100, Imagenet100, CIFAR10, CUB-200, and Scar datasets.
• Effectively improves representative ability and intraclass compactness.

Conclusions:

• Dual-space contrastive learning (DSCL) offers a powerful approach for OWSSL.
• Leveraging complementary information from dual spaces is crucial for handling novel classes.
• DSCL provides a robust framework for advancing OWSSL research and applications.