Interpolation-Based Contrastive Learning for Few-Label Semi-Supervised Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised learning (SSL) effectively utilizes limited labeled data for model training.
• Consistency regularization methods in SSL show promise but degrade with extremely scarce labels.
• Data augmentation in SSL can lead to semantic information drift, hindering performance when labels are minimal.

Purpose of the Study:

• To address the performance degradation of consistency regularization-based SSL methods with very limited labels.
• To propose a novel SSL strategy that improves model accuracy and discriminative capability under extreme label scarcity.
• To enhance the reliability of data augmentation and the effectiveness of embedding learning in SSL.

Main Methods:

• Developed an interpolation-based method for constructing more reliable positive sample pairs in SSL.
• Designed a novel contrastive loss to ensure linear embedding changes between samples, enhancing discriminative power.
• Introduced a strategy that avoids destructive regularization, improving robustness in low-label settings.

Main Results:

• Achieved 88.73% classification accuracy on CIFAR-10 with only two labels per class, outperforming the second-best method (Comatch) by 5.3%.
• Demonstrated significant performance improvements when integrated with existing state-of-the-art SSL algorithms.
• The proposed method shows improved robustness and accuracy, especially in scenarios with extremely limited labeled data.

Conclusions:

• The proposed SSL method effectively tackles the semantic drift issue in data augmentation under extreme label scarcity.
• The novel approach enhances network discriminative capability by enlarging decision boundaries through guided linear embedding changes.
• This strategy offers a significant advancement for semi-supervised learning applications requiring minimal labeled data.