Boosting semi-supervised learning with Contrastive Complementary Labeling.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Semi-supervised learning (SSL) effectively utilizes large unlabeled datasets for deep model training.
• Pseudo-labeling is a common SSL technique, but it often discards low-confidence predictions, potentially wasting valuable data.
• Existing methods struggle to leverage uncertain or low-confidence unlabeled data effectively.

Purpose of the Study:

• To propose a novel semi-supervised learning method that utilizes low-confidence unlabeled data.
• To introduce Contrastive Complementary Labeling (CCL) for improved deep model training.
• To enhance performance in label-scarce settings by maximizing the utility of all unlabeled data.

Main Methods:

• Developed Contrastive Complementary Labeling (CCL), a novel SSL approach.
• CCL leverages low-confidence data by identifying and utilizing complementary labels to form reliable negative pairs.
• Employs contrastive learning to maximize the use of all unlabeled data, including uncertain samples.

Main Results:

• CCL significantly enhances performance over existing advanced SSL methods.
• The method shows particular effectiveness in label-scarce settings.
• Achieved a 2.43% improvement over FixMatch on CIFAR-10 with only 40 labeled data.

Conclusions:

• Contrastive Complementary Labeling (CCL) offers a powerful new approach to semi-supervised learning.
• The method demonstrates the value of utilizing low-confidence data through complementary labels and contrastive learning.
• CCL provides substantial performance gains, especially when labeled data is limited.