Graph-Based Self-Training for Semi-Supervised Deep Similarity Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Semi-supervised learning leverages both labeled and unlabeled data for training deep neural networks.
• Self-training methods offer good generalization but are limited by pseudo-label accuracy.
• Reducing noise in pseudo-labels is crucial for improving semi-supervised learning performance.

Purpose of the Study:

• To propose novel methods for reducing noise in pseudo-labels within self-training frameworks.
• To enhance the accuracy and confidence of predictions in semi-supervised learning.
• To improve the overall performance of deep neural networks using limited labeled data.

Main Methods:

• Developed a similarity graph structure learning (SGSL) model to capture correlations between unlabeled and labeled samples.
• Introduced an uncertainty-based graph convolutional network (UGCN) to aggregate features and quantify prediction uncertainty.
• Proposed a positive and negative self-training framework integrating SGSL and UGCN for end-to-end training.

Main Results:

• The SGSL model facilitates learning more discriminative features, leading to improved prediction accuracy.
• The UGCN reduces pseudo-label noise by selecting samples with low uncertainty and enhances feature discriminability.
• The combined framework with positive and negative pseudo-labels effectively improves semi-supervised learning performance.

Conclusions:

• The proposed SGSL and UGCN methods effectively reduce pseudo-label noise, enhancing self-training in semi-supervised learning.
• The positive and negative self-training framework introduces additional supervised signals, boosting model performance.
• This approach offers a promising direction for improving deep neural network training with limited labeled data.