Bias Calibration for Semi-Supervised Continual Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Image classification is crucial for sensor data analysis in healthcare, environmental monitoring, and industry.
• Dynamic streaming data from sensors presents challenges like new categories and distribution shifts for traditional models.
• Semi-supervised continual learning is vital for handling label scarcity and high annotation costs in sensor data.

Purpose of the Study:

• To address confirmation and relational biases in current semi-supervised continual learning methods.
• To propose a novel Bias Calibration method integrating Confidence-Enhanced Learning and Guided Contrastive Learning.
• To improve image classification performance on sensor streaming data with limited labels.

Main Methods:

• Developed a Bias Calibration method based on nearest-neighbor semi-supervised continual learning.
• Integrated Confidence-Enhanced Learning to generate high-confidence pseudo-labels and mitigate confirmation bias.
• Employed Guided Contrastive Learning to optimize feature representations and reduce relational bias.

Main Results:

• The proposed method significantly outperforms existing approaches on CIFAR-10, CIFAR-100, and ImageNet-100 datasets.
• Demonstrated enhanced classification performance even with partial labeling of sensor streaming data.
• Successfully mitigated dual biases inherent in pseudo-label generation.

Conclusions:

• The Bias Calibration method offers a robust solution for semi-supervised continual learning with sensor data.
• This approach effectively leverages unlabeled data, reducing the need for extensive manual annotation.
• The integration of Confidence-Enhanced Learning and Guided Contrastive Learning shows promise for future advancements in the field.