Using samples with label noise for robust continual learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Leveraging samples with label noise can improve model robustness.
• Existing methods assume a consistent label space, which fails in continual learning due to label shifts.
• Label shifts in continual learning environments can exacerbate noise and degrade performance.

Purpose of the Study:

• To address the limitations of existing methods in continual machine learning with label noise.
• To propose a novel method, Shift-Adaptive Noise Utilization (SANU), for transforming noisy samples into usable data for continual learning.
• To mitigate the label shift problem and enhance model performance in dynamic learning environments.

Main Methods:

• SANU employs a source detection mechanism to identify the correct label space for noisy samples.
• A meta-knowledge representation module is utilized to improve the generalization of the detection process.
• Noisy samples are re-annotated using label guessing and generation strategies to adapt to label shifts.

Main Results:

• SANU effectively mitigates the label shift problem in continual learning.
• The method significantly enhances model performance by utilizing re-annotated noisy samples.
• Experimental results on three continual learning datasets validate the effectiveness of SANU.

Conclusions:

• SANU successfully transforms noisy data into valuable training inputs for continual learning.
• The proposed approach offers a robust solution for handling label noise under label shift conditions.
• This work advances the utilization of noisy data in dynamic machine learning scenarios.