Unsupervised Fusion Feature Matching for Data Bias in Uncertainty Active Learning

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Active learning (AL) aims to select valuable data for model training.
• Traditional uncertainty-based AL methods often suffer from data bias, failing to represent the entire unlabeled dataset.
• Existing solutions incur high training costs or require task-specific redesigns.

Purpose of the Study:

• To propose a novel active learning framework that mitigates data bias.
• To introduce an efficient and adaptable feature-matching-based uncertainty method.
• To reduce the computational burden associated with active learning.

Main Methods:

• Developed unsupervised fusion feature matching (UFFM) for resampling uncertain data.
• Implemented feature matching to remove similar data points and alleviate bias.
• Redesigned classic uncertainty methods for complex visual tasks within the AL framework.

Main Results:

• UFFM demonstrated superior performance compared to existing unsupervised feature matching techniques.
• The proposed uncertainty calculation method outperformed random sampling, traditional uncertainty approaches, and state-of-the-art methods.
• Experimental validation on standard benchmark datasets confirmed the effectiveness of the approach.

Conclusions:

• The proposed feature-matching-based uncertainty method effectively addresses data bias in active learning.
• UFFM offers a computationally efficient and adaptable solution for active learning.
• This work advances the field of active learning by providing a robust and cost-effective approach for data selection.