Adaptive Batch Mode Active Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Active learning reduces manual data labeling effort for classifier training.
• Batch mode active learning (BMAL) selects multiple data points simultaneously.
• Existing BMAL methods often use static or heuristic batch sizes, limiting adaptability.

Purpose of the Study:

• To propose novel optimization-based frameworks for adaptive batch mode active learning (BMAL).
• To integrate batch size and selection criteria into a single adaptive formulation.
• To enhance the efficiency and applicability of active learning in data-rich environments.

Main Methods:

• Developed two optimization-based frameworks for adaptive BMAL.
• Utilized gradient-descent-based optimization strategies.
• Leveraged properties of submodular functions for algorithm derivation.

Main Results:

• Proposed adaptive BMAL algorithms achieve comparable computational complexity to state-of-the-art static BMAL.
• Empirical results on VidTIMIT and MOBIO datasets demonstrate the frameworks' efficacy.
• Validated the potential for real-world biometric recognition applications.

Conclusions:

• The proposed adaptive BMAL frameworks offer an efficient and effective approach to data selection.
• These methods provide a more flexible and robust solution compared to static batch selection.
• The research highlights the suitability of adaptive BMAL for practical biometric recognition tasks.