Bag-Level Aggregation for Multiple-Instance Active Learning in Instance Classification Problems.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Training recognition systems often requires large, weakly annotated datasets.
• Active learning (AL) reduces labeling costs by querying experts for informative instances.
• Existing AL methods are unsuitable for multiple instance learning (MIL) due to data structure.

Purpose of the Study:

• Develop novel active learning strategies for multiple instance learning (MIAL).
• Address the challenge of labeling costs in MIL with limited expert interaction.
• Improve the efficiency of training classifiers on weakly labeled data.

Main Methods:

• Proposed two new MIAL methods: aggregated informativeness and cluster-based aggregative sampling.
• Aggregated informativeness queries bags based on classifier uncertainty and information content.
• Cluster-based aggregative sampling uses hierarchical clustering and considers bag/instance labels.

Main Results:

• Both proposed MIAL methods significantly outperformed reference methods in experiments.
• Achieved comparable performance to single-instance AL methods with fewer queries.
• Demonstrated effectiveness across benchmark datasets from various application domains.

Conclusions:

• The proposed MIAL strategies effectively reduce the number of required queries.
• Appropriate MIAL strategies are crucial for efficient training of recognition systems.
• These methods offer a significant advantage in scenarios with weakly annotated data.