Explicit Metric-Based Multiconcept Multi-Instance Learning With Triplet and Superbag.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multi-instance learning (MIL) is widely used but often implicitly models instance-bag correlations.
• Previous MIL methods have overlooked the significance of negative instances.
• A need exists for explicit and interpretable MIL approaches.

Purpose of the Study:

• To propose a novel metric-based multi-concept MIL approach.
• To explicitly identify instance categories and their correlations.
• To improve the generalization performance and interpretability of MIL models.

Main Methods:

• Developed a triplet-based bag embedding method for explicit instance categorization and attention weighting.
• Introduced an instance correlation metric within a superbag framework to address multi-concept challenges.
• Employed weak supervision for bag embedding.

Main Results:

• Artificial data experiments demonstrated the interpretability of the proposed network.
• Comparison experiments confirmed favorable performance across multiple tasks.
• Ablation studies validated the effectiveness of the presented method.

Conclusions:

• The proposed metric-based multi-concept MIL approach effectively addresses limitations of previous methods.
• The method offers enhanced interpretability and improved generalization.
• This work contributes a more robust and understandable framework for multi-instance learning.