Data-Driven Knowledge Fusion for Deep Multi-Instance Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Multi-instance learning (MIL) is crucial for complex data structures.
• Current deep MIL methods often trap knowledge within algorithms, limiting model development.
• Existing MIL approaches can lead to significant knowledge loss and hinder the creation of more powerful models.

Purpose of the Study:

• Propose a novel data-driven knowledge fusion for deep MIL (DKMIL) algorithm.
• Develop a new interface between data and models to enhance learning ability and scalability.
• Improve the effectiveness of deep MIL by extracting and fusing knowledge from key data samples.

Main Methods:

• Introduced a data-driven knowledge fusion (DKF) module to analyze key sample decision-making.
• Designed a knowledge fusion module to extract valuable information from key samples for model learning.
• Proposed a two-level attention (TLA) module to learn shallow and deep features for effective classification.

Main Results:

• The DKF module provides a new interface between data and models, offering strong scalability.
• Prior knowledge from existing algorithms can be integrated to enhance the model's learning ability.
• The TLA module enables gradual learning of shallow- and deep-level features for improved classification.

Conclusions:

• The proposed DKMIL algorithm offers a novel approach to deep MIL by fusing data-driven knowledge.
• The DKF and TLA modules demonstrate the scalability and efficiency of the proposed architecture.
• Experiments on 62 datasets across five categories validate the effectiveness of the DKMIL approach.