Distilling BlackBox to Interpretable Models for Efficient Transfer Learning.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning Interpretability
• Medical Imaging Analysis

Background:

• Generalizable AI models are crucial for healthcare but struggle with data distribution shifts.
• Fine-tuning AI models requires extensive labeled data in new domains.
• Interpretable AI models often underperform compared to black-box models.

Purpose of the Study:

• To develop an interpretable AI model efficiently fine-tunable to unseen domains with minimal cost.
• To create a mixture of shallow interpretable models that achieve comparable performance to black-box models.
• To leverage pseudo-labeling and fine-tuning for domain adaptation in medical AI.

Main Methods:

• Distilling a black-box model into a mixture of shallow, human-understandable interpretable models.
• Utilizing the domain-invariant assumption for interpretable components.
• Applying pseudo-labeling from semi-supervised learning for target domain concept classification.
• Fine-tuning interpretable models in the target domain for efficient adaptation.

Main Results:

• A mixture of interpretable models achieved performance comparable to black-box models.
• The proposed method enables efficient fine-tuning to unseen domains with minimal computational cost.
• The model demonstrated effectiveness on a large-scale chest X-ray classification dataset.

Conclusions:

• The developed interpretable AI approach facilitates efficient domain adaptation in healthcare.
• This method addresses the challenge of generalizability in AI models for medical applications.
• The interpretable nature of the model aids in understanding AI decision-making in clinical settings.