Learning Transferable Conceptual Prototypes for Interpretable Unsupervised Domain Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks have advanced unsupervised domain adaptation (UDA).
• Current UDA models often lack transparency, hindering applications requiring reliable decisions.
• Existing interpretable UDA methods are typically post-hoc and do not aid model learning.

Purpose of the Study:

• To develop an inherently interpretable UDA method that enhances both knowledge transfer and decision-making.
• To provide effective and intuitive explanations for UDA model behavior.
• To improve UDA performance by integrating interpretability into the learning process.

Main Methods:

• Proposed Transferable Conceptual Prototype Learning (TCPL), an inherently interpretable UDA approach.
• Designed a hierarchically prototypical module for transferring source domain concepts and learning domain-shared prototypes.
• Developed a self-predictive consistent pseudo-label strategy using confidence, predictions, and prototype information for sample selection.

Main Results:

• TCPL provides effective and intuitive explanations for the UDA process.
• The proposed method demonstrates superior performance compared to existing state-of-the-art UDA techniques.
• The approach successfully narrows the domain gap through informed pseudo-labeling.

Conclusions:

• TCPL offers a novel solution for interpretable unsupervised domain adaptation.
• The method effectively balances interpretability with performance enhancement in UDA.
• This work paves the way for more trustworthy and controllable deep learning models in domain adaptation scenarios.