A Visual Analytics Framework for Explaining and Diagnosing Transfer Learning Processes.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Statistical learning models often assume identical data distributions, which is unrealistic and hinders knowledge reuse across domains.
• Transfer Learning (TL) addresses this by modeling inter-domain relationships, crucial for deep learning to reduce data and time demands.
• Existing visual analytics tools lack methods for explaining and diagnosing the knowledge transfer process in deep learning.

Purpose of the Study:

• To present a visual analytics framework for multi-level exploration of transfer learning in deep neural networks.
• To explain how knowledge is transferred from a source model to a target task during deep network training.
• To provide insights into the knowledge transfer process for better model development and application.

Main Methods:

• Developed a multi-aspect visual analytics framework for exploring transfer learning.
• Employed descriptive visualizations incorporating performance metrics.
• Included detailed inspections at statistical, instance, feature, and model structure levels.

Main Results:

• The framework enables a multi-level examination of knowledge transfer in deep neural networks.
• Visualizations facilitate understanding of how learned knowledge is applied to new tasks.
• Case studies on image classification demonstrate the framework's utility for analysts.

Conclusions:

• The proposed framework offers a novel approach to understanding and diagnosing transfer learning in deep learning.
• It supports analysts in exploring the intricacies of knowledge transfer across domains.
• This work contributes to more effective and interpretable deep learning applications.