VAC-CNN: A Visual Analytics System for Comparative Studies of Deep Convolutional Neural Networks.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) have driven significant progress in machine learning applications.
• Existing tools for CNN model comparison are limited, often restricting analysis to two models and lacking customization.
• Visualizing only quantitative features like accuracy is insufficient for understanding CNN model behaviors.

Purpose of the Study:

• To present VAC-CNN, a visual analytics system designed for in-depth inspection and comparative analysis of single or multiple CNN models.
• To overcome the limitations of existing tools by enabling the comparison of a larger number of CNN models (e.g., tens).
• To facilitate a flexible and interactive workflow for evaluating CNN models.

Main Methods:

• Developed VAC-CNN, a visual analytics system with model visualization and explaining support.
• Implemented a highly interactive workflow presenting quantitative and qualitative information.
• Designed for both single model inspection and comparative studies of multiple models.

Main Results:

• VAC-CNN supports in-depth inspection of single CNN models.
• The system enables comparative studies of two or more CNN models, distinguishing itself by handling a larger number of models.
• Demonstrated effectiveness through use cases and a preliminary evaluation study on ImageNet dataset.

Conclusions:

• VAC-CNN provides a powerful tool for evaluating and comparing multiple CNN models.
• The system assists machine learning practitioners, especially novices, in understanding CNN behaviors.
• VAC-CNN enhances the analysis of CNN models through interactive visualization and explanation support.