Towards Better Analysis of Deep Convolutional Neural Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) excel in pattern recognition but lack clear understanding for development.
• Model refinement often involves extensive trial-and-error due to opaque internal workings.

Purpose of the Study:

• To present a visual analytics approach for understanding, diagnosing, and refining deep CNNs.
• To provide insights into the functionality and interactions within deep learning models.

Main Methods:

• Formulating deep CNNs as directed acyclic graphs.
• Developing a hybrid visualization to display neuron facets and interactions.
• Employing hierarchical rectangle packing and matrix reordering for feature visualization.
• Utilizing biclustering-based edge bundling to manage visual complexity.

Main Results:

• The proposed visualization effectively discloses neuron characteristics and interconnections.
• Hierarchical packing and matrix reordering reveal derived features of neuron clusters.
• Edge bundling reduces clutter in visualizing large-scale network connections.

Conclusions:

• The visual analytics approach offers a method for better understanding and refining deep CNNs.
• The developed visualization techniques aid in diagnosing and improving deep learning model performance.
• Evaluation on CNNs shows favorable results, indicating the utility of the approach.