ChannelExplorer: Exploring Class Separability Through Activation Channel Visualization

Area of Science:

• Computer Vision
• Machine Learning
• Data Visualization

Background:

• Deep neural networks (DNNs) excel in vision tasks but their internal workings, especially class separability, are poorly understood.
• Analyzing activation channels and layer contributions is crucial for interpreting DNN behavior.

Purpose of the Study:

• Introduce ChannelExplorer, an interactive visual analytics tool for analyzing DNNs.
• Focus on data-driven insights to explore class separability across model layers.
• Support understanding of diverse model architectures like CNNs, GANs, ResNet, and Stable Diffusion.

Main Methods:

• ChannelExplorer provides a dataset-level overview, progressively drilling down to individual examples.
• Utilizes three coordinated views: Scatterplot View for class confusion, Jaccard Similarity View for activation overlap, and Heatmap View for channel patterns.
• Summarizes activations across model layers to reveal insights into class separability.

Main Results:

• Demonstrated ChannelExplorer's utility in generating ImageNet class hierarchy.
• Successfully used the tool to find mislabeled images within datasets.
• Enabled identification of specific activation channel contributions to model performance.
• Facilitated locating latent states within Stable Diffusion models.

Conclusions:

• ChannelExplorer offers a novel approach to understanding DNNs by focusing on class separability through visual analytics.
• The tool is versatile, supporting various deep learning architectures and use cases.
• Expert user evaluation confirmed the tool's effectiveness in providing actionable insights into DNN behavior.