DendroMap: Visual Exploration of Large-Scale Image Datasets for Machine Learning with Treemaps.

Area of Science:

• Computer Science
• Data Visualization
• Machine Learning

Background:

• Exploring large-scale image datasets is crucial for machine learning (ML) model development.
• Current methods like image grids and t-SNE projections lack scalability and effective organization for massive datasets.
• Insufficient support for user interaction hinders deep data exploration.

Purpose of the Study:

• To introduce DendroMap, a novel interactive visualization approach for large-scale image datasets.
• To address the scalability and organizational limitations of existing ML data exploration techniques.
• To enable users to gain insights into dataset distributions and model performance.

Main Methods:

• Adapted Treemaps, a visualization technique, to organize images based on hierarchical cluster structures.
• Extracted hierarchical cluster structures from high-dimensional image representations.
• Developed interactive features for zooming and exploring data at multiple levels of abstraction.

Main Results:

• DendroMap effectively organizes large image datasets, enabling users to understand overall distributions.
• Case studies on deep learning datasets demonstrated insights into image diversity, underperforming subgroups, and classification errors.
• A user study showed participants preferred DendroMap over a gridified t-SNE for grouping and searching tasks.

Conclusions:

• DendroMap provides an effective and preferred method for interactively exploring large image datasets in machine learning.
• The hierarchical organization and interactive features facilitate deeper data understanding and model analysis.
• DendroMap enhances the discovery of dataset characteristics and potential model improvements.