GlassBox: Visual Debugging for Graph Neural Networks

Overview

GlassBox is an interactive visualization system that lets researchers inspect layer activations, attention weights, and gradient flow simultaneously across a graph neural network’s forward and backward passes, with linked views that highlight where model behavior diverges from expectations.

Approach

Rather than treating explainability as a report generated after training, GlassBox is designed for the debugging loop itself — fast enough to use interactively while iterating on an architecture, and integrated with common deep learning frameworks.

Team & collaborators

Built in collaboration with Diego Reyes and released as open source. Supported by an NSF grant supporting reproducible ML tooling.

Status

Active. A controlled user study with 15 ML researchers found GlassBox roughly halved the time to localize a known injected bug compared to standard logging-based workflows (published at IEEE VIS); current work extends the toolkit to large language models.