Back to Publications
conference · 2023

Interactive Visualization Techniques for Deep Model Debugging

Abhiraj Bibhar, Diego Reyes · IEEE VIS

VisualizationDebugging

Abstract

Debugging deep learning models remains largely a manual, intuition-driven process. We present GlassBox, an interactive visualization system that lets researchers inspect layer activations, attention weights, and gradient flow simultaneously across a model’s forward and backward passes, with linked views that highlight where behavior diverges from expectations.

A user study with 15 machine learning researchers found that GlassBox reduced the time to localize a known injected bug by roughly half compared to standard logging-based debugging workflows.

Key contributions

  • A linked-view visualization system spanning activations, attention, and gradients.
  • A controlled user study demonstrating reduced debugging time versus standard workflows.
  • An open-source release integrated with common deep learning frameworks.
PDF DOI Publisher
Back to all publications

On this page

Authors

  • A Abhiraj Bibhar (this author)
  • D Diego Reyes

Cite this paper

Bibhar, A. & Reyes, D. (2023). Interactive Visualization Techniques for Deep Model Debugging. IEEE VIS.

@inproceedings{bibhar2023interactive,
  title     = {Interactive Visualization Techniques for Deep Model Debugging},
  author    = {Bibhar, Abhiraj and Reyes, Diego},
  booktitle   = {IEEE VIS},
  year      = {2023}
}