ACTIVIS: Visual Exploration of Industry-Scale Deep Neural Network Models

Area of Science:

• Artificial Intelligence
• Machine Learning Visualization

Background:

• Deep learning models achieve high accuracy but lack interpretability.
• Existing visualization tools struggle with the complexity of industrial-scale models and datasets.

Purpose of the Study:

• To develop and evaluate an interactive visualization system for interpreting large-scale deep learning models.
• To address the unique design challenges posed by complex models and big data.

Main Methods:

• Participatory design sessions with over 15 researchers and engineers at Facebook.
• Development and iterative improvement of the ACTIVIS visualization system.
• Integration of multiple coordinated views (computation graph, neuron activation).

Main Results:

• ACTIVIS was developed, deployed, and improved based on user feedback.
• The system enables exploration of deep neural networks at instance- and subset-levels.
• Case studies and usage scenarios demonstrate ACTIVIS's utility with various models.

Conclusions:

• ACTIVIS provides an effective solution for interpreting large-scale deep learning models.
• Interactive visualization is crucial for understanding complex AI systems in industry.
• The system has been successfully deployed on Facebook's machine learning platform.