LossLens: Diagnostics for Machine Learning Through Loss Landscape Visual Analytics

Area of Science:

• Machine Learning
• Data Visualization
• Artificial Intelligence

Background:

• Neural network training optimizes parameters via loss functions.
• Analyzing loss landscapes offers insights into network architecture and learning.
• Global loss landscape visualization is challenging.

Purpose of the Study:

• Introduce LossLens, a visual analytics framework for exploring neural network loss landscapes.
• Enable multi-scale analysis of loss landscapes.
• Enhance model diagnostics through integrated global and local metrics.

Main Methods:

• Developed LossLens, a visual analytics framework.
• Integrated global and local scale metrics.
• Applied framework to case studies involving ResNet-20 and physics-informed neural networks.

Main Results:

• LossLens provides a comprehensive visual representation of loss landscapes.
• Framework facilitates understanding of architectural influences (e.g., residual connections).
• Demonstrated utility in visualizing physics-informed neural network parameter effects.

Conclusions:

• LossLens effectively addresses the challenge of conceptualizing and visualizing global loss landscapes.
• The framework enhances model diagnostics by integrating multi-scale loss landscape information.
• LossLens shows promise for analyzing diverse neural network architectures and applications.