LossPlot: A Better Way to Visualize Loss Landscapes

TL;DR

LossPlot is a user-friendly platform that simplifies the visualization of deep neural network loss landscapes, enabling synchronized manipulation of multiple minimizers and offering customizable visualization features.

Contribution

This work introduces LossPlot, a semi-automated, interactive tool for visualizing loss landscapes with features not available in existing methods.

Findings

Enhanced visualization of loss landscapes with synchronized minimizer manipulation

User-friendly interface with customizable visualization options

Facilitates deeper understanding of neural network training dynamics

Abstract

Investigations into the loss landscapes of deep neural networks are often laborious. This work documents our user-driven approach to create a platform for semi-automating this process. LossPlot accepts data in the form of a csv, and allows multiple trained minimizers of the loss function to be manipulated in sync. Other features include a simple yet intuitive checkbox UI, summary statistics, and the ability to control clipping which other methods do not offer.