LossPlot: A Better Way to Visualize Loss Landscapes
This tool addresses a niche problem for researchers investigating loss landscapes, but it is incremental as it builds on existing visualization methods.
The authors tackled the laborious process of visualizing loss landscapes in deep neural networks by developing LossPlot, a platform that semi-automates this task, offering features like synchronized manipulation of multiple trained minimizers and clipping control not available in other methods.
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.