Gradient Descent Viz is a desktop app that visually demonstrates various gradient descent methods in machine learning. By adjusting settings and exploring different surfaces, users can gain a better understanding of methods like Adam and RMSProp, which handle saddle points more effectively. The app allows users to tune parameters, watch step-by-step cartoon visualizations, and track elements like gradients, momentum, and paths. It is built in C++ using Qt and is cross-platform compatible. The code structure includes classes responsible for UI layout, plot interactions, animations, and mathematical implementations of each descent method. Contributions to turn the project into a web app are encouraged.
https://github.com/lilipads/gradient_descent_viz