Machine Learning Optimization Simulator

Training a machine learning model means searching a high-dimensional loss surface for a point where the loss is as low as possible. This simulator visualizes that search directly in 3D: a bumpy landscape stands in for a model's loss function, with a deep global minimum and two shallower local-minimum traps carved into it. Three balls — one per optimizer — start from the same point and race downhill using SGD, Momentum and Adam's real update rules. Tune the learning rate and momentum, toggle gradient noise to mimic mini-batch stochasticity, and watch which optimizer converges fastest, which overshoots, and which gets stuck in a shallow trap instead of reaching the true minimum.