Adam wins on most surfaces: it adapts the lr per-parameter. SGD is slow but sometimes generalises better. The Rosenbrock valley is intentionally hard - flat in one direction, steep in the other.
Try: lr=0.001 - SGD barely moves. Increase to 0.01 and watch Adam sprint while Momentum builds speed gradually.
Optimizer Race - Interactive Visualization
Modern ML uses Adam, not plain SGD - but why? This visualization races SGD, Momentum, RMSProp, and Adam from the same starting point on a challenging loss surface contour plot. Adam reaches the minimum fastest by maintaining per-parameter learning rates and first/second moment estimates. Momentum overshoots and oscillates. Plain SGD trudges along slowly, often following a zigzag path.
Watch SGD, Momentum, RMSProp, and Adam race to minimum simultaneously
See color-coded paths on the loss contour map
Adjust learning rate to see all methods affected differently
Understand why adaptive methods (Adam, RMSProp) are preferred
See Momentum overshooting and SGD zigzagging vs Adam's smooth path
Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.