Simulation run on this device. Scenario values are not sent anywhere — the engine is fully local.
math simulation · free
Gradient Descent Visualizer
Step descent down a loss surface you can see: lr, steps, and the where-it-lands trace.
Why does a big learning rate bounce out of the ravine?
Engine 1.0.0 · simulation
0.05
Stable below 0.1 on this surface — beyond that, y overcorrects.
25
- Final loss
- 1.86e-2
- Steps simulated
- 25
- Verdict
- Still descending — needs more steps (or bigger lr)
Method
- Minimizes f(x, y) = x² + 10y² — a ravine that punishes big steps along the steep axis.
- Update rule x ← x − lr·∂f/∂x per axis; stability on this surface breaks near lr ≈ 0.1.
- The path drawing is the exact trajectory of those updates.