GRADIENT DESCENT EXPLORER
The optimizer is the same point in both views: height is loss in 3D, while the contour map exposes direction and path shape in 2D.
ready
3D
loss surface · drag to orbit
2D
contour map · drag starting point
Surface
Convex bowl
Curved ravine
Local minima
Learning rate
0.080
Momentum
0.60
Gradient noise
0.00
Take one step
Run
Reset path
Iteration
0
Position (x, y)
—
Loss
—
Gradient magnitude
—
Try it:
raise the learning rate until the path crosses the ravine, then add momentum and watch the oscillation change.
current point
optimization path
lower loss