Optimizers ·
Loss Landscapes
Watch optimizers descend a 2-D parameter space; the height is the loss. Gradients are computed numerically, so the dynamics are real, not scripted.
—
—
Terrain
Optimizers
Hyperparameters
learning rate
1.0×
momentum β
0.90
speed
2×
Terrain noise
off
magnitude
0.35
frequency band
1.3
Regenerate field
Run
Pause
Reset
Random start
Live · loss & status
step 0
Loss vs step · time–loss trade-off
high
loss
low
drag to orbit · scroll to zoom ·
click the surface to drop a new start point