Weight space

Every point on this surface is a possible model. Its height is the loss: how wrong it is.

Hover to raise peaks (high error). Hold Shift to dig valleys (low error). Click to place a model. Run gradient descent to watch it roll toward the lowest ground.

wt+1  =  wt  −  η · ∇J(wt)  +  β · vt
Learning rate η0.050
Momentum β0.70
Step0
Loss J(w)0.000
|∇J|0.000
Place a model and press run.
Hover to raise peaks · Shift to dig valleys · Click to place a model · Run gradient descent Drag to sculpt the surface · Tap to place a model · Run gradient descent