Adam Optimizer Landscape Simulator

Visualizing Non-Stationary Dynamics & Failure Modes
2D Gradient Contour Surface Saddle Point
Step: 0
Adam
SGD + Momentum
RMSprop
SGD
Adam Active
Loss: 0.00
Oscillation: 0.00
SGD + Momentum Active
Loss: 0.00
Oscillation: 0.00
RMSprop Active
Loss: 0.00
Oscillation: 0.00
SGD Active
Loss: 0.00
Oscillation: 0.00
Why "Throwing Adam At It" Fails
On anisotropic or non-stationary landscapes, Adam's second-moment estimation ($v_t$) can cause step scaling to overshoot narrow valleys or diverge when gradients shift faster than moments decay. Click anywhere on the canvas above to drop comparison trajectory probes!
Enjoy this tool? Build your own with Super