Question 1

Why run three optimizers at once?

Accepted Answer

From the same start and learning rate, each algorithm takes a different path and converges at a different speed. SGD is simple but oscillates in valleys; Momentum keeps direction with inertia; Adam adapts step size per parameter.

Question 2

What is the Rosenbrock valley?

Accepted Answer

f(x,y)=(1−x)²+10(y−x²)² forms a narrow curved valley—like many neural network loss surfaces. It shows how optimizer and learning rate choices can diverge or crawl slowly toward the minimum at (1,1).

Question 3

How do I change the start point?

Accepted Answer

Click the loss canvas to move the white start marker; all three paths reset from there. Switching terrain presets resets to that terrain's default start.

Question 4

Why does the saddle terrain have two minima?

Accepted Answer

f(x,y)=(x²−1)²+y² is a double well: a saddle at (0,0) and minima at (-1,0) and (1,0). Compare which valley each optimizer enters from your start point and how they cross the saddle.

Loss Landscape Lab

Loss terrain

Optimization paths

Loss Landscape Lab

Loss terrain

Optimization paths