A model has a single parameter , and its loss is
Two training runs are launched from different initial settings:
| Run | Starting setting |
|---|---|
| A | |
| B |
Each run behaves like a marble released on the surface, in thick fog. It can see nothing beyond its own feet: it moves continuously in whichever direction makes decrease, and it halts the instant the ground beneath it goes flat. It has no momentum, so it cannot climb over anything, and it never jumps.
The two runs use the same loss and the same rule. Only their starting points differ.
How much higher is the final loss of the worse run than the final loss of the better run?
Report the gap (worse minus better) in loss units. The arithmetic comes out exact, so no rounding is required.