09.2 First, we get the values

Content and key concepts

On the previous page, let's draw the calculation relationships.
We're going in the order of addictions: you can't calculate the sum before the sum is added. The correct values don't prove the derivatives are correct.
And we're going to save the behavior and we'll go back to the error, and the deposits of the different paths to the total weight add up. It's an automatic differentiation, not a new way to guess the direction.
Для изменения рисунка включитеJavaScript; расчёт и теория доступны далее.
Values are not yet derivatives
Forward follows the topological list. In the example a=2x3=6,z=6+1=7,L=7x7=49.Leaves x,w,b remain the same, optimizer is not called.
Host
Parents
Operation
Value
a
x,w
x×w
6
z
a,b
a+b
7
L
z,z
z×z
49
"Dale" shows one named node. Future values are not marked as past. A clean computing core can pre-form the deterministic snapshots, but UI shows the state of the selected frame, not mixes it with the final answer.
Experiment.
Replace x by 0:a=0,z=1,L=1. Then b/0:L=0. After the change of JSON, the navigator and the backway are dropped; the old result is not given as relevant.
Boundary.
The coincidence of the loss does not prove that the backward is correct. You can compute the correct forward and lose the contribution of the common node when disferentiation. Check derivatives are needed separately. If you change leaf.data between forwards and backwards without count, the saved local derivatives will become obsolete. The training editor reschedules the graph for a new entry.
You can lose the accumulation with the right forward.
How does this matter continue?
Python examples of Chapter 09
Academic base