After one exact parameter update, the same toy example can be run forward again with only w11 changed. This lesson recomputes the immediate arithmetic effect without making a training or convergence claim.

highlighted = computed this step

Use the updated weight

Now keep the same input and change only w11 to 3/2. The first hidden weighted sum uses x1=1, x2=2, unchanged w12=1, and b1=-1.

z1new=3/21  +  12    1z_{1}^{new}=3/2\cdot1\;+\;1\cdot2\;-\;1
Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.forward pass (all exact)quantityrulevaluez13/2*1 + 1*2 - 15/2h1ReLU(z1)5/2z21*1 - 1*2 + 0-1h2ReLU(z2)0yhat1*5/2 + 1*0 + 05/2L(yhat - 3)^21/4

Recompute the hidden activation

The first hidden pre-activation is 3/2*1 + 1*2 - 1 = 5/2. Since that is positive, ReLU stays open and h1 becomes 5/2. The second hidden activation remains 0.

z1new=5/2,h1new=5/2z_{1}^{new}=5/2,\quad h_{1}^{new}=5/2
Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.forward pass (all exact)quantityrulevaluez13/2*1 + 1*2 - 15/2h1ReLU(z1)5/2z21*1 - 1*2 + 0-1h2ReLU(z2)0yhat1*5/2 + 1*0 + 05/2L(yhat - 3)^21/4

Recompute output and loss

With h1=5/2 and h2=0, the output is yhat=5/2. The target is y=3, so the same-example loss is (5/2 - 3) squared = 1/4.

Lnew=(5/2    3)2=1/4L^{new}=(5/2\;-\;3)^{2}=1/4
Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.forward pass (all exact)quantityrulevaluez13/2*1 + 1*2 - 15/2h1ReLU(z1)5/2z21*1 - 1*2 + 0-1h2ReLU(z2)0yhat1*5/2 + 1*0 + 05/2L(yhat - 3)^21/4

Compare on the same example

The old loss on this same row was 1. After changing only w11, the recomputed loss is 1/4. This is the immediate arithmetic effect on the same example only: not training, not convergence, not learning, and no generalization claim is made.

Lold=1,Lnew=1/4L^{old}=1,\quad L^{new}=1/4
Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.Check the Updated Forward PassSame example, only w11 changed before recomputing the forward pass.forward pass (all exact)quantityrulevaluez13/2*1 + 1*2 - 15/2h1ReLU(z1)5/2z21*1 - 1*2 + 0-1h2ReLU(z2)0yhat1*5/2 + 1*0 + 05/2L(yhat - 3)^21/4