0% found this document useful (0 votes)
6 views21 pages

Backpropagation Weight Update Process

The document outlines a neural network training process, detailing the forward and backward pass calculations for weight updates. It includes specific values for weights, biases, and outputs, as well as the calculation of total error and its minimization through gradient descent. The final results show an improvement in error after multiple iterations of backpropagation.

Uploaded by

Aditya Deore
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
6 views21 pages

Backpropagation Weight Update Process

The document outlines a neural network training process, detailing the forward and backward pass calculations for weight updates. It includes specific values for weights, biases, and outputs, as well as the calculation of total error and its minimization through gradient descent. The final results show an improvement in error after multiple iterations of backpropagation.

Uploaded by

Aditya Deore
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

ê

Given - x =0.05
x 2 = 0·10

W4
W1 =0.15, W2 = 0.2, W3 = 0.25, W1
W5=0·4, W6= 0.45; W7=0.5

C Forward Pass
-;) Shi
Outh, =

120
43 20.25
the W2=012

W 0.55
Sha
11
1
1+e-03775
=

WG=0.47

6 Out b2
8

5055 Mi
;) Soi
Expected Output
01 = 0·01, 02 =
0.99 = 0.3, b, =
0.35,0
)

, Wg=0.55, b2 = 0.6

w1 x + W2x2 +b,
=

= 0.15 × 0.05 + 0.2 X 0.10 +0.35 =


0.3775

0.5932

Wz x + Wy x 2 + bj = 0.25 × 0.05 +0.3×0·1


+0.35
0.3925
1

Ite
0.3925
= 0·59689

W5, +wo h2 + b2 = 0.4x0.5933 +


0.45x
1.1059
0.5968970x

Outo
1
=

èl·1059
07514

14) So2 = with, +ligh 2 + b2 =


0.5x0.5933+0·55x0·59689
='1'
112250
+0.6

Out 02
S
1+1.225
Calculating total error
-
E
tota
l
EO,
Eoi
E02

E02
=
=

= 0·7729

ди
= '/1⁄2
Squared Frror fu

(target - putput)"
1⁄2 (target 0,- Out o)2 =
2 (0.01 -0.7514)2
2

= 0.2748

E total = Eoi +E02


=

1 1⁄2 (target 02 - Out o2)2 =


1/2 (0.99-0.772
0.02356
1.
= 0.2748 +0.02356 . Etotal =
0.298
Backward pass
Goal is to update each weight in
My
so that they cause actual off to be closer to
target off
error for each off neuron.
thereby minimizing i) so ws, we want to know how
much change in w
I Exokal (ive of Exotal [Link] partial derivat affects the total
error de

(this is
Gradient w.r.t. Ws)
ο ως
W5)

By applying chain rule, we know


that:
703

Exotal
Etotal
*
d out o
*

awp
dout of
Iso,
2W5
O
Now, need to figure out each piece in this equat.
First, how much does the total error change
with
respect to the Olp ?
Estota
l

1⁄2 (target - Out,)2 +
1⁄2 (target -Out (2)2
)
*

I Etoral £0, 2 + 1 (target-Outo,)


2 =" + -1 +0 constant 2 Outo = -
(target - Outo1)
= - (target - Outo1) =
-(0.01 -0.7514)
A 0.7414
ilp
Next, how much does op of O, change
w.r.t its `/p
*. SOT.

The partial derivative of the sigmoid


this ex
Outo,
1
Ité So,
Doutor

So,
Outo, (1 - Outo1) =
0.7314(-0.7514)
1

= 0.1868

Finally. How much does total i/p of 0,


change wirtus
two
6

So1 = wsouth, + wg Out h2 +


b2 * @ So1 = 1x Outh, x
w(i-1)
Jws
=

1x Outh, x W{ > +0 +0 =
Outh,
0.5933
20:0

Futting it all together


in eqn
E total
aws
0.7414 * 0.18680-59336
=0.0822

Now, to decrease the error, we subtract this value


from the current weight (multiplied with learning
rate = 0.5 consider)
Ws - n d Exotal
w5
1

or

Ws
new
= ws = 2 *
ws

Ws
ne
new

θως
=0·4- 0.5× 0.0822
= 0·3589

with repeating above steps to get we, w,


& Wg: -
WE new =
W7 new
W8 new =
news
= 0.4087

0.5113

0.5614

ii) Hidden
Layer
1

2861

we continew backpropagat for


these
with original weights & all are updated
after getting all new
weights!
Now we calculate new values for w,, w2, &
W3 &W,
I E Letal
I
Estotal
*

d out hi
a Shi
*
2

dw,
ωι

Jouth,
Here we use similar process but
slightly diff has accout the fact that the
Olp of each hidden layer neuron contributes
to the opp of multiple olp neurons. We
know that Outb. affects both Out
& Out 02 Therefore the
Etotal need to take into
consideration its
Out b
effect on the both off neurons:
Etotal 2 out hi
1

DEOL +
Jouth,
d to2
Out he
Jouth,
starting with 1st part of
equation
JEO, aso,
*
Jouth,
2

we can calculate
EOL
Eor
using
values we

calculated
earli
slova Eo,
aso,
2
* Outo,
0.7416 × 0.1868
=

Iso,
Outo
a 30,
0.1385

05:0
And 50, is equal to
ws:
= Ws "Outh, + wo⋅ Out b2 + b2
*
до
Out),
So,
aso1 =
Douthi
w6"
280

Ws to to
+0+05
0.45

804-0

Putting them in eq" 04

Outhe

Followi
ng
same process for
@EO2,

egn
Eo
= 0.1385 × 0.4
= 0.0554
JEO2, we get! Jout 12

E total
JEO,
+

@
Outh,
2 E02
a
Outbr

2 Out 2
= −0·0191

2 Out hi
0.0554 + -0.0191 = 0.0364
Now, that we have IE total
20uth,
&
the a
20uth,
s
ashi
2 Outh,
J Shi
J

, we need to find
Two for each weight
cut the
Outh =
He-sh,

Out (1- Outh,) = 0.59


33 (1-0.5933)
h1
= 0.2413
We calculate the partial
derivative of total silp to "I w.r.t. W,
the same way we did for the off news,
= W1.x, + Wz⋅ x2 +
b1
Shi asn
B

2, +0 +0 = 0.05

it all together
Dw.
Puttin
g
I Etotal
Etotal x
2 Outh, x
asni

dash, dw,
деб
до
1

op

=X

Out hi
0.0364 0.2413 * 0.05
= 0.00044

We can now update w:


W1 new
= w1 - n. a Etotal
=
1

0.1498
| W
= 0.15-0.5 × 0.00044
W3
Repeating this for W2, wz &
w4=
W2
new
0.1996

Ws new
=

0·2498

8.2995
W4 new
=

we update all
weights! when we
ard the 0.05 & 0
Now, we
forward
On
the new was
fed
i/p originally, the erro
was 0.29837! After applying
this 1st round of backpropagat error is
now down to 0.29103.. But after
repeating proces 1000's time we get the
opp as expected.

You might also like