0% found this document useful (0 votes)
214 views26 pages

Perceptron Network

The document discusses perceptron networks, including: 1) The perceptron learning rule is more powerful than the Hebb rule and uses an iterative weight adjustment process. 2) The original perceptron has three layers - sensory, associator, and response units. 3) A single layer perceptron consists of a single neuron and is limited to performing pattern classification with only two linearly separable classes. The training continues until no errors occur.

Uploaded by

John
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)
214 views26 pages

Perceptron Network

The document discusses perceptron networks, including: 1) The perceptron learning rule is more powerful than the Hebb rule and uses an iterative weight adjustment process. 2) The original perceptron has three layers - sensory, associator, and response units. 3) A single layer perceptron consists of a single neuron and is limited to performing pattern classification with only two linearly separable classes. The training continues until no errors occur.

Uploaded by

John
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
You are on page 1/ 26

c

How the perceptron learning rule is


better than the Hebb rule.
Layer structure in the original
perceptrons.
Learning and training algorithms in
the perceptron network
Architecture, algorithm and the
application procedure of the
perceptron net.
Derivation of perceptron algorithm
for several outputclasses

Perceptron • Applications of multilayer


perceptrons.

Networks
4.1 Introduction
119621,and Minsky and Papert [19881,developed large class of artificial neural
Frank Rosenblatt
Perceptrons. 'Ibe perceptron learning rule uses an iterative weight adjustment that is
networks called
than the Hebb rule. The perceptrons use threshold output function and the McCulloch-
more powerful
neuron. Their iterativelearningconvergesto correctweighty,i.e. the weightsthat
Pitts model of a output value for the training input pattern. The original
perceptron is found to have
produce the exact associator and response units as shown in Fig. 4.1.
sensory,
three layers,
61
l ien rptron Networks

sensory Associator Response


Unit Unit Unit

Original Perceptron

Thesensory and association units have binary activations and an activation of + I , 0 or— I iQug,edfor
theresponseunit. All the units have their con•espondingweighted interconnections.Training in rrrceptron
u.illcontinueuntil no error occurs. ms net solves the problem and is also used to learn the classifica-
non.The perceptrons are of two types: single layer and multi layer perceptrons. A detailed study about
thepcrccptronnetworks is made in this chapter.

4.2 Single Layer Perceptron


patterns
A single layer perceptron is the simplest form of a neural network used for the classification of
and
thatare linearly separable. Fundamentally, it consists of a single neuron with adjustable weights
linearly
bias.Rosenblattfound that if the patterns used to train the perceptronare drawn from two
form of
separableclasses, the perceptron algorithm convergesand positionsthe decision surface in the
limited to perform-
a hyperplanebetween the two classes. The perceptronbuilt around a single neuron is
for the per-
ingpatternclassification with only two classes. Also classes have to be linearly separable
ceptionto work properly.
is that, it is concerned
The basic concept of a single layer perceptron as used in pattern classification
the perceptron network very
mth only a single neuron. The linearity and the integrity learning makes
simple.Training in the perceptron continues till no error occurs.

4.2.1 Architecture
Fig. 4.2.
Thearchitectureof the single layer perceptron is shown in
sensory, associator and response units. The input to
As we have already studied, the perceptron has
unit, which is a binary vector. Since only the
the responseunit will be the output from the associator
adjusted, the concept is limited to single layer
weightbetween the associator and the response unit is
networkas discussed in Section 2.7.
unit and the response unit is shown. The
In the architecture shown in fig. 4.2, only the associator
response unit are
sensorunit is hidden,because only the weights between the associator and the
n. There always exists a com-
adjusted.The input layer consists of input neurons from Xl Xi x
through weighted intercon-
monbias of 'l'. The input neurons are connected to the output neurons
interconnectionsbetween
nections.This is a single layer network because it has only one layer of
input signal received and performs the
the input and the output neurons. This network perceives the
classification.
62 to

Fig. 4.2 Architectureof Single Layer Perceptron

4.2.2 Algorithm
To start the training process, initially the weights and the bias are set to zero. The initial weights of the
network can be formulated from other techniqueslike Fuzzy systems, Genetic Algorithm etc. It is also
essential to set the learning rate parameter, which ranges between 0 to 1. Then the input is presented, the
net input is calculated by multiplying the weights with the inputs and adding the result with the bias
entity. Once the net input is calculated, by applying the activation function the output of the network is
also obtained. This output is compared with the target, where if any difference occurs, we go in for
weight updation based on perceptron learning rule, else the network training is stopped. The algorithm
can be used for both binary and bipolar input vectors. It uses a bipolar target with fixed threshold and
adjustable bias.
The training algorithm is as follows:
Step 1: Initialize weights and bias (initially it can be zero). Set learning rate to 1).
Step 2: While stopping condition is false do Steps 3—7.
Step 3: For each training pair s:t do Steps 4—6.
Step 4: Set activations of input units.
Xi= sj for i = 1 ton
Step 5: Compute the output unit response.

The activation function used is,


1, if Yin > 0
Y = f(Yin) = 0, if —0 S y_in SO
-1, if
63
6: wcight« and biae ate updated
if the target iq not
If t y and the value of X,ig not eqtjnj to 'he rew.nse.
'em,

el.se

Anew) W'(old)
b (new)=
(old)'
Step 7: Test for stopping condition.
stopping conditions may be the weight
changes.
vote:
1. Only weights connecting active input units (Xi#0)
are updated.
2. Weights are updated only for patterns that do not
produce the correct value of

4.2.3 Application Procedure


This procedure enables the user to test the network performance.The network should be trained with
sufficient number of training data and using the testing data its performance can be tested. The applica-
tion procedure used for testing perceptron network is as follows.
Step 1: The weights to be used here are taken from the training algorithm.
Step 2: For each input vector x to be classified do Steps 3 4.
Step 3: Input units activations are set.
Step 4: Calculate the response of output unit,

0
Y-in >
0 If —0 < y_in SO
-1, If Y-in O

4.2.4 Perception Algorithm for SeveralOutput Classes


The perceptron network for single output class is extended for several output classes. Here there exist
more number of output neurons, but the weight updation in this case also is based on the perceptron
learning rule. The algorithm is as follows:
Step 1: Initialize the weights and biases. Set the learning rate
Step 2: When stopping condition is false, perform Steps 3 7.
Step 3: For each input training pair, do Steps 4 6.
Step 4: Set activation for the input units,
Xi= Sifor i = 1 ton
94
forj = I to
output of each output Y ini¯ m.
Step S: Compute the

O. if
l. if
and bias atv•to updated for j e I to m and i 1 to n.
Step 6:
It N, t, and v, 0. then

b(old) at,
cisc If
= wv(old)

Ihat IS.the and weights remain unchanged.


Step 7: Te€t for stopping condition.
The •topping condition may be the weight changes.

Example 4.1 Develop a perceptron for the AND function with bipolar inputs and targets.
Solution The training pattern for AND function can be,

Input Target
b t
1 1 1 1

-1 1 1 -1
1 -1 1 -1
-1 -1 1 -1
Step 1: Initial weights WI= = 0 and b = 0, a = 1, 0 = 0.
Step 2: Begin computation.
Step 3: For input pair (l, l): l, do Steps 4—6
Step 4: Set activauons of input units

Step 5: Calculate the net input.


m = b + E Xi Wi
Applying the activation,
l, if
Y-in
if -OSy_in SO
-1, if
Therefore y = O.
65
6: and V
Kinee t y. the new weightq
Rte.

ncw wctghts and bias arc (l I i).


algorithmic steps are rcpcatcd for
all the input Vectorswith
calculatedwctghts. their initial wetghtg the re•nougly
By prcscnting all the input vectors, the
updated weights are shown in table
telov:
Input Net Output Target Weight Changes Weight'
AWI AW2 Ab
(0

ms completes one epoch of the training.


The final weights after the first epoch is completed are, WI= 1, w2 = 1, b = I
Weknow that b + + =0

(-1)

x2= — + 1 is the separating line equation.


The decision boundary for AND function trained by
perceptron network is given as,
In a similar way, the perceptron network can be developed for logic functions OR, NOT, AND NOT etc.

Example 4.2 Develop a perceptron for the AND function with binary inputs and bipolar targets without
bias up to 2 epochs. (Take first with (0,()) and next without (0,0)).
Solution Initializing the weights to be, WI w2 = 0 and the bias is neglected here (because the problem is
stated without bias). Hence a I and threshold 0 = O.
(a) With(0,0) and without bias.
Intnkiuctionto ScunJl Nctuorks

net input

O, if
if y _in < — O
weight change,
Awi = and
New weight is.
(new) = w(old) + Aw
Epoch 1:
Input Net Output Target Weight Changes Weights
AWI
(0 0)

The separating line for 1st and 2nd input are, Xl+ x2 = 0 and x2 = 0 respectively.
Epoch 2:
The initial weights used are the final weights from the pervious iteration.

Input Net Output Target Weight Changes Weights


AWI AW2
(0 0)

Without bias for the given inputs, the final weights obtained are same as that for with bias and the equation
of separating line also remains same. Thus the equations remain same and are given by,
for 1st input
for 2nd input
(b) Without bias and (0,0)
Epoch 1:
Input Net Output Target Weight Changes Weights

(0 0)

are (0,0) and the separating line are x2 =


also the final weights and x2 = 0.
In this case,
Perceptron Networks 67

Epoch 2:
The final weights from Epoch I are used here as initial weights,
Input Net Output Ilu•get Weight Changes Weights
AWI

(0 0)

Here also, the weights are same as that of previous epoch.


The separating line here also, without bias
I st input

2nd input

Thus from all this, it is clear that without bias the convergence does not occur. Even after neglecting (0, 0),
the convergence does not occur.

Example 4.3 Using the perceptron learning rule, find the weights required to perform the following clas-
sifications. Vectors (1 1 1 1), (—1 1 —1 —1) and (1 —1 —1 1) are members of class (having target value l);
vectors (1 1 1 —1) and (1 —1 —1 1) are not members of class (having target value —1). Use learning rate of 1
and starting weights of 0. Using each of the training and vectors as input, test the response of the net.
Solution The initial weights are assumed to be zero and the learning rate as 1.
The updation is done according to perceptron learning rule,
If y t, weight change, Aw = octXi& Ab = Oct.

New weights are, (new)= W(old)+ Aw


(new) (old)
If t = y, no weight change
By using the above, the below tabulation is formed, where, y_ in= b +

and y = f(Yin)is the activation applied.


to Neurol

Input Output Weight Chnnges


Weights
AWA Ab

-1
Epoch 2: initial --9
00020
00020
Initial —i
-1
-1
The final weights from Epoch 1 are used as the initial weights for Epoch 2. Thus the output is equaltotarget
by training for suitable weights.
Testing the response of the net
The final weights are,
For the 1st set of input, WI= 0, w2 = 0, w3 = 0, w4 = 2, b = 0, and
For the 2nd set of input, WI =— 1, 1 w3 = —1, w4 = -- 1, b = 1

The net input is, Yin= b + Ex


For the 1st set of inputs,
(i) (1 111 1)
Enl
Applying activation,

Applying activation, Y2= f(y-jn2) =


For 2nd set of inputs,
-11)
Applying activation,
(ii) (1 -111)
Y—in2

Applying activations,
69
calculated teq output matcheq with tArget the Input

4.4 Fot the following noiQYvet•ionq of


training patternq. identify the reqpr.n•.eof network hy
it into comet, incotn•ct ot indefinite.
(0-1 1). (0 1-1). (001). (0 0-1). (0 10). 01).
0-1), (1 -1 (1 00). (1 1 0), (0-1 0). (l l)
So/gtton concept fot this pmblcm is
If + > 0, then the response is correct.

If + + < 0, then the response is incorrect.


If + x2w2 + x3w3 = 0, then the response is indefinite or undetermined.
Say if the weights taken from the bipolar step function are, WI= O, w: 2. w, 2
For (0—1 1), Xl = 0, x2=— 1, x3
+ x2w2+ x3w3= 0 + 2 + 2 = 4 > 0, so the response is correct
Vector Resp«nse
+ X2W2+ X3W3
(0 0) Correct
(0 Correct
(0
(0
Correct
(0 -1)
(0 0)
Correct

0) Correct
(1
Undetemined
(1 0)
(1 0)
Undetermined

bipolar inputs and


MATLAB program for perceptron net foranAND function with
Example 4.5 Write a
targets.
is given as
Solution The truth table for the AND function

is given as follows.
The MATLAB program for the above table
70 Introduction to Neural Networks
6.
Program
*Perceptron for AND funtion
clear:
clc:

alpha-input( 'Enter Learning rate-' ) :


theta-input( 'Enter Threshold val ue=' ) :
con-I :
epoch-O :
while con
con=0:
for
if yin>theta
end
if yin <=theta &yin>--theta

if yin<-theta
end
if y-t(i)
con-I;
for J-1:2

end
end
epoch-epoch+l :
end
disp( 'Perceptron for ANDfunt)on•)
disp(' Final Wenght
matru').
disp(w):
dl€p( •FInal Bias').
disp(b):
Output
Enter Learmrtgrate-I
Enter Threshold value-C.S
Perceptron for ANDtunvon
F Ina 1 Weight n;atmx
1 1

Final Elas
Perr eptrrpnNetworks

4.6 Write a MATLAB pmgram to recognize


the number 0, l. 2, A 5/3 matrix forms the
For any valid point it is takcn as I and invalid point it is
and when the test data taken O.The net has to be trained to
all the nutnbcrs is given. thc network has to recognize the particular nurn-

The numbers are fonncd from the 5 x 3 matrix and thc input
Olgtton data file is determined. The input data
are given. The data are stored
and the test data files in a file called 'reg.mat% When the test data is given.
recognized then it is + l, and if the pattern is not recognized,
ifthcpatternis it is L
pgt* - reg.mat
01 1 11 1 1;

1100111011;
0100000000;
1011100111;

1010001010;
0100000000;

0100000000;
0010000000;
0001000000;
0000100000;
0000010000;
0000001000;
0000000100;
0000000010;
0000000001;]

11001;
Neural

01001.

10111;
100;
0101 0;
10011;

cleer:

nať ) :
IȚut•Icd.A' :cd.B•:cd.C'
:cd.D' :cd.E' :cd.F' :cd.G' ;cd.H' :cd. I
' :cd.0' ' •
lor

output(i

end

for J- l: 2

eise

test-tcd.K' :cd.Ľ
net-newp(aw. 10, ' M';cd.N'
hardl im' ) : :cd.0' •
net. tra i npararn.
net. epochs-1000 •
trainparam.goal-0 •
net-trai net. i nput.
output) ;
Y-sim(net.test);
Perceptron Networks 73

1-0.

end
end

s=sprintf( 'Test Pattern is Recognisedas Xd' , i , 1-1):


disp(s);
else
s=sprintf( 'Test Pattern is NotRecognised', i):
disp(s):
end
end

Output
TRAINC, Epoch 0/1000
TRAINC. Epoch 25/1000
TRAINC. Epoch 50/1000
TRAINC. Epoch 54/1000
TRAINC. Performance goal met.
Test Pattern 1 is Recognisedas 0
Test Pattern 2 is Not Recognised
Test Pattern 3 is Recognisedas 2
Test Pattern 4 is Recognisedas 3
Test Pattern 5 is Recognisedas 4
1 Performanceis 0, Goal is O

0.9
0.8
8 0.7
0.6
8 0.5
2 0.4
0.3
02
0.1

o 5 10 15 20 25 30 35 40 45 50
54 Epochs
TrainingCurve
Intnxiuction to Neural

Example 4.' With a suitable cxantple demonstrate the pcrccptron learning law with its d
using MATLAB. Give the output in graphical form ecision
regions
Solution lhe following example demonstrates the pcrccptron learning law.
Program
clear
p S: dimensionality of the augmentedinput space
N SO. numberof travvng patterns - size of the training epoch
PARTI • Generationof the training and validation sets.

nn round(

wht wht = wht/norm(wht);


wtt
(wht*X 0):

XP= X(pr.
Z projection of the weightvector
co = cl = find(D==1):
Z co and cl are vectors of pointers to input patterns X
Z belonging to the class 0 or 1, respectively.
figure(l). cIf reset
'o', )(p(l,cl), Xp(2,cl), 'x')
Z The input patterns are plotted on the selected projection
Z plane. Patterns belongingto the class 0, or 1 are marked
Z with 'o' or 'x' respectively
axis(axis), hold on
Z The axes and the contentsof the current plot are frozen
Z Superimposition of the projection of the separation
plane on the
Z plot. Theprojection is a straight line. Fourpoints
lying onthis
Z line are foundfromthe line equation wp. x -0

S= +wp(3))•
LJ. [L grid, drawnow
% PART 2: Learning
eta = 0.5; %The training gain.
wh = 2*rand(1,p)-1:
Pemeptron Networks 75
Randominitialisation of the weight vector with
t fromthe range ( •1. •1) An exampleot an initial
value;
vector fol lows
projectlon of the initial decision planewhichis
orthogonal
to wh is plotted as previously
- projection of the weightvector

) L). (L S(2.:))). grid on. drawnow


t Maximum numberof training epochs
[C+I. zeros(1.C)): X Initialization of the vector of the total stunsof squarederrors over an

X The matrix WWwill store all weight


vector whoneweight vector per rowof the matrix WW
Xc is an epoch counter
cw total counter of weight updates

L). [L S(2.:)). drawnow


for n =
eps = D(n) - , n)) >= 0); X eps(n) = d(n) - y(n)
wh = wh+ eta*eps*X(: . n)' : X The Perceptron Learning Law
cw = cw + l;
Wk(cw. : ) = wh/norm(wh):X The updatedand normalised weight vector is stored in for
featureplotting

= E(c) + abs(eps) X lepsl = epsA2


end:
wp= wh([pr X projection of the weight vector
S= +wp(3)) .
L]. [L SC. drawnow
end :
Z After every pass throughthe set of training patterns the projection of the current decision
Planewhich is determinedby the current weight vector is plotted after the previous projection has
beenerased.
"(l:cw, pr);

wht-
-0.4078 0.8716 -0.0416 0.2684 0.0126

10 6 6 4 6 3 4 4 4 2 o o
Nctuv»rk.s
Introductionto Neural
76

0.4

0.2

0.2

0.4 oo o
-0.6 O

-0.8

0.5 1
-0.5

Output for Perceptron Leaming Law Example

Example 4.8 With a suitable example simulate the perceptron learning network and
anes. Plot the points assumed in the respective quadrants using different symbols for identification.
Solution Plot the elements as square in the first quadrant, as star in the second quadrant, as
third quadrant, as circle in the fourth quadrant. Based on the learning rule draw the decision boundaries.
Program

p2=t121' . Z- class 1. first quadrantwhenweplot the elements, square


-11' . p4-t2 -21' , Z- class 2, 4th quadrantwhenweplot the elements. circle
21• p6=t-2 1]' Z- class 3, 2ndquadrant whenwe plot the elements. star
p7=t-1 -l l' . pact-2 -21' - class 4. 3rd quadrantwhenwe plot
the elements.diamond
ZNow.lets plot the vectors
hold on
.P1(2) . .P2(1) , 'ks' .p3(2) . .p4(2) . 'ko s )
.P6(2) 'k*'
.p8(1) .p8(2)
hold
3 -3 31)Zset nice ans onthe figure
Z- class l. first quadrant
1J' ; 1)' Z- class 2. 4th quadrant whenweplot the elements. square
01'; t6-C1 . Z- class 3, 2nd whenweplot the elements.circle
quadrantwhenweplot the
1]' ; 11' •.Z- class 4. 3rd elements, star
quadrantwhenweplot
nets simulate perceptronlearning the elements.diamond
Perceptron Networks 77

netp-newp(R.2): tnetp is perceptron network


with 2 neurons and 2 nodes. hardlimit transfer
function. perceptronrule learning
*Definethe input matrix and target matrix
pg[pl p? p3 p4 p5 p6 p? p8):

yesnm(netp.P) twell. that is obvioulsy not good. Y


train is not equal P
let's
'*Now.
netp.trainParam.epochs20: X let's train for 20 epochs
netp train(netp.P.T): *train.
seemsthat the training is finished after 3 epochs
andgoal is met. Lets check by simulation
yl-sim(netp.P)
is the sameas target vector. so our networkis trained
weightsandbiases after training
Otthe
w=netp.IÅ{I. *weights
*bias
%decisonboundaries are lines perepndicular to weights
{ke assumehere that input vector y]'
Xboundary generated by neuron 1
Xboundary generated by neuron 2
net's plot input patterns with decision boundaries
fi gure
hold on
, 'ks' .p2(1) .p2(2) . 'ks' .p3(1) . 'ko' .p4(1) .p4(2) . 'ko' )
.p5(2) . 'k*' , 'k*' , 'kd' .p8(1) .p8(2) . 'kd' )

3 -3 nice axis onthe figure


plot(x.y. 'r' 'b' )Xhereweplot boundaries
hold off
Z SEPARATE BOUNDARIES
Zadditional data to set decision boundariesto separate quadrants
p9=t10.05]' : 1]' ;

-0.05]' : p12-[O.05 :
til-t3:t12-t4:

P13-[-10.05]' 1]' :
t 13-t5:t14-t6:

-0.05)' :

netp=newp(R.2. 'hardl imi . 'learnp' ) :


XDefinethe input matrix an target matrix
P-[pl P2 P3 P4 P5 p6 p7 p8 p9 P12 P13 p14 P15 P16]:
•(1.2). tboundarygenerated by neuron 1
tboundarygenerated by neuron 2

. , .p€(l) .p6(2). 'k*' .p7(1) . 'kd' .p8(1) .p8(2) . 'kd' )

. .P14(1) .p14(2) . 'k*' . 'kd' .p16(1) .p16(2) . 'kd' )

c 3 -3 3) •set nice axis on the figure


'r y.yl. •b' )Xhereweplot boundaries

Cuerent v lot released

Epoch 0/20
3/20
W INC. Per formance goal met

TRAINC, Epoch 0/5000


Perceptron Networks 79
TRAINC.Epoch 50/5000
TRAINC,Epoch 75/5000
TRAINC.Epoch 92/5000
TRAINC. Performance goal met.

-20 . 0000 -1.0000


-1.0000 -20 . 0000

Figure No. 1
fue kdÜ Ylew Insert 1001s %ndow Help
loci: Endo.

Plot of the Given Vectors Plot of Input Pattems with Decision Boundaries
Figure No •r With T kAl NC
tue »ew loots
PeHormmc:2
is O,God is O

e 0.6

pet Boo•üares
to Neurol

a MAII„AB fov pattcrn classification using perceptron network.


Etgrttplc 4.0 NVntc
pattctll. the input vectors and noisy vectors from patterns as shown below st
the net unth a notsy train
it in a •mat file.

-1-1-1 -1-11 -11-1 -111

111
Input vectors

Noisy vectors

Solution The input vectors and the noisy vectors are stored ina mat file, say class.mat, and the requireddata
is taken from the file. Here a subfunction called charplot.m is used. The MATLAB program for this is given
below.
Program
ZPerceptronfor patternclassification
clear:
clc•.
ZGet the data from file
data-open( 'class .mat' ) ;
x-data . s: Zinput pattern
t-data . t: ZTarget
ts-data. ts ; ZTesting pattern
n-15:

Zlnitialize the Weightmatrix

Xlntitalize learning rate andthresholdvalue


al pha=l:
Perceptron Networks 81

theta-0:
tP10t for Input Pattern
f' gure( I ) :

end
end
55 0 25]):
title( ' Input Pattern for Training') ;
con=l ;
epoch=0 :
while con
con=0:
for
for j=l:m
for i=l:n

if yin(j)>theta
end
if yin(j) <=theta
&

if yin(j)<-theta

if y(l.
w-w:b-b:

con-l•
for j-l:m
for i-l:n
end

end
82 Introduction to Neural Networks

epoch-epoch*
l,
end
(hsp( 'Numberof Epochs: ) ;
disp(epoch):
tTest1nqthe networkwith test pattern
XPlot for test pattern
fngure(2)•.

for J-1:4

end
end
55 0 251);
title( 'Noisy Input Pattern for Testing' ) ;
for 1-1:8
for

for

end
if yin(j)>theta
end
if yin(j) <=theta
&
end
if yin(j)<-theta
end

for i-1:8

end

end
XPlot for test outputpattern
figure(3);

for i-1:2
for J-1•.4
83
Perceptron Networks

k-k+l:
end
end
55 0 25));
fied OutputPattern' ) :
title( 'Classi
Subprogram used :
.ys. row.col )
functioncharplot(x.xs

for VI : row
for

end
end
for
for

plot(j+xs-l.ys-i+l.'r'):
hold on
else

hold on
end
end
end

Output

Numberof Epochs

Pattem Used Tra


Neocorks
Introduction to Neural
Patt•m testing
toowy

10

Training
Noisy Input Pattern used for
Pitn

15

10

to

Classified Output Pattem

4.3 Brief Introduction to Multilayer Perceptron Networks


ultilayer perceptron networks is an important class of neural networks. The network consists Ofa set
f sensory units that constitute the input layer and one or more hidden layer Of computation modes.The
nput signal passes through the networkin the forward direction. The network of this type is called
ultilayer perceptron (MLP).
The multilayer perceptronsare used with supervisedlearning and have led to the successful
ropagation algorithm. The disadvantageof the single layer perceptron is that it cannot be extendedto
ulti-layered version. In MLP networks there exists a non-linear activation function. The widely
on-linear activation function is logistic sigmoid function. The MLP network also has various layers
idden neurons. The hidden neuronsmake the MLP network active for highly complex tasks. layers
the network are connected by synaptic weights.The MLP thus has a
high computational
85
Perceptron Networks
A disadvantageof MLP may also be the presence of non-linearityand complex connections of the
networkwhich leads to highly complex theoretical analysis. Also the existence of hidden neurons makes
process tedious.
thelearning
The Ml-p networks are usually fully connected networks. There are various multilayer preceptron
networkswhich includes Back propagation network, Radial basis function network etc. These are dealt
forthcoming chapters.
in detailin the

Theperceptrontraining algorithm discussed above can be implemented on a digital computer or on any


electronichardware and the network becomes, in a sense self-adjusting. Rosenblatt's proof has been a
majormilestone and has a great importance in the field of neural network. The single layer perceptron
networkwas discussed with its architecturealgorithm,examples etc. An overview of multi-layer
perceptronsis also given in this chapter. It should be noted that there is no proof that the perceptron
trainingalgorithmis faster than simply trying all possibleweight adjustments,but in some cases, the
perceptronnetwork provides a superior result.

ReviewQuestions
4.1 What are the three layers in the original perceptron?
4.2 Briefly discuss on the learning rule of a perception network.
4.3 Explain the working of the perceptron net.
4.4 What is the activation function used in the perception network?
4.5 Explain the architecture of the perceptron net used for pattern classification.
4.6 State the perceptron learning rule convergence theorem.
4.7 Explain the algorithm used for training the perceptron net.
4.8 What are the stopping conditions used to stop the progress of the training algorithm?
4.9 How is perceptron net used in the aspect of linear separability?
4.10 Compare perceptron and Hebb net.
4.11 State the application algorithm used in perceptron net.
4.12 Write the perceptron training algorithm for several output classes.
4.13 Give a brief note on multilayer perceptrons.

Exercise Problems
4.14 Form a perceptronnet for OR function with binary input and output. Compare it with the
results obtained using bipolar input and targets.

You might also like