Classification of Iris data set

Mentor:
Assist. prof. Primo Potonik
Student:
Vitaly Borovinskiy

Ljubljana, 2009
1. Problem statement

Fishers Iris data base (Fisher, 1936) is perhaps the best known
database to be found in the pattern recognition literature. The data
set contains 3 classes of 50 instances each, where each class refers
to a type of iris plant. One class is linearly separable from the other
two; the latter are not linearly separable from each other.
The data base contains the following attributes:
1). sepal length in cm
2). sepal width in cm
3). petal length in cm
4). petal width in cm
5). class:
- Iris Setosa
- Iris Versicolour
- Iris Virginica
Fishers Iris data base is available in Matlab (load fisheriris) and in
Internet (for example, on http://archive.ics.uci.edu/ml/datasets/Iris).
The goal of the seminar is to demonstrate the process of building a
neural network based classifier that solves the classification problem.
During the seminar various neural network based approaches will be
shown, the process of building various neural network architectures
will be demonstrated, and finally classification results will be
presented.
2. Theoretical part

In this seminar classification problem is solved by 3 types of neural

networks:
1) multilayer perceptron;
2) radial basis function network;
3) probabilistic neural network.
These network types are shortly described in this seminar. Each of
these networks has adjustable parameters that affect its
performance.

2.1Multilayer perceptron
Multilayer perceptron is a multilayer feedforward network.

Feedforward networks often have one or more hidden layers of

sigmoid neurons followed by an output layer of linear neurons.
Multiple layers of neurons with nonlinear transfer functions allow the
network to learn nonlinear and linear relationships between input
and output vectors. The linear output layer lets the network produce
values outside the range -1 to +1.
In this seminar the transfer functions of hidden layers are hyperbolic
tangent sigmoid functions. Network architecture is determined by
the number of hidden layers and by the number of neurons in each
hidden layer.
The network is trained by the backpropagation learning rule.

2.2 Radial basis function network

Radial basis function network is a feedforward network.

Radial basis function networks consist of two layers: a hidden radial

basis layer of S1 neurons, and an output linear layer of S2 neurons.
Each radial basis layer neurons weighted input is the distance
between the input vector and its weight vector. Each radial basis
layer neurons net input is the element-by-element product of its
weighted input with its bias. Each neurons output is its net input
passed through radial basis transfer function.
Radial basis function network is created iteratively one neuron at a
time. Neurons are added to the network until the sum-squared error
falls beneath an error goal or a maximum number of neurons has
been reached.
Design parameter of radial basis function network is spread of radial
basis transfer function.

2.3 Probabilistic neural network

Probabilistic neural network is a feedforward network. It is specialized
to classification.

When an input is presented, the first layer computes distances from

the input vector to the training input vectors and produces a vector
whose elements indicate how close the input is to a training input.
The second layer sums these contributions for each class of inputs to
produce as its net output a vector of probabilities. Finally, a
competitive output layer picks the maximum of these probabilities,
and produces a 1 for that class and a 0 for the other classes.
Design parameter of probabilistic neural network is spread of radial
basis transfer function.
Little or no training is required for probabilistic neural network (except
spread optimization).
3. Practical part

3.1 Cross-validation
In this seminar a cross-validation procedure is applied to provide
better generalization of neural network classifiers. To perform the
cross-validation procedure input data is partitioned into 3 sets:
1) training set;
2) validation set;
3) test set.
The training set is used to train the network. The validation set is used
to validate the network, to adjust network design parameters. The
test set is used to test the generalization performance of the
selected design of neural network.
To ensure a correct comparison of different types of neural networks
the division of input data into training, validation and test sets is
performed by independent part of code (see Appendix) and the
division result is stored.
The partitioning of input data is performed randomly with a certain
ratio of input entities to be stored as training set, validation set and
test set (0.7, 0.15 and 0.15 respectively).

3.2 Multilayer perceptron

As soon as the architecture and the performance of multilayer
perceptron are determined by the number of hidden layers and by
the number of neurons in each hidden layer these are the network
design parameters that are adjusted. The correct classification
function is introduced as the ratio of number of correctly classified
inputs to the whole number of inputs.
Multilayer perceptrons with 1 and 2 hidden layers are investigated.
The procedure of adjusting the number of neurons in hidden layers is
organized as a grid search (see Appendix). With each combination
of numbers of neurons in the hidden layers the multilayer perceptron
is trained on the train set, the value of correct classification function
for the train set is stored. The validation set is used for standard early
stopping procedure, the value of correct classification function for
the validation set is stored as well.
The values of the correct classification function are plotted versus
the corresponding number of neurons in the hidden layer.

Fig. 1. Correct classification function for multilayer perceptron with 1 hidden layer. Blue line
training set; green line validation set
Fig. 2. Correct classification function for multilayer perceptron with 2 hidden layers (2 ortogonal
projections of surface). Training set

Fig. 3. Correct classification function for multilayer perceptron with 2 hidden layers (2 ortogonal
projections of surface). Validation set

The number of neurons that ensures the best generalization is

chosen. The training and simulation of the chosen model of
multilayer perceptron is performed on joined training and validation
sets, the value of correct classification function is calculated.
Finally, the generalization performance of the network is simulated
on the test set. The corresponding value of correct classification
function is calculated.
For the training of the multilayer perceptron BFGS algorithm is used,
as it is known that for small networks quasi-Newton algorithms are
preferable.

3.3 Radial basis function network

To obtain the optimal performance of the radial basis function
network spread parameter is adjusted. The procedure of adjusting
the spread is organized as a search (see Appendix). The search is
performed in 2 iterations with different range of varying parameter
and different search step.
For each value of spread the radial basis function network is
simulated for the train and validation sets, the values of correct
classification function for the train set and for the validation set are
stored.
The values of the correct classification function are plotted versus
the spread.
 Fig. 4. Correct classification function for radial basis function network. Blue line training set;
green line validation set

The value of spread parameter that ensures the best generalization

is chosen. The radial basis function network is built and simulated on
joined training and validation sets, the value of correct classification
function is calculated.
Finally, the generalization performance of the network is tested on
the test set. The corresponding value of correct classification
function is calculated.

3.4 Probabilistic neural network

To obtain the optimal performance of the probabilistic neural
network spread parameter is adjusted. The procedure of adjusting
the spread is organized as a search (see Appendix). The search is
performed in 2 iterations with different range of varying parameter
and different search step.
For each value of spread the probabilistic neural network is
simulated for the train and validation sets, the values of correct
classification function for the train set and for the validation set are
stored.
The values of the correct classification function are plotted versus
the spread.

Fig. 5. Correct classification function for radial basis function network. Blue line training set;
green line validation set

The value of spread parameter that ensures the best generalization

is chosen. The probabilistic neural network is built and simulated on
joined training and validation sets, the value of correct classification
function is calculated.
Finally, the generalization performance of the network is tested on
the test set. The corresponding value of correct classification
function is calculated.

3.5 Results comparison

In the following table the values of correct classification function
obtained by supplying different sets of input data into the chosen
models of neural networks and processing their outputs are
presented.
Table 1

Neural networks
Radial basis
Sets of inputs Multilayer Probabilistic
function
perceptron neural network
network
training +
99.483% 99.225% 98.450%
validation
test 96.825% 100% 95.238%
4. Possible ways to improve the performance of discussed neural
networks

From working on my seminar and from discussions with my mentor I

see the following ways to improve the performance of neural
networks investigated:
1. Input data pre-processing
1) Partitioning of the input data for the cross-validation procedure
has effect on the neural network performance. Indeed, even
when the same values of division ratios are kept (0.7/0.15/0.15)
and the whole data set is partitioned randomly again, the
values of the correct classification function change:
Table 2

Neural networks
Radial basis Probabilistic
Sets of inputs Multilayer
function neural
perceptron
network network
training +
100% 99.483% 100%
validation
test 96.825% 96.825% 95.238%
This probably happens because the number of inputs is very
small and the performance of the network is very sensitive to
the way the original set is partitioned. The partitioning of original
set can also be optimized. The possible way is to divide the
original data set into a number of small sets and to search
through them for the one that ensures the best generalization
being used as validation set.
2) The principal component analysis can be applied to the
original data set to reduce its dimensionality.
2. Multilayer perceptron adjustment
While adjusting the number of neurons in hidden layers of
multilayer perceptron the results of grid search appear not
unique. For example, for a single hidden layer the plots of the
correct classification function versus number of neurons are
different each time the search is performed:

This probably happens because on every search run the

training is finished in a different local minima of network
performance function.
Applying a different network training function doesnt solve the
problem. Still, the BFGS algorithm is used for training instead of
standard Levenberg-Marquardt algorithm, because the
training is performed slightly faster.
5. Conclusions

1. Classification performance of all 3 investigated types of neural

networks is acceptable.
2. Radial basis function network exhibits better generalization
performance then multilayer perceptron and probabilistic
neural network.
3. Small number of inputs effect crucially on the generalization
performance of neural network classifier.
Appendix

1. Multilayer perceptron Matlab code

close all; clear; clc
%% load divided input data set
load divinp.mat
% coding (+1/-1) of 3 classes
a = [-1 -1 +1]';
b = [-1 +1 -1]';
c = [+1 -1 -1]';
% define training inputs
trainInp = [trainSeto trainVers trainVirg];
% define targets
T = [repmat(a,1,length(trainSeto)) repmat(b,1,length(trainVers))
repmat(c,1,length(trainVirg))];
%% network training
trainCor = zeros(10,10);
valCor = zeros(10,10);
Xn = zeros(1,10);
Yn = zeros(1,10);
for k = 1:10,
Yn(1,k) = k;
for n = 1:10,
Xn(1,n) = n;
net = newff(trainInp,T,[k n],{},'trainbfg');
net = init(net);
net.divideParam.trainRatio = 1;
net.divideParam.valRatio = 0;
net.divideParam.testRatio = 0;
%net.trainParam.show = NaN;
net.trainParam.max_fail = 2;

valInp = [valSeto valVers valVirg];

VV.P = valInp;
valT = [repmat(a,1,length(valSeto)) repmat(b,1,length(valVers))
repmat(c,1,length(valVirg))];

net = train(net,trainInp,T,[],[],VV);%,TV);

Y = sim(net,trainInp);

[Yval,Pfval,Afval,Eval,perfval] = sim(net,valInp,[],[],valT);

% calculate [%] of correct classifications

trainCor(k,n) = 100 * length(find(T.*Y > 0)) / length(T);
valCor(k,n) = 100 * length(find(valT.*Yval > 0)) / length(valT);
end
end
figure
surf(Xn,Yn,trainCor/3);
view(2)
figure
surf(Xn,Yn,valCor/3);
view(2)
%% final training
k = 3;
n = 3;

fintrain = [trainInp valInp];

finT = [T valT];

net = newff(fintrain,finT,[k n],{},'trainbfg');

net.divideParam.trainRatio = 1;
net.divideParam.valRatio = 0;
net.divideParam.testRatio = 0;

net = train(net,fintrain,finT);

finY = sim(net,fintrain);

finCor = 100 * length(find(finT.*finY > 0)) / length(finT);

fprintf('Num of neurons in 1st layer = %d\n',net.layers{1}.size)
fprintf('Num of neurons in 2nd layer = %d\n',net.layers{2}.size)
fprintf('Correct class = %.3f %%\n',finCor/3)
%% Testing
% define test set
testInp = [testSeto testVers testVirg];
testT = [repmat(a,1,length(testSeto)) repmat(b,1,length(testVers))
repmat(c,1,length(testVirg))];

testOut = sim(net,testInp);
testCor = 100 * length(find(testT.*testOut > 0)) / length(testT);
fprintf('Correct class = %.3f %%\n',testCor/3)

% plot targets and network response

figure;
plot(testT')
xlim([1 21])
ylim([0 2])
set(gca,'ytick',[1 2 3])
hold on
grid on
plot(testOut','r')
legend('Targets','Network response')
xlabel('Sample No.')

2. Radial basis function network Matlab code

close all; clear; clc
%% load divided input data set
load divinp.mat
% coding (+1/-1) of 3 classes
a = [-1 -1 +1]';
b = [-1 +1 -1]';
c = [+1 -1 -1]';
% define training inputs
trainInp = [trainSeto trainVers trainVirg];
% define targets
T = [repmat(a,1,length(trainSeto)) repmat(b,1,length(trainVers))
repmat(c,1,length(trainVirg))];
%% choose a spread constant (1st step)
spread = 2.1;
Cor = zeros(2,209);
Sp = zeros(1,209);
Sp(1,1) = spread;
for i = 1:209,
spread = spread - 0.01;
Sp(1,i) = spread;
% choose max number of neurons
K = 40;
% performance goal (SSE)
goal = 0;
% number of neurons to add between displays
Ki = 5;

% create a neural network

net = newrb(trainInp,T,goal,spread,K,Ki);

% simulate RBFN on training data

Y = sim(net,trainInp);

% define validation vector

valInp = [valSeto valVers valVirg];
valT = [repmat(a,1,length(valSeto)) repmat(b,1,length(valVers))
repmat(c,1,length(valVirg))];
[Yval,Pf,Af,E,perf] = sim(net,valInp,[],[],valT);

% calculate [%] of correct classifications

Cor(1,i) = 100 * length(find(T.*Y > 0)) / length(T);
Cor(2,i) = 100 * length(find(valT.*Yval > 0)) / length(valT);
end
figure
pl = plot(Sp,Cor/3);
set(pl,{'linewidth'},{1,3}');
%% choose a spread constant (2nd step)
spread = 1.0;
Cor = zeros(2,410);
Sp = zeros(1,410);
Sp(1,1) = spread;
for i = 1:410,
spread = spread - 0.001;
Sp(1,i) = spread;
% choose max number of neurons
K = 40;
% performance goal (SSE)
goal = 0;
% number of neurons to add between displays
Ki = 5;

% create a neural network

net = newrb(trainInp,T,goal,spread,K,Ki);

% simulate RBFN on training data

Y = sim(net,trainInp);

% define validation vector

valInp = [valSeto valVers valVirg];
valT = [repmat(a,1,length(valSeto)) repmat(b,1,length(valVers))
repmat(c,1,length(valVirg))];
[Yval,Pf,Af,E,perf] = sim(net,valInp,[],[],valT);

% calculate [%] of correct classifications

Cor(1,i) = 100 * length(find(T.*Y > 0)) / length(T);
Cor(2,i) = 100 * length(find(valT.*Yval > 0)) / length(valT);
end
figure
pl = plot(Sp,Cor/3);
set(pl,{'linewidth'},{1,3}');
%% final training
spr = 0.8;
fintrain = [trainInp valInp];
finT = [T valT];
[net,tr] = newrb(fintrain,finT,goal,spr,K,Ki);

% simulate RBFN on training data

finY = sim(net,fintrain);

% calculate [%] of correct classifications

finCor = 100 * length(find(finT.*finY > 0)) / length(finT);
fprintf('\nSpread = %.3f\n',spr)
fprintf('Num of neurons = %d\n',net.layers{1}.size)
fprintf('Correct class = %.3f %%\n',finCor/3)
% plot targets and network response
figure;
plot(T')
ylim([-2 2])
set(gca,'ytick',[-2 0 2])
hold on
grid on
plot(Y','r')
legend('Targets','Network response')
xlabel('Sample No.')
%% Testing
% define test set
testInp = [testSeto testVers testVirg];
testT = [repmat(a,1,length(testSeto)) repmat(b,1,length(testVers))
repmat(c,1,length(testVirg))];

testOut = sim(net,testInp);
testCor = 100 * length(find(testT.*testOut > 0)) / length(testT);
fprintf('\nSpread = %.3f\n',spr)
fprintf('Num of neurons = %d\n',net.layers{1}.size)
fprintf('Correct class = %.3f %%\n',testCor/3)

% plot targets and network response

figure;
plot(testT')
ylim([-2 2])
set(gca,'ytick',[-2 0 2])
hold on
grid on
plot(testOut','r')
legend('Targets','Network response')
xlabel('Sample No.')

3. Probabilistic neural network Matlab code

close all; clear; clc
%% load divided input data set
load divinp.mat
% coding the classes
a = 1;
b = 2;
c = 3;
% define training inputs
trainInp = [trainSeto trainVers trainVirg];
% define targets
T = [repmat(a,1,length(trainSeto)) repmat(b,1,length(trainVers))
repmat(c,1,length(trainVirg))];
%% choose a spread constant (1st step)
spread = 1.1;
Cor = zeros(2,109);
Sp = zeros(1,109);
Sp(1,1) = spread;
for i = 1:109,
spread = spread - 0.01;
Sp(1,i) = spread;
% create a neural network
net = newpnn(trainInp,ind2vec(T),spread);

% simulate PNN on training data

Y = sim(net,trainInp);
% convert PNN outputs
Y = vec2ind(Y);

% define validation vector

valInp = [valSeto valVers valVirg];
valT = [repmat(a,1,length(valSeto)) repmat(b,1,length(valVers))
repmat(c,1,length(valVirg))];
Yval = sim(net,valInp,[],[],ind2vec(valT));
Yval = vec2ind(Yval);

% calculate [%] of correct classifications

Cor(1,i) = 100 * length(find(T==Y)) / length(T);
Cor(2,i) = 100 * length(find(valT==Yval)) / length(valT);
end
figure
pl = plot(Sp,Cor);
set(pl,{'linewidth'},{1,3}');
%% choose a spread constant (2nd step)
spread = 0.25;
Cor1 = zeros(2,200);
Sp1 = zeros(1,200);
Sp1(1,1) = spread;
for i = 1:200,
spread = spread - 0.0001;
Sp1(1,i) = spread;
% create a neural network
net = newpnn(trainInp,ind2vec(T),spread);

% simulate PNN on training data

Y = sim(net,trainInp);
% convert PNN outputs
Y = vec2ind(Y);

Yval = sim(net,valInp,[],[],ind2vec(valT));
Yval = vec2ind(Yval);

% calculate [%] of correct classifications

Cor1(1,i) = 100 * length(find(T==Y)) / length(T);
Cor1(2,i) = 100 * length(find(valT==Yval)) / length(valT);
end
figure
pl1 = plot(Sp1,Cor1);
set(pl1,{'linewidth'},{1,3}');
%% final training
spr = 0.242;
fintrain = [trainInp valInp];
finT = [T valT];
net = newpnn(fintrain,ind2vec(finT),spr);

% simulate PNN on training data

finY = sim(net,fintrain);
% convert PNN outputs
finY = vec2ind(finY);

% calculate [%] of correct classifications

finCor = 100 * length(find(finT==finY)) / length(finT);
fprintf('\nSpread = %.3f\n',spr)
fprintf('Num of neurons = %d\n',net.layers{1}.size)
fprintf('Correct class = %.3f %%\n',finCor)
% plot targets and network response
figure;
plot(T')
ylim([0 4])
set(gca,'ytick',[1 2 3])
hold on
grid on
plot(Y','r')
legend('Targets','Network response')
xlabel('Sample No.')
%% Testing
% define test set
testInp = [testSeto testVers testVirg];
testT = [repmat(a,1,length(testSeto)) repmat(b,1,length(testVers))
repmat(c,1,length(testVirg))];

testOut = sim(net,testInp);
testOut = vec2ind(testOut);
testCor = 100 * length(find(testT==testOut)) / length(testT);
fprintf('\nSpread = %.3f\n',spr)
fprintf('Num of neurons = %d\n',net.layers{1}.size)
fprintf('Correct class = %.3f %%\n',testCor)

% plot targets and network response

figure;
plot(testT')
ylim([0 4])
set(gca,'ytick',[1 2 3])
hold on
grid on
plot(testOut','r')
legend('Targets','Network response')
xlabel('Sample No.')