Возможности среды Matlab для построения нейронных сетей, использование предоставленными возможностями, страница 8

%    along with 10 sets of noisy letters.

numNoisy = 10;

alphabet2 = [alphabet repmat(alphabet,1,numNoisy)+randn(35,26*numNoisy)*0.2];

targets2 = [targets repmat(targets,1,numNoisy)];

net2 = train(net,alphabet2,targets2);

%    ...and finally finishes.

pause % Strike any key to finish training the network...

% SET TESTING PARAMETERS

noise_range = 0:.05:.5;

max_test = 100;

network1 = [];

network2 = [];

% PERFORM THE TEST

for noiselevel = noise_range

fprintf('Testing networks with noise level of %.2f.\n',noiselevel);

Testing networks with noise level of 0.00.

errors1 = 0;

errors2 = 0;

for i=1:max_test

x = alphabet + randn(35,26)*noiselevel;

% TEST NETWORK 1

y = sim(net1,x);

yy = compet(y);

errors1 = errors1 + sum(sum(abs(yy-targets)))/2;

% TEST NETWORK 2

yn = sim(net2,x);

yyn = compet(yn);

errors2 = errors2 + sum(sum(abs(yyn-targets)))/2;

echo off

Testing networks with noise level of 0.05.

Testing networks with noise level of 0.10.

Testing networks with noise level of 0.15.

Testing networks with noise level of 0.20.

Testing networks with noise level of 0.25.

Testing networks with noise level of 0.30.

Testing networks with noise level of 0.35.

Testing networks with noise level of 0.40.

Testing networks with noise level of 0.45.

Testing networks with noise level of 0.50.

pause % Strike any key to display the test results...

%    DISPLAY RESULTS

%    ===============

%    Here is a plot showing the percentage of errors for

%    the two networks for varying levels of noise.

clf

plot(noise_range,network1*100,'--',noise_range,network2*100);

title('Percentage of Recognition Errors');

xlabel('Noise Level');

ylabel('Network 1 _ _   Network 2 ___');

%    Network 1, trained without noise, has more errors due

%    to noise than does Network 2, which was trained with noise.

echo off

End of APPCR1

2.9.5.  Cancer detection

Этот пример демонстрирует использование нейронной сети для обнаружения рака по данным общей спектрометрии в карте пациента. Цель – построить классификатор, который сможет различать раковых и контрольных пациентов по данным общей спектрометрии.

Методология, использовавшаяся в данной демонстрации – выбор сокращенного набора отличий или «особенностей», которые могут использоваться для определения раковых и контрольных пациентов с использованием классификатора. Такими особенностями будут уровни интенсивности ионов для определенных значений массы/заряда.

load OvarianCancerQAQCdataset.mat

whos

Name          Size                Bytes  Class     Attributes

MZ        15000x1                120000  double             

Y         15000x216            25920000  double             

grp         216x1                 15552  cell               

[feat,stat] = rankfeatures(Y,grp,'CRITERION','ttest','NUMBER',100);

P = double(Y(feat, :));      % Inputs

Cidx = strcmp('Cancer',grp); % Logical index vector for Cancer samples

T = double(Cidx)';           % Targets

rand('seed', 672880951)

net = newff(P,T,5);                        % create neural network

[net,tr] = train(net,P,T); % train network

testInputs = P(:,tr.testInd);

testTargets = T(:,tr.testInd);

out = round(sim(net,testInputs));           % Get response of trained network

diff = [testTargets - 2*out];

detections = length(find(diff==-1));        % cancer samples classified as cancerous

false_positives = length(find(diff==1));    % cancer samples classified as normal

true_positives = length(find(diff==0));     % normal samples classified as normal

false_alarms = length(find(diff==-2));      % normal samples classified as cancerous

Nt = size(testInputs,2);                    % Number of testing samples

fprintf('Total testing samples: %d\n', Nt);

%classification matrix

cm = [detections false_positives; false_alarms true_positives]

Total testing samples: 43

cm =

24     0

2    17

cm_p = (cm ./ Nt) .* 100                   % classification matrix in percentages

cm_p =

55.8140         0

4.6512   39.5349

fprintf('Percentage Correct classification   : %f%%\n', 100*(cm(1,1)+cm(2,2))/Nt);

fprintf('Percentage Incorrect classification : %f%%\n', 100*(cm(1,2)+cm(2,1))/Nt);