load alphachars

X=alphabet;
T=eye(26);

nn=linspace(0.05,0.6,15);

for j=1:15
    X=[X alphabet+nn(j)*randn(size(alphabet))];
    T=[T eye(26)];
end

net=feedforwardnet(25);
net.divideParam.trainRatio=0.75;
net.divideParam.valRatio=0.10;
net.divideParam.testRatio=0.15;
%net.trainFcn='trainrp';

net=train(net,X,T);

nn=[0,nn];
Misclass=zeros(16,1);
for j=1:16
    Out=sim(net,alphabet+nn(j)*randn(size(alphabet)));
    for k=1:26
        [vv,ii]=max(Out(:,k));
        if ii~=k
            Misclass(j)=Misclass(j)+1;
        end
    end
    Misclass(j)=Misclass(j)/26;
end
figure(2)
plot(nn,Misclass);
title('Percent error versus Noise');

%% A nice way to visualize what Matlab has produced.
%
% ...  This is just for fun
%
% The result of the code below is to build a noisy test set, then see how
% Matlab classifies it.  For each test letter, we show the noisy letter,
% the actual letter, and what letter Matlab used.  The "pause" command
% below halts the code until any button is pressed.

figure(3)
Test=alphabet+0.4*randn(size(alphabet));
Out=sim(net,Test);
mm=colormap(gray);
mm=mm(end:-1:1,:);


for j=1:26
subplot(1,3,1)
colormap(mm)
imagesc(reshape(Test(:,j),5,7)');
axis equal; axis off
title('Input')
subplot(1,3,2)
imagesc(reshape(alphabet(:,j),5,7)');
axis equal; axis off
title('Actual')
[vv,ii]=max(Out(:,j));
subplot(1,3,3)
imagesc(reshape(alphabet(:,ii),5,7)');
axis equal; axis off
title('Matlab Says');
pause
end