Strange results with Matlab Neural Network Toolbox

Created and trained the network with NNToolbox, exported the weights and biases. The network works perfect, using sim(net) or just net(x). But when I try to reproduce the result with pen and paper, using the same weights and biases, the output is completely different. For testing I created a single layer net with PURELIN activation function. All weights and biases set to zero. According to all ANN books and wikipedias the result would be zero for all inputs, but Matlab says it's 0.5! If we set the weights to 1 and the biases to 0 (with PURELIN function the output must be the arithmetic sum of inputs), the result is even stranger: for [0;0] input it's -0.0611, for [1;1] it's 1.0605 and for [2;2] the output 2.1821. I am confused: I always believed Matlab and took its results "as is", but I can't explain that.

Answer 1

[Greg Heath]

By default, the feedforward nets transform the data using MAPMINMAX.

Therefore you have to either disable this default or use MAPMINMAX in your calculation.

See the documentation. For example,

help NEWFF

doc NEWFF

Similarly for

FEEDFORWARDNET, FITNET AND PATTERNNET

Hope this Helps.

Greg

[IEfremov]

Thanks, that explains everything!

Answer 2

[Greg Heath]

Please do not email followups. Stay within Answers.

If the following does not help, reply here, not via email.

% On Mar 31, 9:09 pm, "Siva " <mailto:[e-mail] [e-mail]> wrote:

> I am trying to determine the final equation for the trained network.-----SNIP

close all, clear all, clc

help feedforwardnet

[ x, t ] = simplefit_dataset;

[ I N ] = size(x)

[ O N ] = size(t)

whos

figure

hold on

plot(x,t,'o','LineWidth',2)

drawnow

H = 10

net = feedforwardnet(H);

net = train(net,x,t);

view(net)

y = net(x);

plot(x,y,'r','LineWidth',2)

drawnow

IW = net.IW{1,1}

LW = net.LW{2,1}

b1 = net.b{1}

b2 = net.b{2}

xmin = min(x)

xmax=max(x)

xn = -1+ 2*(x-xmin)/(xmax-xmin) ;

h =tansig(IW*xn+b1*ones(1,N));

yn = LW*h+b2;

tmin = min(t)

tmax = max(t)

yhat = tmin +(tmax-tmin)*(yn +1)/2;

plot(x,yhat,'g','LineWidth',2')

mse(y-yhat)

Hope this helps.

Greg

[IEfremov]

Thanks a lot! And sorry for spamming.