How neural network output is calculated ?

For example

    function Example_3()

    net = feedforwardnet(1);

    p = [[2;2] [1;-2] [-2;2] [-1;1]]
    t = [1 2 1 2]

    net.layers{1}.transferFcn = 'logsig';
    net.layers{2}.transferFcn = 'logsig';
    net = train(net,p,t);

    wb = formwb(net,net.b,net.iw,net.lw);
    [b,iw,lw] = separatewb(net,wb);

    end

If

iw are -1.60579942154570 and 5.53933429980683

lw is -24.9335783159999

biases are -1.16445538542225 for the hidden neuron

and 7.83599936414935 for the output neuron

I was able able to calculate the correct values when using a perception but not with neural networks for some reason

I calculate the output using logsig(logsig((IW1*Input1)+(IW2*Input2)+bias1)*LW1+bias2)

If I input -1 and 1, how is the output calculated as 1.9994?

Shouldn't the output be between 0 and 1 because of the logsig ?

[Greg Heath]

When you choose examples, it is worthwhile to 1st consider continuous functions like ones in the help and doc documentation AND include plots of the inputs, targets and targets vs inputs.

Answer 1

[Greg Heath]

When you calculate the output of a net you have to take into account that values in the calculations are scaled, by default,

Hope this helps

Thank you for formally accepting my answer

Greg

[AlAwadhi]

Edit: ignore this, typing this to you made me notice my error. Thank you so much. Hello Greg,

First of all thank you for answering

I already did that. I can get the correct answers if the output is between -1 and 1. If I change the outputs to be between any other numbers, my results are different than matlabs. For example I get 9.5 and matlab gets 8.3. The range is correct but the answers are different.

I have a neural network with 1 hidden layer and it uses tansig as an activation function. The output layer also uses tansig

This is what I did in my excel sheet

1.First I scaled the inputs with

 Scaled input = (tanhMax-tanhMin)*(input-inputMin)/(inputMax-inputMin) + tanhMin

2.Each hidden neuron is equal to the

 HiddenNeuron = tanh(sum(inputs*inputWeights))

3.Each output neuron is equal to the

 OutputNeuron = tanh(sum(hidden*hiddenWeights))

4.I scaled back the outputs to be in the same range as my target range

 Final output = (TargetMax-TargetMin)*(output- tanhMin)/(tanhMax-tanhMin) + TargetMin

I'm not sure what I did wrong, but everything is good only if the target range is supposed to be between -1 and 1.

Thank you

Answer 2

[Greg Heath]

If your outputs are constrained to [ 0, 1 ] use SOFTMAX

If your outputs are constrained to [ -1, 1 ] use TANH

Otherwise use LINEAR

Hope this helps.

Thank you or formally accepting my answer

Greg