Can I use weights and bias to manually verify the feedforwardnet?

I have 20 input and 20 target. I build feedforwardnet with one hidden layer of 10 neurons. I got "input weight (IW)" size of 10*20, "layer weight (LW)" size of 20*10. First bias size 10*1, second bias size 20*1.

I am trying to use Excel to manully calcuate the output step-by-step, so as to fully understand the procedure. What I do in Excel is (MMULT is the matrix multiplication function in Excel):

• MMULT(20 input, IW) = 10 elements
• 10 elements + First bias 10*1 = 10 elements
• tanh(10 elements) = 10 elements
• MMULT(10 elements, LW) = 20 elements
• 20 elements + second bias 20*1 = 20 elements

I can do the calculation based on the above steps. However, the final results differ a lot from the MatLab's prediction. I can get the exact result by using Pytorch's weights and bias. Is "Tanh" the default activation function in feedforwardnet?

Answer 1

[Steven Lord]

Don't forget the pre- and post-processing steps.

net = feedforwardnet(10);
net.inputs{1}

ans = 
    Neural Network Input

              name: 'Input'
    feedbackOutput: []
       processFcns: {'removeconstantrows', 'mapminmax'}
     processParams: {1x2 cell array of 2 params}
   processSettings: {1x2 cell array of 2 settings}
    processedRange: []
     processedSize: 0
             range: []
              size: 0
          userdata: (your custom info)

In this case that would be removeconstantrows and mapminmax. Also see the net.outputs property.