Where is the problem in my neural network script? I cannot update the weights and biases after training.

I am trying to design a custom neural network model to solve a problem. Here is the script:

net= network;
net.numInputs = 1;
net.inputs{1}.size = 1;
net.numLayers = 2;
net.layers{1}.size = 5;
net.layers{2}.size = 1;
net.inputConnect(1,1) = 1;
net.inputConnect(2,1) = 0;
net.biasConnect(1) = 1;
net.biasConnect(2) = 0;
net.layerConnect(2,1) = 1;
net.layers{1}.transferFcn = 'logsig';
net.layers{2}.transferFcn = 'purelin';
net.inputWeights{1,1}.initFcn = 'initzero';
net.inputWeights{1,1}.learn = 1;
net.inputWeights{1,1}.learnFcn = 'learngd';
net.layerWeights{2,1}.initFcn = 'initzero';
net.layerWeights{2,1}.learn = 1;
net.layerWeights{2,1}.learnFcn = 'learngd';
net.biases{1}.initFcn = 'initzero';
net.biases{1}.learn = 1;
net.biases{1}.learnFcn = 'learngd';
net.outputConnect = [0 1];
net.initFcn = 'initlay';
net.trainFcn = 'traingd';
net.performFcn = 'sse';

And here is the the diagram for the neural network:

Before I trained the network, I typed

net.IW{1}; net.b{1}; net.LW{2,1}

in the command window and checked that the values were all initially zero. However, when I trained the network, say

train(net, 0.1, 1)

Even though there was a pop-up window showing the process,

it seemed that the weights and biases did not update as I typed net.IW{1}, net.b{1} and net.LW{2,1}. Also, the output was always zero, no matter what value I input using

sim()

Could anyone help me to indicate the problem in the script or the procedure in training? Thank you.

[Greg Heath]

Choose a MATLAB example.

 help nndatasets
 doc nndatasets

Post all code and results.

Answer 1

[Greg Heath]

train(net, 0.1, 1)

You cannot train with a single point. You need input and target row vectors or matrices.

[Eric Leong]

Hi Greg,

Thanks for your answer and suggestion. So I have two additional questions. First, when I use simplefit_dataset as inputs and targets in the MATLAB examples and making up the same structure, the weights can be updated. But this is not the case while using my script.

And also, I am also curious why single points cannot be used in training in my case. I have tried that using a single point would work if I omit the second linear layer.

Could you explain why those things happen? Thank you.

[Greg Heath]

 For designing an I-H-O net with Ntrn input/target pairs 
you have 

Nw = (I+1)*H+(H+1)*O unknown weights and biases to 
be estimated with

 Ntrneq = Ntrn*O training equations

A necessary condition for a STABLE minimum MSE solution is that the number of unknowns do not exceed the number of equations

 Ntrn >= 1 + (I+O+1)*H/O

PS: Replace logsig with tansig and center your inputs. I like standardization with zero-mean/unit-variance inputs. Then you can check for outliers before training. However, the MATLAB default just transforms INPUTS AND TARGETS to [ -1 1 ]

Hope this helps.

Greg

[Eric Leong]

Thanks for your information, Greg.