I tried to create a neural networks but what's wrong ?

This is the code i tried :

net = newff(P,T,S)
net = newff(P,T,S,TF,BTF,BLF,PF,IPF,OPF,DDF)

Description

newff(P,T,S) takes,

• P - RxQ1 matrix of Q1 representative R-element input vectors.
• T - SNxQ2 matrix of Q2 representative SN-element target vectors.
• Si - Sizes of N-1 hidden layers, S1 to S(N-1), default = [].

(Output layer size SN is determined from T.) and returns an N layer feed-forward backprop network.

This a description that i found .So i follow the instructions to create a neural network with one hidden layer composed of 6 nodes .

trainInput=[1 0.4 0.2 0.7]
trainOutput=[0.8 ]
chrom=[0.5 0.6 0.8 0.6 0.7 0.9 0.8 0.4 0.5 0.9 0.9;
        0.1 0.7 0.6 0.7 0.9 0.5 0.9 0.2 0.4 0.5 0.9];
     X=chrom(1,:);
net=newff(minmax(trainInput'),trainOutput',6);

trainInput =

    1.0000    0.4000    0.2000    0.7000

trainOutput =

    0.8000

>> net.IW

ans = 

      [6x0 double]
      []
  >> net.LW

ans = 

                []    []
      [0x6 double]    []

>> net.b

ans = 

      [6x1 double]
      [0x1 double]

I didn't underdtand what does this notation means :

net.IW

ans = 

      [6x0 double]
      []

it seems there is no connection with input ,i entered 4 input why i just have 2 cell ?

[Greg Heath]

The description you mention is either innaccurate or you have misinterpreted it.

Where, exactly, is ist?

Try to duplicate the results from the one or more of the demos and/or examples in the NN TBX.

What version do you have?

newff is now osolete, do you have either

1. The later versions newfit (regression & curvefitting) or newpr ( classification & pattern recognition)

or

2. The latest versions fitnet (regression & curvefitting) or patternnet ( classification & pattern recognition)

Answer 1

[Walter Roberson]

One cell per layer, it appears to me.

You can view the contents by requesting net.b{1} and net.b{2}

Answer 2

[Greg Heath]

1. You need the same number of training targets as you have inputs.

 [I N ] = size(trainInputs)

 [ O N ] = size(trainOutputs)

This will result in Neq = N*O training equations

2. Correct your I-H-O network creation command

 net = newff(trainInput, trainOutput, H)

This will automatically result in Nw = (I+1)*H+(H+1)*O random initial weights. For (I,H,O) = (1,6,1), Nw = 12+7 = 19

3. You need to train the net to solve the Neq equations from (trainInputs,trainOutputs)for Nw unknown final weights.

net = train(net,trainInputs,trainOutputs);

4. Don't expect good results unless 0.7*N >> Nw = 19. If you cannot substantially increase the amount of training data,decrease H.

5. Try one or more of the examples or demos in the documentation.

Hope this helps.

Greg

Answer 3

[Mariem Harmassi]

think u for answering i have resolved the problem .and my code can genrate good results . But how can I randomly divide my dataset into k parts ??

Answer 4

[Greg Heath]

Please do not use the answer space for comments and/or asking additional unrelated questions.

The zeros in the weight vector cell components do not occur when the input and output data matrices have the same number of columns. I have not checked the source code to determine why.

The answer to your second question is:

help cvpartition

doc cvpartition

The documentation contains examples of implementing k-fold crossvalidation with the classify function.

However, there are no NN examples in the documentation. Nevertheless, I have written straightforward code that implements NN k-fold crossvalidation.

Try searching in the Newsgroup archives.

Thank you for reconsidering your official accepted answer.