Using 5-fold cross validation with neural networks

I am trying to use k-fold with my neural networks to compare them with their 3 way split equivalents. I have a 150x4 dataset and since it is a very small amount I am trying to see whether 5-fold would allow the ANN to give better results since if I understood correctly Matlab will then pass 2 training sets 2 testing and a validation containing the respective number of rows after sorting the data randomly.

I have been going through some code examples however every time I try and implement a different example (from matlab or other websites) I either get an error using horzcat "Dimensions of matrices being concatenated are not consistent." or it doesn't work, which isn't making any sense since all 4 inputs and the output have 150x1 structure. I have tried applying k-fold to the individual columns (imported as a numeric matrix) and also as a collective matrix however all I get in the workspace are a testIdxs, a trainIdxs a and a k and an "Index exceeds matrix dimensions" in the command window

Can anyone explain how this needs to be done or share a link where k-fold is done on a dataset and then passed to a neural network for training?

Answer 1

[Greg Heath]

I just wrote a tutorial in the NEWSGROUP

http://www.mathworks.com/matlabcentral/newsreader/view_thread/340857

Hope this helps.

Thank you for formally accepting my answer

Greg

[Yogini Prabhu]

hi Greg

the page dosnt exist

Answer 2

[Greg Heath]

You probably just need to transpose your input and target matrices. However, to check previous posts consider

ANSWERS Hits

 greg neural crossvalidation    36
 greg neural cross-validation   61
 greg neural cross validation   67

NEWSGROUP Hits

 greg neural crossvalidation    32
 greg neural cross-validation   55
 greg neural cross validation   72

Additional references can be obtained by removing one or both of the leading search words

Hope this helps.

Greg

[Anthony Scicluna]

Thanks for the replies, I transposed the data and tried the following code:

k = 5;
cv = cvpartition(length(Temperature),'kfold',k);
for i=1:k
    trainIdxs{i} = find(training(cv,i));
    testIdxs{i}  = find(test(cv,i));
      trainMatrix{i} = [Temperature(trainIdxs{i}) HeartRate(trainIdxs{i}) Gender(trainIdxs{i}) Age(trainIdxs{i}) targets(trainIdxs{i})];
      validMatrix{i}  = [Temperature(testIdxs{i})  HeartRate(testIdxs{i}) Gender(testIdxs{i}) Age(testIdxs{i}) targets(testIdxs{i})];
  end

However I'm not sure if it has worked properly and how I would pass the information to the neural network.

the trainMatrix variable in the workspace is 1x5 cell comprising 1x600 and the validMatrix variable is also a 1x5 cell and comprises 1x150. As I said before though I think it isn't working properly since when I view the data all the five rows (four inputs and target) are placed in the same row.

[Yogini Prabhu]

those pages are not available now