Trining a neural network with leave one out crossval method

Hello there,

I am new at neural networks and matlab. I am tring to train a network but i have less data available with me, so I am trying with leave-one-out method. But i am unable to find a way. Is there any direct method of training with leave-one-out training in matlab environment or what way should i follow. Thank you.

Answer 1

[Jalaj Gambhir]

Hi,

'Leave-one-out' is a cross validation method. You can generate cross validation indices for train and test set using cvpartition, specifying 'LeaveOut' parameter. This would generate partitions of n-1 training samples and 1 test sample.

>> load fisheriris;
>> x = meas;
>> y = species;
>> c = cvpartition(y,'LeaveOut')

This generates

c = 

Leave-one-out cross validation partition
   NumObservations: 150
       NumTestSets: 150
     TrainSize: 149  149  149  149  149  149  149  149  149  149 ...
      TestSize: 1  1  1  1  1  1  1  1  1  1 ...

For each partition 'i', you can generate train and test samples as:

>> x_train = x(training(c,i),:);
>> y_train = y(training(c,i),:); % You might want to convert this to one-hot-encoded vectors
>> x_test = x(test(c,i),:);
>> y_test = y(test(c,2),:); % You might want to convert this to one-hot-encoded vectors

Then you can use this train and test data to train a neural network using tools like nnstart which are perfect for beginners. Look at an example here.

[Juan Manuel Miguel]

Thank you Jalaj, it was very useful for me. I think you meant y_test = y(test(c,i),:); instead of y_test = y(test(c,2),:); didn't you?

Thank you