How can change the performance function in neural network ?

I have Matlab versions R2013a and R2012b. I want to customize the performance function for neural network online training using 'adapt' function. The default performance function for both adapt/train functions is 'mse'. How can I change this function for adapt function.

What I tried so far: I tried changing the 'performFcn' to a .m file I wrote. Using the standard template of mse.m. I wrote all the necessary functions and the dcos function for gradient error function as well. But everything gave random errors - which mostly I traced back to the problem that latest versions of Matlab dont allow overriding 'mse' for adapt function. Is it true ?

Irrespective of that - how do I customize the performance function for neural network ?

Thanks

Anush

[Greg Heath]

What performance function do you want to implement?

What nndataset do you want to use for an example?

 help nndatasets

[Anush]

Hi Greg, Say, I will use the simple iris_dataset and I want to implement the cross entropy function. I have included my code snippet below.

    load iris_dataset.mat;
    netSize = [15 7];

    %Dummy data for initialization - all hardcoded for this example
    initData = ones([4 3]);
    initLabel = [1 0 0;0 1 0;0 0 1];
    neuralNet = feedforwardnet(netSize,'trainscg');
    neuralNet.performFcn = 'crossEntropy'; 
    neuralNet = configure(neuralNet,initData,initLabel);

    % Adapt the network for each data
    for i = 1:size(irisInputs,2)
    [neuralNet,predictedOutput(:,i)] =     adapt(neuralNet,irisInputs(:,i),irisTargets(:,i));
    end

This code works perfectly fine for an online update of the neural network using 'mse' as performance function. Now I have mentioned the performance function to be 'crossEntropy'. How do I go about writing the crossEntropy.m file the following function:

    eta.*(-irisTargets(:,i).*log(predictedOutput(:,i)))

Thanks

Anush

[Anush]

Any help in this regard ?

[Giorgio]

Hi Anush, Did you solve your problem ?? i'm working on MATLAB R2014b and i'm stuck to your same problem.

Thanks

Answer 1

[Greg Heath]

The PATTERNNET default is crossentropy. There is nothing more for you to do except make sure that the target matrix columns are from eye(c) for c classes..

help crossentropy

doc crossentropy

type crossentropy

Hope this helps

Thank you for formally accepting my answer

Greg