I am trying to replace the objective function of neural network to customised one

I am trying to estimate the parameters using nueral networks with multiple layres for this equation; y = x1*x2*(alpha1) + x3*(alpha2), where x1, x2 and x3 are the inputs and y is the output. The neural network architecture is as follow, three inputs (x1,x2,x3) and two outputs (alpha1,alpha2) with multiple hidden layers. The objetive function will be as follow: e = [y - (x1*x2*(alpha1) + x3*(alpha2))], however, I can't change the perform.Fnc in matlab to customized one. I get the following error [performFcn" must be '' or the name of a network performance function], How can I solve this problem ? How can I use the estimated outputs from the NN as parameters inside the objective function? How to change the net.perform.Fcn from built-in ones to the one that I created ?

Answer 1

[Matt J]

This talks about supplying a custom model function for NN training:

https://www.mathworks.com/help/deeplearning/ug/train-network-using-model-function.html#TrainDeepNetworkInFunctionalFormExample-9

[Baker Shehadeh]

Thank you for your response, I went throught this example but I coudln't find what I am looking for. To be more specific, Can I chnage the "net.perform.Fcn" and assgin it to other function that I created instead of built-in ones like 'mse' ? I tried to assign it but it produced the following error (performFcn must be ' ' or the name of a network performance function)

[Matt J]

Perhaps you are refering to the net.performFcn mentioned here,

https://www.mathworks.com/help/deeplearning/ug/neural-network-object-properties.html#bss4hk6-52

[Matt J]

This post might be more relevant,

https://www.mathworks.com/matlabcentral/answers/461848-custom-performance-function-for-shallow-neural-networks-using-mse-package

[Baker Shehadeh]

Thank you. It's more relevant than the previous one, however, it didn't solve my problem since I don't want to create the custom fucntion based on specific package and I couldn't tweak the existing files to what I want. Can I create my own package and use it for my own specific case?

[Matt J]

No idea. Can you not migrate to the framework of the Deep Learning Toolbox for your NN work? It seems to be much more amenable to the kind of customization that you're after.