I want to use a loss function I created with a shallow neural network.

I want to use the app's neural network fitting to create a network that modifies only the loss function used for training.

In the app, MSE is used to evaluate the error.

I would like to use my own loss function instead of using MSE to evaluate the error.

When trainning, I want to use the loss function I created, but I don't know how to create the loss function and how to incorporate the created loss function.

How can I create a loss function in matlab? How can I use the loss function I created during training?

Answer 1

[Rushil]

Hello

The Neural Net Fitting App is used for creating shallow networks to solve simple data fitting tasks, most of which involve numbers. So, it may not be possible to add a custom loss function using the Neural Net Fitting app.

As an alternative, one may consider using the Deep Network Designer app, which allows designing a network visually. Consider the following steps to train a network with a custom loss function:

1) Design a network using the Deep Network Designer app. The network can be exported as a “dlnetwork” object to the MATLAB workspace.

2) Define a custom loss function as a separate MATLAB function file (for example, myCustomLoss.m):

function loss = myCustomLoss(Y, T) 
    loss = mean(abs(Y - T), 'all'); % mean absolute error (MAE) 
end 

3) Implement a custom training loop in MATLAB. In a loop, use the exported network and call the custom loss function during training. Below is an example of how it could be done:

for epoch = 1:numEpochs 
    % forward pass 
    dlY = forward(net, dlX); 
    
    % compute custom loss 
    loss = myCustomLoss(dlY, dlT); 

    % compute gradients and parameter update 
    gradients = dlgradient(loss, net.Learnables); 
    net = dlupdate(@(w, g) w - learningRate * g, net, gradients); 
end  

To find more information about the app and programmatic workflows in the Deep Network Designer, refer to the following documentation link:

https://mathworks.com/help/deeplearning/ref/deepnetworkdesigner-app.html

Hope it helps