How to Define Custom Regression Output Layer with Robust (Huber) Loss

The following example provides a framework to define a custom regression layer for MAE loss:

https://www.mathworks.com/help/deeplearning/ug/define-custom-regression-output-layer.html

How can you define a similar layer for Huber loss to reduce the effects of outliers? I am specifically asking for when the feature input is a simple numeric array.

Thank you.

Answer 1

[Onur Kilic]

I realized that I hadn't correctly formatted the predictions made by the network with dlarray. Below is the working regression layer based on Huber loss in case anyone needs it.

classdef huberRegressionLayer < nnet.layer.RegressionLayer ...
        & nnet.layer.Acceleratable % (Optional)
    
    methods
        function layer = huberRegressionLayer(name)           
            % layer = huberRegressionLayer(name) creates a
            % robust regression layer based on huber loss
            % and specifies the layer name.
            
            % Set layer name.
            layer.Name = name;
            % Set layer description.
            layer.Description = 'Huber loss';
        end
        function loss = forwardLoss(layer,Y,T)
            % loss = forwardLoss(layer, Y, T) returns the Huber loss between
            % the predictions Y and the training targets T.
            dlY = dlarray(Y,'CB');
            loss = huber(dlY,T,"TransitionPoint",1);
        end
    end
end