Neural Network Loss Function: Mean (absolute) Cubic Error

Hello,

for my neural network, it's very important to not have a high error-range, i.e. a higher mean-error is better than a higher error-range.

That's why I'd like to implement a different loss function. My network has a regressionLayer Output which computes loss based on mean squared error. To increase the weight of errors that lie further away, I'd like to change that into a mean cubic error.

The standard loss function of the regression Layer is and I'd like to perform a tiny change to or alternatively .

Is that possible in a not so complicated way?

Thank you for your help in advance,

Best regards

Answer 1

[Torsten]

You want the error to be negative if t_i < y_i ?

This won't work: The loss function should always be non-negative.

[ThomasP]

oh true, thanks, then it has to be either or the absolute value