1

I'm working with a simple MLP which uses MSE as a loss function for its scalar target "y_true". However, MSE does not really capture the error of "y_pred" compared to "y_true".

In this case, a more suitable loss function would be based on another variable, lets call it "y_eval", as follows

loss_1(y_true,y_pred,y_eval) = max(0,y_true - y_eval) if y_pred <= y_eval 
                               else max(0,y_eval - y_true),

or even squared

loss_2(y_true,y_pred,y_eval) = max(0,y_true - y_eval)^2 if y_pred <= y_eval
                               else max(0,y_eval - y_true)^2.

What would be the best way to implement it? And is this a valid loss function? I'm thinking about creating a new target [y_true, y_eval] with shape (2,) and going from there... Thanks for any advice!

michael
  • 11
  • 2
  • I think I found a satisfying answer to my question here - [keras-loss-function-with-additional-dynamic-parameter](https://stackoverflow.com/questions/50124158/keras-loss-function-with-additional-dynamic-parameter) – michael Oct 16 '22 at 07:47

0 Answers0