Hi, for a specific machine learning problem, is it possible to me to choose a function, like f(a) to evaluate a, while my loss function is actually generated from g(f(a))? If yes, are there any examples? If not, why I cannot use it? Thanks a lot.

I’m not sure if I have ever seen a use case of an evaluation function `f(a)`

and a loss `g(f(a))`

, but I don’t see why it couldn’t work assuming both methods are differentiable.

Note that you are often using a metric function `f(a)`

and a loss `g(a)`

such as the accuracy as the metric function and e.g. cross-entropy loss as the loss function.