beta-NLL

[gdevos010]

Is your feature request related to a current problem? Please describe. beta-NLL from the paper "On the Pitfalls of Heteroscedastic Uncertainty Estimation with Probabilistic Neural Networks"

Describe proposed solution We could create a new likelihood

or

It could be added to the _nlloss() method https://github.com/unit8co/darts/blob/155c653a5243af3dd56faad9ffa539bd58c8e050/darts/utils/likelihood_models.py#L126-L132

[hrzn]

Nice! Seems like a great addition indeed. Adding to the backlog. It can perhaps be implemented by accepting an optional beta parameter in the GaussianLikelihood.

[hrzn]

Nice! Seems like a great addition indeed. Adding to the backlog.

[gdevos010]

I think it would look something like this

nll = -out_distr.log_prob(target).mean() 
if self.beta > 0.0:
    variance = out_distr.variance
    nll = nll * (variance.detach() ** self.beta)
return nll