Differentiable version of FNMR@FMR metric to use it as loss

[AlekseySh]

A paper for inspiration: link.

[deepslug]

Any updates/progress on this?

[AlekseySh]

@deepslug Nope, not enough resources. Do you want to try it? If so, the idea is that we want to adapt the approach from this paper and make FNMR@FMR metric differentiable.

[deepslug]

Unfortunately, I don't have enough resources either, but this is something I’d bring on top of my (and hopefully your) list!

[AlekseySh]

Got it!

[deepslug]

As I mentioned in my previous comment on the post, I work in the field of biometrics and am keen on seeing the differential version of FNMR@FMR as it can directly optimize the metric. Given the recent active development of OML, I wanted to add this comment to bump up the thread :)

[AlekseySh]

@deepslug thank you for you comment. I'd like to add that we've already implemented similar idea in OML. There is SurrogatePricisonLoss -- differentiable version of Precision metric. Experiments showed it was able to perform on SOTA level. So, it would be interesting to apply similar idea to FNMR@FMR.

Contributors are welcome!