Evaluation data statistics

[Zafar Takhirov]

Suppose we use LiRA during the unlearning process. To speedup the process by using offline LiRA, would it be possible to get the distributions of the `in` set, as well as some representative `out` set distributions?
If not, would computing `N` models count towards the final evaluation?

[Fabian Pedregosa]

Hi Zafar,

1. The unlearning function signature provides a retain, forget, and validation set, from which in theory you could estimate these distributions. We don't plan to provide any other information besides this

2. Yes, anything done in the unlearning function runs towards runtime.

Kind regards

On Fri, Jul 28, 2023 at 11:24 PM Zafar Takhirov <z.ta...@gmail.com> wrote:

Suppose we use LiRA during the unlearning process. To speedup the process by using offline LiRA, would it be possible to get the distributions of the `in` set, as well as some representative `out` set distributions?
If not, would computing `N` models count towards the final evaluation?

--
You received this message because you are subscribed to the Google Groups "unlearning-challenge" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unlearning-chall...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/unlearning-challenge/9790d960-294a-4404-bbc1-b39b75ff29a3n%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.