Training History Availability for Unlearning

[Martin VW]

Hello,

After reading the challenge instructions, I wonder if it is correct to assume that no information can be acquired during the training of the model ?

For instance, one might want to know how many times each data point in the forget set was sampled, and at which epoch, or revert the model to a previous checkpoint to then apply an unlearning function, as it is done in some parts of the (federated) unlearning literature.

Not having these options could limit the scope of possible approaches to tackling the unlearning challenge, especially since these quantities seem reasonably easy to evaluate in most real-world cases.

Please confirm if my understanding of the challenge is correct and if there are any possible solutions to the limitations I have mentioned.

Kind regards,

Martin Van Waerebeke

[Eleni Triantafillou]

Hi,

Thank you for your question!

Your understanding is correct. And we are aware that this setup does exclude a particular set of approaches (though we believe that, even within this scope, there are plenty of ideas to try, and we believe that this constitutes a great starting point for a first competition on the topic :)). 

We will consider broadening the scope in future iterations of this challenge, thanks for the feedback.

Thanks!

--
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/c4e021a2-5efa-4d22-9179-68b265cc18aen%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.