Access to initial weights

[Mohit Burkule]

The unlearning function only has access to the "trained" weights in net variable

def unlearning(net, retain, forget, validation)

Would it be possible to have access to initial weights inside unlearn function during the competition (randomly generated or pre-trained, if you used pre-trained weights)

There is a chance that the forget-samples caused the model to converge into a specific minimum. Had the model seen only the retain-samples, it would have potentially converged into a different minimum. Having access to the initial weights allows us to experiment with how they update in response to the retain-samples versus a combination of both retain and forget samples. In case the initial weights are not available, the first checkpoint of the model's training might also prove to be useful.

I understand that we have these initial weights included in the starter kit provided, but it would be immensely beneficial to have access to them even during the competition when we only have access to variables passed to the unlearn function. 

[Bhargav Kowshik]

Interesting idea Mohit, I like this idea too!

Having access to both the foundational model as well as the pre-trained model should open up opportunities to understand how the training dataset (forget + retain) modified the model weights from that in the foundational model to that pre-trained model.

[Fabian Pedregosa]

Thanks for the suggestion! This is a very interesting idea. We'll consider it in future versions.

Kind regards

--
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/8a956d38-8687-4f83-aac6-74a1deed3e9bn%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.