Inference after deletion

[harelber]

Hi,
I tried to find the proper place to run inference after the deletion of some concepts. I wanted to infer something that is not identical to the original conecpt that was deleted. How can I do that?
Thanks!

[ericlormul]

Hi, it's the same as forgetting. As the concept has been forgotten after training, it is going to generate something that's different from the original concept. The caveat is what it generates depends on how strong is the forgetting process. For example, if you only do one step training of forgetting, it's going to be the identical concept. As the step increases, the concept will be less prominent.

[Yrefael]

Hi @ericlormul,
Following @harelber question, let's say I want to delete the concept "a house", and then I want to see how the model generates pictures of "a window". How practically would I do that?

[ericlormul]

Just use "a house" as the target concept for forgetting. Then, use the trained model(trained to forget 'a house') to generate 'a window', see how it will generates.