Adding some more resources

[meg-huggingface]

https://arxiv.org/abs/2003.10933 "Learn to Forget: Machine Unlearning via Neuron Masking", I believe would fit under "Model-Agnostic Approaches". It also comes with another evaluation metric called "FORSAKEN".

[meg-huggingface]

https://arxiv.org/abs/2206.06520 "Memory-Based Model Editing at Scale" is also relevant (neural-network specific)

[meg-huggingface]

https://arxiv.org/abs/1908.04319 "Neural Text Generation with Unlikelihood Training" is language-model specific Code is here: https://github.com/facebookresearch/unlikelihood_training

[adasken]

Thank you for the suggestion, we will review and update accordingly.

[meg-huggingface]

https://arxiv.org/abs/2205.10770 "Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models" is LM-specific, but is potentially useful as a Survey.

[meg-huggingface]

Slightly tangential, since this is more about concepts than directly learned facts, but still relevant just in case you want to add: https://arxiv.org/abs/2004.07667 , "Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection" Similarly, https://arxiv.org/abs/2201.12091 , "Linear Adversarial Concept Erasure"