Determining collateral effects of deleting specific information from neural networks

Determine which records or pieces of information in a trained neural network will be affected when attempting to remove a particular item of training data, given that the correlations among internal representations are unknown.

Background

The authors highlight that information in neural networks is often entangled across parameters. Attempting to delete one item may inadvertently affect other items due to unknown correlations in internal representations.

This uncertainty poses risks for reliable machine unlearning, as erasing one individual's data could alter or erase information about others, undermining accuracy and legal compliance with the Right to be Forgotten.