Reading path · 4 papers
Unlearning was never only about privacy
The modern literature opens with the right to be forgotten. The founding paper does not. What was lost, and what recovered it?
-
1Towards Making Systems Forget with Machine Unlearning
Coins the term, three years before GDPR, and gives THREE motivations: privacy, SECURITY -- undoing training-data pollution attacks, where 3% mislabelled data significantly degrades a detector -- and USABILITY. Also separates deletion from privacy in a line worth quoting: an operation that outputs the entire dataset in the clear supports deletion and is certainly not private.
-
2Making AI Forget You: Data Deletion in Machine Learning
Narrows productively. Frames deletion as a computational problem with an Omega(n/m) amortised lower bound, and its four design principles -- linearity, laziness, modularity, quantization -- map the whole field from its second paper. Modularity is SISA two years early.
-
3Machine Unlearning
The narrowing completes. Opens on data-protection regulation, defines unlearning as distributional indistinguishability from retraining, and makes the right to be forgotten the organising motivation. Security and usability drop out of the frame.
-
4Towards Unbounded Machine Unlearning
Widens again, and this is the useful observation. Its three applications -- removing biases, resolving confusion from mislabelled data, and user privacy -- are close to Cao and Yang's original three. RC is recovery from label pollution; RB is the usability and correctness motivation.
What this leaves open
SCRUB claims to be the first to argue that forgetting is application-dependent. Against the EVALUATION literature that is fair. But the breadth was present in the founding paper and was lost in between, which makes the contribution a recovery rather than a discovery -- a more interesting story, and one the review can tell because it read the 2015 paper rather than citing it secondhand.