Reading path · 3 papers
One method, three papers: relabeling
A short lineage worth following end to end, because it shows how a simple idea accretes machinery -- and because implementing the first gives us most of the third.
-
1Amnesiac Machine Learning
Introduces the idea as one of three methods: relabel the forget set with random incorrect labels and retrain briefly. Distinct from the batch- subtraction variant the same paper is better known for, and logging-free -- the authors themselves recommend it to owners worried about storage cost.
-
2SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
Adds a weight saliency mask. Equation 6 IS random labelling on the forget set plus a retain-set cross-entropy regulariser, with updates restricted to weights whose forget-loss gradient magnitude exceeds a threshold -- set, the authors report, at the median. The lineage is visible in the objective, not inferred from a summary.
-
3Model Sparsity Can Simplify Machine Unlearning
The contrast case, and the reason to read the two together. Same research group; SalUn criticises its l1-sparse baseline for looking strongest on unlearning accuracy alone while sacrificing retain and test accuracy. Checked against Jia's own Table 2, that critique holds -- and both papers are honest, since Jia report the 17x speed-up that buys the loss.
What this leaves open
Implementing Graves' relabeling variant next semester yields SalUn's objective minus the mask, so SalUn becomes a small increment rather than a fresh build. Whether the mask earns its complexity in OUR setting is untested: SalUn's evaluation is 10% and 50% random forgetting, while Phase 2 forgets 0.25%.