- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper proposes Caffeine, an approximate machine unlearning method that operates without access to the retained data. At each saved training checkpoint, it uses two forget-set gradient evaluations to estimate curvature along the gradient direction. It scales the gradient by the inverse of this estimate, then combines the checkpoint updates using a learning-rate-weighted signed RMS rule to update the final model. The authors evaluate class deletion on three image datasets and two text datasets, and random-subset deletion on two image datasets. The reported results show a better trade-off between forget accuracy and utility than the evaluated approximate baselines, together with substantial runtime savings over retraining on the image tasks.
Critical strengths
- The method addresses a meaningful constraint with a relatively simple procedure. Preserving utility without access to retained data is a practical challenge. Caffeine estimates its update scale from forget-set gradients, avoids second-order automatic differentiation, and can modify the full network. These properties make it a potentially useful approach when training checkpoints are available.
- The reported efficiency–utility trade-off is promising. For example, on TinyImageNet class deletion, Caffeine achieves 46.37% test accuracy compared with 47.00% for retraining, while reducing runtime from 4,103 to 170 seconds. If stronger evaluations confirm its forgetting effectiveness, this efficiency improvement would be a meaningful contribution.
Critical weaknesses
- The evaluation does not sufficiently distinguish unlearning from changes in prediction behavior. Class deletion is assessed primarily through near-zero forget accuracy. Suppressing the deleted class’s output can achieve this without removing the relevant training information from the representation. The paper criticizes final-layer methods for leaving such information intact, but does not provide a corresponding representation or relearning evaluation to establish its own advantage. The random-subset results provide limited additional evidence. In Tables 6 and 7, the original model already has an MIA AUC of approximately 0.52, suggesting that the chosen attack has little ability to distinguish members before unlearning. Consequently, an AUC near 0.5 after applying Caffeine is weak evidence of successful removal. This limitation affects the paper’s central claim, rather than merely the breadth of its evaluation.
- The signed RMS aggregation is insufficiently justified, and the ablations do not establish its contribution. Equation (11) takes the sign from the weighted sum of historical updates and the magnitude from their RMS. For a coordinate with two equally weighted updates of and , the resulting update is approximately , despite the near cancellation of the two directions. Thus, the rule does not inherently produce small updates in oscillatory regions or establish the claimed trust-region behavior. The paper does not directly compare this aggregation with a weighted mean, nor does it include a well-tuned gradient-ascent baseline under the same data-access constraint and computational budget. It is therefore difficult to determine whether the gains come from curvature estimation, historical checkpoints, or simply a different update scale. These controls are particularly important given the severe degradation of several evaluated baselines.
Questions for authors
- Clarification of the experimental protocol: Does retraining start from an initialization that has not been trained on the deleted data? The statement that all methods, including retraining, start from the same original models and checkpoints is ambiguous. Also, which dataset and splits are used for the five-class “IMDB” experiment? Reference [35] points to the Stanford Sentiment Treebank.
- Evidence for the aggregation rule: What precisely does Remark 3 establish about signed RMS aggregation? How does that result constrain the update applied to the final model, beyond a bound relative to historical update magnitudes? Do you have comparisons against weighted-mean aggregation and fixed-step gradient ascent using the same checkpoints and comparable computational budgets? These comparisons would materially affect my assessment of the method’s contribution.
- Evidence for forgetting: Do you have results from a membership-inference attack with demonstrated discriminatory power on the original model, or representation/relearning evaluations against the retrained model? Please also report variability across runs for the AUC results and reconcile the discrepancies between Section 4.3 and Table 6. This would help determine how much evidence the reported attack provides. These questions concern evidence needed to support the current claims, rather than requests to extend the work to new applications. The supplied PDF does not include the cited supplementary material; if it already addresses these points, please identify the relevant results.
3 comments
I appreciate the authors' efforts in addressing my concerns. The additional representation-level evaluation, frozen linear probe, and SST-5 MIA experiment provide stronger evidence of the forgetting effectiveness. The clarification of the retraining protocol and the additional aggregation ablations are also helpful.
However, some concerns remain. In particular, the curvature estimation is still not directly validated, and the poor performance of some baselines (e.g., SFMU and HFOL) remains insufficiently explained. These issues leave some uncertainty about the technical validity and reliability of the experimental comparisons.
Overall, the rebuttal has improved the paper considerably, but I believe some important concerns remain unresolved. Therefore, I maintain my original rating.
We thank the reviewer for acknowledging that the rebuttal has improved the paper considerably. The review listed two weaknesses and three questions. As we read the reviewer's comment, these are now addressed: evidence of forgetting (W1, Q3) by the frozen-backbone probe, the SST-5 MIA and the standard deviations over three training seeds; the aggregation rule (W2, Q2) by Proposition 3 and by the weighted-mean and fixed-step comparisons, which the review said "would materially affect my assessment"; and the protocol (Q1).
Neither remaining point was among the review's questions. We answer both below and respectfully ask the reviewer to reconsider the rating in light of them. All appendix references are to the Supplementary Material PDF (pages 14-24); the probe table cited is now Table 13.
1. Direct validation of the curvature estimate. This is in the paper (Appendix F.4, Table 12, page 24 of the supplementary PDF, added in the latest revision). Caffeine uses only the scalar . We compare its two-gradient estimate with the exact at every checkpoint. The exact value is computed with Hessian-vector products by double backpropagation; it is the same number the materialized Hessian would give.
| Checkpoint (epoch) | Exact | (ours) | Relative error |
|---|---|---|---|
| 10 | 189.31 | 183.29 | 3.18% |
| 20 | 232.93 | 218.97 | 5.99% |
| 30 | 242.05 | 231.19 | 4.49% |
| 40 | 88.29 | 88.30 | 0.01% |
| 50 | 47.38 | 51.06 | 7.77% |
The error is below 8% at every checkpoint (smaller CNN of Table 11, CIFAR-10, class 0). Replacing by the exact inside Caffeine changes test, forget and retain accuracy by at most 0.15 points (Table 11).
2. Baseline performance. The review mentioned the degradation of several baselines as the reason for requesting controls. Those controls are now in Table 10 (page 22). They use no third-party code, and the same model and checkpoints as Caffeine:
| Variant | Test | Forget | Retain |
|---|---|---|---|
| Retrain | 85.48 | 0.00 | 89.98 |
| Fixed-step ascent, | 75.40 | 16.98 | 87.29 |
| Fixed-step ascent, | 68.63 | 2.84 | 83.02 |
| + weighted mean | 82.22 | 10.45 | 88.20 |
| Caffeine | 81.13 | 0.05 | 87.86 |
Our conclusions therefore do not rest on SFMU or HFOL: every result is measured against retraining on the same model, and the controls above show the same picture. SFMU, GradTao and SCRUB run from the authors' released code with their recommended hyperparameters (Appendix D, Table 8, page 20). HFOL's released code targets the small models of its paper, and we adapted it to ResNet-18 without changing the algorithm. We did not tune any baseline further and say so.
Why these two fail. Neither method was designed for class deletion: both papers evaluate sample-level or random-subset deletion.
SFMU. In the authors' own setting the released code reproduces the results reported in their paper (CIFAR-10, linear classifier on ResNet-18 features, 10% random-subset deletion: they report 70% test, 71% remaining and 68% forget accuracy; we obtain for SFMU, 79.3% test, 83.20% forget, and 83.14% retain accuracy, so its behavior here is a property of the method under class deletion, not of our setup. SFMU modifies only the final layer. Its backbone is identical to the original model's, so the deleted data remains in the features: a new linear head recovers the deleted class from SFMU's features at 88.62%, exactly as from the original model's (Table 13, page 24). A final-layer correction of the size SFMU applies does not remove a whole class.
HFOL. Its paper evaluates a small CNN and a linear model; ResNet-18, with 11.7M parameters, is outside that range. HFOL serves a deletion by adding the stored vector of every deleted sample to the model, followed by calibrated noise (Appendix D). Each vector is an approximation, and the errors and the noise accumulate with the number of samples deleted at once. With 5,000 samples on CIFAR-10 the accumulated update collapses the model, under class deletion (10.33% test accuracy) and under random-subset deletion (10.30%); on TinyImageNet it leaves the model unchanged (61.94% forget accuracy against 62.43%). It runs out of memory for two or more classes and cannot run on DistilBERT, whose attention kernel does not support the double backward pass it needs.
We thank the reviewer again for the constructive feedback.
Dear Reviewer,
The authors have kindly provided detailed responses to your comments. We would appreciate it if you could take a moment to review their rebuttal and share any further feedback or updates to your evaluation.
Thank you for your time and thoughtful consideration.