- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper investigates the overcorrection phenomenon in gradient-ascent-based machine unlearning, where the forgetting quality initially improves but subsequently deteriorates as unlearning continues. To explain this behavior, the authors develop a three-stage theoretical analysis using a two-layer ReLU convolutional neural network under a signal-noise data model. The analysis characterizes when the unlearned model approaches the retrained reference, how long it remains in a favorable forgetting regime, and how its behavior eventually deviates. Based on these findings, the authors propose a loss-threshold-based learning-rate adjustment strategy to mitigate overcorrection. Experiments on synthetic data, CIFAR-10, CIFAR-100, and Tiny ImageNet demonstrate non-monotonic forgetting behavior and suggest that the proposed strategy can delay deterioration.
Critical strengths
-
The paper investigates an important limitation of gradient-ascent-based unlearning by focusing on forgetting quality relative to retraining. The proposed three-stage analysis offers a potentially useful perspective on the dynamics of unlearning.
-
The overcorrection phenomenon is demonstrated across multiple image classification benchmarks and different forgetting configurations. The experiments provide useful evidence that excessive gradient ascent can deteriorate forgetting quality.
Critical weaknesses
-
There appear to be inconsistencies in the signal-noise decomposition and its initialization between Eq. (4.1) and Appendix E.2. Additionally, the gradient-ascent update normalization differs between Eq. (2.3) and the corresponding appendix derivations. These issues raise concerns about the correctness of the theoretical analysis and the claimed three-stage characterization.
-
The theoretical analysis focuses on KL divergence, whereas the main real-world experiments evaluate KS distance without establishing a rigorous connection between the two. Moreover, the proposed mitigation strategy is essentially a learning-rate reduction rule, and comparisons with simple alternatives such as early stopping and standard learning-rate scheduling are missing. The experiments also provide limited evidence regarding retained utility and overall unlearning effectiveness.
Questions for authors
-
Could the authors clarify the initialization of the signal-noise decomposition in Eq. (4.1) and Appendix E.2, and explain the different gradient-ascent normalization factors used in the main text and appendix? How do these differences affect the validity of Theorem 3.2?
-
What theoretical or empirical evidence supports transferring the KL-divergence-based analysis to the KS-distance-based evaluation? Can the authors directly verify the predicted three-stage behavior using the theoretical metric on real-world datasets?
-
How does the proposed learning-rate adjustment compare with early stopping and conventional learning-rate schedules under matched initial learning rates and computational budgets? In particular, can the authors clarify the different initial learning rates used for CIFAR-10 in Figures 2 and 4 and report the corresponding retained-data performance?
3 comments
Q3. Can the authors directly verify the predicted three-stage behavior using the theoretical metric on real-world datasets?
The predicted three-stage behavior can also be verified using the theoretical metric, KL divergence, on real-world datasets. We conduct additional experiments on CIFAR-100 to evaluate the evolution of KL divergence during unlearning. As shown in Table 2, the KL divergence first decreases and then increases, demonstrating that the overcorrection phenomenon persists under the theoretical metric on real-world data.
| Stage | Initial | Minimum | Final |
|---|---|---|---|
| KL | 1.000 | 0.376 | 0.802 |
Table 2: KL divergence at different stages of GA on CIFAR-100.
Q4. How does the proposed learning-rate adjustment compare with early stopping and conventional learning-rate schedules under matched initial learning rates and computational budgets? In particular, can the authors clarify the different initial learning rates used for CIFAR-10 in Figures 2 and 4 and report the corresponding retained-data performance?
Our proposed learning-rate adjustment strategy achieves lower KS distance and higher retain accuracy (RA) than early stopping and conventional learning-rate schedules. We conduct additional experiments on CIFAR-100, comparing our strategy with early stopping at the best-performing checkpoint (which is difficult to identify in practice), standard step decay, and cosine decay under matched initial learning rates and computational budgets. We also report the corresponding retain accuracy. As shown in Table 3, our strategy outperforms all these baselines in both KS distance and RA, demonstrating its effectiveness in improving forgetting quality while preserving retained-data performance.
| Method | KS Distance ↓ | RA (%) ↑ |
|---|---|---|
| GA | 0.760 | 37.72 |
| Early Stopping | 0.194 | 67.29 |
| Step Decay | 0.410 | 60.11 |
| Cosine Decay | 0.394 | 60.85 |
| Our Strategy | 0.186 | 67.53 |
Table 3: Comparison of learning-rate adjustment strategies on CIFAR-100 in terms of KS distance and retain accuracy (RA).
We also clarify that the discrepancy in the reported initial unlearning learning rates for CIFAR-10 between Figures 2 and 4 is solely a typographical error. The actual initial learning rate used in both experiments is .
References
[1] Zhang, R., et al. Negative preference optimization: From catastrophic collapse to effective unlearning. arXiv, 2024
Dear Reviewer GoVz,
As the discussion period is approaching its deadline, we would like to briefly follow up on our rebuttal. In response to your constructive suggestions, we have clarified the intermediate reparameterization steps omitted for brevity when describing the initialization of the unlearning process. We have also provided both empirical and theoretical evidence supporting the strong consistency between KL-divergence and KS-distance dynamics in our setting. Furthermore, we have conducted additional experiments comparing our proposed learning-rate adjustment strategy with early stopping and conventional learning-rate schedules, demonstrating its superiority in improving forgetting quality while preserving retained-data performance.
We believe that these clarifications and additional results have addressed all of your concerns. Since we will no longer be able to respond after the discussion closes, we would greatly appreciate it if you could briefly let us know whether any concerns remain. We would be happy to provide any final clarification before the deadline.
Thank you again for your time and valuable feedback.
Best regards, Authors of Submission 31009
Thanks for your constructive feedback! We address your questions and concerns as follows.
The initialization of the signal-noise decomposition in Eq. (4.1) and Appendix E.2 is consistent, with some intermediate reparameterization steps omitted for brevity. The unlearned model is initialized from the final weights of the base model trained on both the forgetting and retained data, i.e., wj,su,0=wj,sb,tend. In Eq. (4.1), the decomposition is expressed relative to wj,su,0, so the initial coefficient increments are zero. In Appendix E.2, we equivalently absorb the fixed coefficients inherited from the base model into the unlearning coefficients, with a corresponding shift of the decomposition's reference point. Specifically, we define γj,su,t(μf)=γj,sb,tend(μf)+Δγj,su,t(μf), and similarly for ρj,s,ifu,t(ξf,if). Since the initial increments are zero, this gives γj,su,0(μf)=γj,sb,tend(μf) and ρj,s,ifu,0(ξf,if)=ρj,s,ifb,tend(ξf,if), as stated in Appendix E.2. The inherited coefficients remain fixed, so this reparameterization preserves the same weight trajectory and coefficient update equations.
Regarding the gradient-ascent normalization factors, the correct denominator during unlearning is nfm, since gradient ascent is performed only on the forgetting dataset. The denominator m(nf+nr) appearing in the proof of Lemma E.2, including Eqs. (E.9)–(E.11), is a typographical error carried over from the base-model training derivation. Importantly, the update equations stated in Lemma E.2 are correct and consistent with Eq. (2.3) and Lemma 4.1 in the main text, all using η/(nfm). Thus, correcting this typographical error and clarifying the equivalent parameterizations do not change the underlying unlearning dynamics or the conclusions of Theorem 3.2.
In our setting, both empirical and theoretical evidence support the strong consistency between KL-divergence and KS-distance dynamics. We have added synthetic experiments with KS distance, and the results in Table 1 show that KL and KS exhibit highly consistent dynamics. For a more fine-grained comparison, we report their favorable intervals, defined as the steps at which the normalized metric is below 0.1, and their turning points. The favorable intervals are [2907, 2937] for KS and [2893, 2944] for KL, with corresponding turning points at steps 2922 and 2925, respectively. These results demonstrate that the favorable intervals substantially overlap and the turning points are nearly identical.
Table 1: KL divergence and KS distance on synthetic data.
From a theoretical perspective, Lemma 4.2 shows that a small signed-margin gap ∣zu−zr∣, where zu=yffu(xf) and zr=yffr(xf), implies a small KL divergence. Meanwhile, in the binary setting, the truth ratio is q(z)/(1−q(z))=exp(z), which is a strictly monotone transformation of the signed margin. Therefore, closeness of the signed-margin distributions is equivalent to closeness of the corresponding truth-ratio distributions, and hence a smaller KS distance. Our KL analysis thus captures the same underlying mechanism measured by KS: in the first stage, the unlearned margin approaches the retrained margin; in the second stage, they remain close; and in the third stage, they deviate, producing the same decrease-then-increase pattern. This combination of KS-based evaluation and KL-based theoretical analysis has also been adopted in [1].