acceptodds
Under review as a conference paper at ICLR 2027

Not All Samples Leave Equal Imprints: Influence-Aware Unlearning in Continual Learning

Abstract

Machine unlearning aims to remove the influence of specific data from a trained model, motivated by data regulations such as the right to be forgotten. Existing methods implicitly assume that all samples are encoded with similar strength, ignoring differences in their imprint on the model, and therefore apply uniform unlearning updates. This assumption breaks in class-incremental learning (CIL), where training data leave highly non-uniform imprints on the model. As a result, uniform unlearning leads to over-removal of weakly encoded samples and insufficient forgetting of strongly encoded ones, degrading overall performance. To address this, we propose IMPRINT, an imprint-aware unlearning framework for CIL. Our key idea is to align unlearning intensity with sample-wise imprint, estimated via the logit margin as a simple yet effective proxy. This enables adaptive scaling of the forget loss and seamless plug-and-play integration with existing unlearning methods. Extensive experiments show that IMPRINT consistently improves both forgetting and retention performance in CIL settings, significantly reducing the gap to retraining while outperforming state-of-the-art baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Continual Learning, Machine Unlearning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GmkdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies machine unlearning in class-incremental learning (CIL). The authors observe that a CIL model encodes different samples with very different strength: classes from early tasks are usually weakly encoded while classes from the most recent task are strongly encoded. Applying a uniform unlearning update to all forget samples therefore over-removes weakly encoded samples and under-removes strongly encoded ones. To address this, the authors propose IMPRINT, which estimates the imprint of each forget sample using its logit margin and converts the margin into a per-sample weight through a sigmoid. The weighting is model-agnostic and plug-and-play, and can be applied on top of existing unlearning methods such as RL, SalUn, AMUN and BS. The appendix also gives the margin a geometric interpretation and a lower bound connecting it to parameter-space removal effort. Experiments on CIFAR-100, Tiny-ImageNet, ImageNet-1k, a ViT setting, iCaRL, Task-IL/Domain-IL and 20 Newsgroups show that IMPRINT achieves the smallest average gap to retraining among the compared methods, with negligible runtime overhead.

Critical strengths
  1. The problem is concrete and well motivated. The variation in sample-wise imprint strength has direct empirical support: the class-level correlations in Figures 1 and 3, and the agreement with IMU's influence ranking in Table 2. This is a real issue that has not received explicit attention in the unlearning literature, and the paper gives a clear illustration of the failure case of uniform unlearning in Figure 2.

  2. The supporting analysis is complete relative to the simplicity of the method. The appendix derives the connection between margin, decision-boundary distance and parameter-space removal effort, and backs it with measured correlations of 0.960 against the first-order removal effort and 0.959 against an empirical boundary-crossing search. The mean-matched uniform control in Table 16 isolates the effect of sample-dependent weighting from simply rescaling the loss. The method requires no additional data or reference model, and the overhead in Figure 6 is negligible. The improvement holds across backbones, datasets, buffer sizes, ten baselines and several continual-unlearning dynamics experiments in Section E.

Critical weaknesses
  1. The main text is somewhat imbalanced in coverage. Most headline experiments use class-wise forgetting on CIFAR-100 and Tiny-ImageNet; the ImageNet-1k and ViT results use FOSTER only, and random data forgetting is discussed briefly.
  2. The margin proxy assumes the logit scale is roughly comparable across tasks. Appendix B.7 audits the natural task-wise scale, finding a ratio of only 1.032, and runs a 4x stress test in Table 15; however, the stress test only applies a monotone, task-level scaling to the margin computation. Whether the ordering stays stable under other drift patterns, such as drift concentrated in a few classes rather than whole tasks, is not really covered. I suggest adding a short discussion of such failure modes.
  3. The sensitivity ablations over the temperature, the forgetting ratio and the CIL training budget are all run on CIFAR-100 with DER++ only. The proxy comparison does extend to Tiny-ImageNet, as shown in Table 11, but I suggest extending at least the temperature ablation to one more backbone to make the robustness claims easier to accept.
Questions for authors
  1. The temperature sweep in Table 17 is on CIFAR-100 only. How sensitive is IMPRINT to tau on Tiny-ImageNet and ImageNet-1k, where the logit distributions can be quite different from CIFAR-100?
  2. Section 3.5 mentions upsampling the retain set to match the effective size of the forget set. How many upsampling iterations are typically needed, and does the unlearning time scale badly when the forget set is much larger than the replay buffer?
  3. The sample-wise margin is recomputed at every training epoch. Have you observed any instability in this adaptive process, such as oscillating weights for strongly encoded samples, and does a coarser update frequency, say every few epochs, work about as well?
  4. The scale audit in B.7 is task-wise. Have you checked whether a per-class logit-scale drift, one not aligned with task boundaries, occurs in practice, and if so, whether the margin ordering survives it?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Gmkd (1/3)

We sincerely thank the reviewer for the time, effort, and valuable feedback.


W1: Experimental Coverage

The revised manuscript now includes the following results.

1. ImageNet-1k plug-and-play comparison.

We compare RL, SalUn, AMUN, and BS with their IMPRINT variants on ImageNet-1k using ResNet-18 and FOSTER.

Table 7. ImageNet-1k plug-and-play results with FOSTER.

MethodFA ↓RA ↑MIA ↑FT ↓RT ↑Avg. Gap ↓
Retrain0.0039.43100.000.0043.18-
RL0.0018.2099.840.0021.059.09 ± 2.79
RL+IMPRINT0.0034.1899.990.0037.942.50 ± 0.42
SalUn0.0017.80100.000.0021.289.11 ± 2.05
SalUn+IMPRINT0.0034.25100.000.0038.462.38 ± 0.16
AMUN14.5837.8974.7316.7242.4313.25 ± 2.02
AMUN+IMPRINT4.0944.0789.994.9948.635.84 ± 0.73
BS6.6423.4486.638.0429.0211.64 ± 0.33
BS+IMPRINT0.5335.8898.990.6541.201.57 ± 0.46

IMPRINT reduces the mean Avg. Gap for all four methods on ImageNet-1k.

2. ViT with DER++.

We evaluate IMPRINT on CIFAR-100 with ViT and DER++, extending the submitted paper's ViT–FOSTER analysis. We compare uniform RL and SalUn with their IMPRINT variants at the default τ=0.5\tau=0.5 across five forgetting configurations.

Table: CIFAR-100 with ViT and DER++ (τ=0.5\tau=0.5).

MethodFA ↓RA ↑MIA ↑FT ↓RT ↑Avg. Gap ↓
Retrain0.0080.14100.000.0073.51-
RL0.0269.3899.990.0062.694.32 ± 0.97
RL+IMPRINT0.8876.3599.830.3169.871.76 ± 0.31
SalUn0.0370.4399.980.0063.913.87 ± 1.22
SalUn+IMPRINT0.7975.3399.780.3168.962.14 ± 0.80

IMPRINT reduces the mean Avg. Gap and improves retained-data performance for both methods, extending the ViT evidence to another CIL strategy.

3. Random forgetting on Tiny-ImageNet.

We extend random forgetting from CIFAR-100 to Tiny-ImageNet with ResNet-18 and DER++. Random forgetting selects individual samples across classes, producing forget sets with diverse class memberships and imprint strengths.

Table: Random forgetting on Tiny-ImageNet with ResNet-18 and DER++.

MethodFARAMIATAAvg.Gap4 ↓
Retrain24.9533.9868.3425.12-
RL18.9722.4795.1017.7012.92
RL+IMPRINT31.7933.8571.4428.213.29
SalUn21.6424.4994.0019.4311.04
SalUn+IMPRINT34.2835.7365.2328.894.49

IMPRINT reduces the average gap to retraining and improves retained-data performance, extending the evidence to sample-level deletion on another dataset. Together, these evaluations broaden coverage across the tested scenarios, datasets, CIL methods, and backbones. We have moved the ImageNet-1k plug-and-play comparison to Section 4.2 (Table 7) and added the ViT–DER++ and Tiny-ImageNet random-forgetting results to Appendix D.1 (Table 36) and Appendix D.2 (Table 37), respectively.


W2 & Q4: Robustness to Per-Class Logit-Scale Variation

We examine natural class-wise scale variation and controlled class-dependent distortions on CIFAR-100, using a replay buffer capacity of 2,000 and three forgetting configurations.

1. Natural variation and margin ordering.

We measure each sample's logit scale as the standard deviation of its output logits and define scs_c as the median scale among samples of class cc. We compare raw margins with class-scale-normalized margins:

minorm=migsyi,m_i^{\mathrm{norm}} = m_i \frac{g}{s_{y_i}},

where gg is the geometric mean of class-level scales.

  • P95/P05: the 95th-to-5th percentile ratio of class-level scales before unlearning, or their relative changes during unlearning.
  • Task-explained variance: the fraction explained by task means, computed from class log scales before unlearning and log-scale changes during unlearning.
  • Spearman's ρ\rho: ranking agreement between raw and normalized margins.
  • Top-10% overlap: the percentage of the raw-margin top decile also appearing in the normalized-margin top decile.

Table: Natural class-wise scale variation and margin ordering.

AnalysisP95/P05Task-explained varianceSpearman ρ\rho ↑Top-10% overlap ↑
Before unlearning, 3 seeds1.45–1.555.60–8.35%0.9945–0.995980.25–82.20%
During unlearning, epochs 1–101.31–1.337.35–8.94%0.9773–0.981289.63–91.63%

Class-wise scale variation occurs, with task means explaining only a small fraction. Overall ranking agreement remains high, although normalization changes some of the highest-ranked samples.

AuthorsOpenReview ↗

Response to Reviewer Gmkd (2/3)

2. Controlled within-task perturbations.

We apply three class-dependent scaling patterns within each original ten-class task:

Table: Scaling assignments within each original ten-class task.

PatternScaling assignments within each original task
Balanced 4×4\timesFive classes receive 22; five receive 0.50.5.
Sparse 4×4\timesOne class receives 40.9≈3.48224^{0.9} \approx 3.4822; nine receive 4−0.1≈0.87064^{-0.1} \approx 0.8706.
Inverse sparse 4×4\timesOne class receives 4−0.9≈0.28724^{-0.9} \approx 0.2872; nine receive 40.1≈1.14874^{0.1} \approx 1.1487.

Each pattern has a maximum-to-minimum scaling ratio of 44 and a geometric mean of 11. The sparse patterns distinguish one class from the other nine, directly testing variation concentrated within tasks.

For a sample with label yiy_i, we modify its margin and weight as

\widetilde{\omega}_i = \sigma\left(\frac{\widetilde{m}_i}{\tau}\right).$$ Base forget/retain losses and evaluation predictions use the original logits; perturbations affect only IMPRINT's weighting signal. We report ranking agreement between unperturbed and perturbed margins alongside unlearning performance. **Table: Effects of class-dependent scaling.** | Condition | Spearman $\rho$ ↑ | Top-10% overlap ↑ | FA ↓ | RA ↑ | MIA ↑ | FT ↓ | RT ↑ | Avg. Gap ↓ | |---|---:|---:|---:|---:|---:|---:|---:|---:| | Retrain | — | — | 0.00 | 64.36 | 100.00 | 0.00 | 55.62 | — | | **Native IMPRINT** | 1.0000 | 100.00% | 0.02 | 62.40 | 99.86 | 0.01 | 55.51 | **0.48 ± 0.01** | | Balanced $4\times$ | 0.9204 | 45.83% | 0.08 | 61.64 | 99.75 | 0.04 | 54.64 | 0.81 ± 0.05 | | Sparse $4\times$ | 0.9709 | 66.44% | 0.04 | 62.03 | 99.81 | 0.01 | 54.92 | 0.66 ± 0.03 | | Inverse sparse $4\times$ | 0.9756 | 91.80% | 0.08 | 61.90 | 99.75 | 0.01 | 54.81 | 0.73 ± 0.15 | Global rank agreement remains high, but top-decile membership changes substantially under balanced and sparse scaling—by approximately one-third under sparse scaling—and is better preserved under inverse-sparse scaling. Avg. Gap increases under all three perturbations, while the other metrics remain close to native performance. These results support continued effectiveness under the tested distortions, not invariance of ranking or performance. **3. Scope and failure modes.** Positive class-dependent scaling preserves margin signs and within-class ordering but can reorder samples across classes. Independently scaling or shifting output-class logits can also change the strongest competitor or margin sign, potentially weakening raw-margin weighting; these more general distortions are outside our tests. We have added the class-wise logit-scale analysis and discussion of failure modes to Appendix B.8 (Tables 18 and 19). ### W3 & Q1: Temperature Sensitivity Across Datasets and CIL Methods We extend the sweep to $\tau\in\{0.25,0.5,1.0\}$ on Tiny-ImageNet (ResNet-18, DER++), ImageNet-1k (ResNet-18, FOSTER), and CIFAR-100 (ViT, FOSTER and DER++), considering five forgetting configurations. **Table: Temperature sensitivity: Tiny-ImageNet, ResNet-18, DER++.** | Method | τ | FA ↓ | RA ↑ | MIA ↑ | FT ↓ | RT ↑ | Avg. Gap ↓ | |---|---:|---:|---:|---:|---:|---:|---:| | Retrain | — | 0.00 | 56.78 | 100.00 | 0.00 | 44.01 | - | | RL | — | 0.00 | 25.02 | 99.99 | 0.00 | 26.80 | 9.80 ± 1.20 | | **RL+IMPRINT** | 0.25 | 0.08 | 47.34 | 99.61 | 0.01 | 39.67 | **2.85 ± 1.26** | | RL+IMPRINT | 0.5 | 0.17 | 42.98 | 98.95 | 0.03 | 36.04 | 4.60 ± 1.02 | | RL+IMPRINT | 1 | 0.05 | 31.94 | 99.30 | 0.00 | 26.83 | 8.55 ± 4.10 | **Table: Temperature sensitivity: ImageNet-1k, ResNet-18, FOSTER.** | Method | τ | FA ↓ | RA ↑ | MIA ↑ | FT ↓ | RT ↑ | Avg. Gap ↓ | |---|---:|---:|---:|---:|---:|---:|---:| | Retrain | — | 0.00 | 39.43 | 100.00 | 0.00 | 43.18 | — | | RL | — | 0.00 | 20.76 | 100.00 | 0.00 | 23.57 | 7.66 ± 4.15 | | RL+IMPRINT | 0.25 | 0.00 | 34.53 | 100.00 | 0.00 | 38.22 | 1.97 ± 0.31 | | RL+IMPRINT | 0.5 | 0.00 | 34.55 | 100.00 | 0.00 | 38.19 | 1.97 ± 0.35 | | **RL+IMPRINT** | 1 | 0.00 | 34.88 | 100.00 | 0.00 | 38.61 | **1.83 ± 0.21** | **Table: Temperature sensitivity: CIFAR-100, ViT, FOSTER.** | Method | τ | FA ↓ | RA ↑ | MIA ↑ | FT ↓ | RT ↑ | Avg. Gap ↓ | |---|---:|---:|---:|---:|---:|---:|---:| | Retrain | — | 0.00 | 91.63 | 100.00 | 0.00 | 89.00 | - | | RL | — | 0.00 | 77.93 | 100.00 | 0.00 | 75.61 | 5.42 ± 2.56 | | **RL+IMPRINT** | 0.25 | 0.08 | 89.86 | 99.97 | 0.19 | 87.79 | **0.66 ± 0.59** | | RL+IMPRINT | 0.5 | 0.04 | 89.08 | 99.98 | 0.09 | 86.94 | 0.95 ± 0.62 | | RL+IMPRINT | 1 | 0.01 | 85.85 | 99.99 | 0.01 | 83.61 | 2.24 ± 0.61 |
AuthorsOpenReview ↗

Response to Reviewer Gmkd (3/3)

Table: Temperature sensitivity: CIFAR-100, ViT, DER++.

MethodτFA ↓RA ↑MIA ↑FT ↓RT ↑Avg. Gap ↓
Retrain—0.0080.14100.000.0073.51-
RL—0.0269.3899.990.0062.694.32 ± 0.97
RL+IMPRINT0.251.1575.4299.840.4269.052.18 ± 0.26
RL+IMPRINT0.50.8876.3599.830.3169.871.76 ± 0.31
RL+IMPRINT10.3174.9399.900.0568.582.12 ± 1.29

Temperature sensitivity varies across settings: smaller temperatures yield lower mean gaps on Tiny-ImageNet and ViT–FOSTER, ImageNet-1k shows little variation, and τ=0.5\tau=0.5 performs best on ViT–DER++. RL+IMPRINT achieves lower reported mean gaps than uniform RL at every tested temperature in all four settings, supporting τ=0.5\tau=0.5 as an effective shared default. We have added the temperature sweeps across datasets and backbones to Appendix C.1 (Table 22).


Q2: Retain-Set Upsampling and Computational Cost

We evaluate 10- and 30-class deletion on CIFAR-100 using RTX 3090 GPUs, a replay buffer capacity of 2,000, and batch size 256. Upsampling counts equivalent full passes through available retained samples per epoch. Elapsed times include setup and validation through the first epoch's validation checkpoint.

Table: Retain-set reuse and elapsed time on CIFAR-100.

Number of forget classesForget samplesForget/retain batches per epochRetain-set upsampling (×)Elapsed time through first-epoch validation
104,50018 / 182.2844.02 s
3013,50053 / 538.91710.34 s

The two settings require approximately 2.3×2.3\times and 8.9×8.9\times retain-set reuse per epoch. Each forget minibatch is paired with exactly one retain minibatch, giving E⌈F/B⌉E\lceil F/B\rceil minibatch pairs for FF forget samples, batch size BB, and EE epochs. At fixed BB and EE, optimization workload scales linearly with forget-set size; a smaller retain set increases reuse without adding optimization steps or a nested loop. We have clarified retain-set upsampling in Section 3.5 and added the reuse and runtime measurements to Appendix C.4 (Table 29).


Q3: Stability of Adaptive Logit Margin

Native IMPRINT recomputes sample-wise margins and weights at every forget minibatch; The submitted version described epoch-wise updates. We examine stability and coarser refresh schedules on CIFAR-100 with DER++ and a replay buffer of 2,000.

1. Adaptive-weight stability.

We track fixed evaluation samples in the top quartile of pre-unlearning margins, representing initially strongly imprinted samples, and evaluate their weights at epoch checkpoints.

Table: Adaptive-weight stability for initially strongly imprinted samples.

DiagnosticResult
Samples with repeated margin-sign changes0.484%
Samples with repeated extreme weight switches (≤0.1 ↔ ≥0.9)0.000%
Mean adjacent absolute weight change0.00382

Repeated margin-sign changes are rare, no repeated extreme weight switches are observed, and mean adjacent-epoch changes are small. These diagnostics provide little evidence of persistent oscillations at the recorded epoch checkpoints.

2. Weight-refresh frequency.

For k∈{2,3,5}k\in\{2,3,5\}, refresh epochs begin at epoch 1 and recur every kk epochs. During each refresh epoch, margins and weights are recomputed at every forget minibatch using the native implementation and cached. Intervening epochs continue minibatch optimization using the corresponding cached weights. Thus, k=2k=2 alternates native recomputation in epochs 1, 3, 5, … with cached weights in epochs 2, 4, 6, …. Results are averaged over three paired deletion requests.

Table: Effect of weight-refresh frequency on CIFAR-100.

MethodFA ↓RA ↑MIA ↑FT ↓RT ↑Avg. Gap ↓
Retrain0.0064.36100.000.0055.62—
RL2.1557.3299.381.2150.693.19 ± 0.35
IMPRINT (native)0.0262.4099.860.0155.510.48 ± 0.01
IMPRINT (k=2k=2)0.1662.9699.500.0355.800.48 ± 0.12
IMPRINT (k=3k=3)0.0261.8899.870.0054.700.71 ± 0.15
IMPRINT (k=5k=5)2.2963.2096.880.8655.991.56 ± 0.21

Alternating native recomputation and cached epochs (k=2k=2) matches the native mean Avg. Gap of 0.480.48; k=3k=3 and k=5k=5 increase it to 0.710.71 and 1.561.56. At k=5k=5, forgetting weakens despite relatively high retain performance. One possible explanation is that outdated weights become less aligned with the model's current imprint. These results support periodic recalibration.

We have clarified the weight-update frequency in Section 3.5 and added the stability diagnostics and refresh-frequency comparison to Appendix C.5 (Table 30).

Thank you again for your time and effort. We would be happy to address any remaining questions or concerns the reviewer might have.

Reviewer_MoT6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies post-hoc unlearning for models trained with rehearsal-based class-incremental learning. The authors observe that CIL leaves samples with uneven imprints: older-task samples tend to have small or negative logit margins, while recent-task samples are strongly encoded. They argue that uniform forget losses therefore over-perturb the model on weakly encoded samples and hurt retention. The proposed method, IMPRINT, scales each forget sample's loss by ωi=σ(mi/τ)\omega_i = \sigma(m_i/\tau), where mim_i is the current logit margin, and is applied on top of RL, SalUn, AMUN and BS. Experiments on CIFAR-100, Tiny-ImageNet and ImageNet-1k with ResNet-18 and ViT-B/16, using DER++, FOSTER and iCaRL, show that in the main comparisons IMPRINT gives the smallest average gap to a retrained model, largely by keeping retain accuracy close to the retrained reference. The appendix adds a geometric interpretation of the margin, a lower bound on the parameter change needed to flip a prediction, and several control experiments.

Critical strengths
  1. The proposed method is simple and effective.

  2. The paper has a useful analysis of CIL models.

Critical weaknesses
  1. The source of the gains is not isolated. Appendix B.1 links the margin to decision-level removal effort, but the experiments do not show which part of the weighting produces the improvement. With τ=0.5\tau = 0.5 the initial weights are strongly separated. Table 12 reports a 10th-percentile weight of 0.0004 and a median of 0.968, and a margin of 2 already gives a weight of about 0.98. This suggests that much of the benefit may come from suppressing nonpositive-margin samples rather than from differentiating among strongly imprinted ones. A hard correctness gate 1[mi>0]\mathbb{1}[m_i > 0] would test this and is missing. The temperature study covers only 0.5 and 1. Corollary 1 motivates margin-dependent removal effort under local assumptions, but it does not explain why a quickly saturating sigmoid is the right allocation rule.

    Appendix B.8 also states that gradients propagate through the weights. When ωi\omega_i is not detached, the gradient contains the additional term ℓi ωi(1−ωi) τ−1∇θmi\ell_i\,\omega_i(1-\omega_i)\,\tau^{-1}\nabla_\theta m_i. Taken alone, a gradient-descent step using this term lowers the true-class margin to first order. Its effect goes beyond scaling the base-loss gradient and could contribute to the forgetting gains. The main text does not specify whether the weights are detached or how often they are refreshed. Section 3.5 mentions recomputation at each epoch, while Appendix B.8 records weights per minibatch. A stop-gradient ablation is needed. The mean-matched control in Table 16 does not settle the question. The authors note that it also removes the weight-gradient effect, and Eq. 59 includes an unexplained factor of 0.5, so the coefficients may not actually be matched.

  2. The evaluation does not yet support the broader claims. The main forget classes were chosen to span a wide range of margins, which is the same signal IMPRINT uses. Results with uniformly sampled forget classes are needed to judge how representative the gains are. The random-sample deletion in Table 7 is a different setting and does not answer this. The forgetting claims should also be separated from the retention gains. Against RL, class-wise FA is already near zero, so most of the improvement comes from retention. On the homogeneous low-margin forget set in Table 8, adding IMPRINT to RL raises FA from 0.21 to 3.43 and lowers MIA from 99.95 to 94.53, even though the average gap improves.

    Several baselines have very low retain accuracy. For example, SCRUB obtains RA of 3.86 on CIFAR-100 at buffer size 500 and 0.14 on ImageNet-1k. Appendix A.1 states that the baselines use the hyperparameters from their original papers, with learning rates of 0.01 or 0.001, except for GA with ViT, which uses 10−510^{-5}. It does not report any search for this CIL setting. This weakens the broad comparison against ten baselines, although the paired comparisons with each base method remain informative.

    Finally, a nonpositive margin indicates weak current decision support, not that the sample's influence has been removed. Appendix B.1 reports a median weight of 0.0053 for the initially nonpositive-margin group in the analyzed setting, so this distinction matters for the forgetting claims. The paper does not define the MIA attack or its inputs. The reported value is also close to its ceiling for several methods that already reach near-zero FA, which limits its ability to separate them. Appendix E.1 shows that IMPRINT reduces forgetting relapse after subsequent learning compared with RL and SalUn, but some relapse remains. A clearer MIA protocol and a separate comparison of initially low-weight samples against the retrained model would strengthen the claims.

Questions for authors
  1. Appendix B.8 states that gradients propagate through the weights. Does this hold for all main experiments and base methods, and how often are the weights refreshed? Please compare with an otherwise identical variant that applies a stop-gradient to ωi\omega_i.
  2. How does a hard correctness gate 1[mi>0]\mathbb{1}[m_i > 0] compare with IMPRINT under the same optimization settings? Does a wider temperature sweep show a benefit from continuous weighting beyond correctness-based suppression?
  3. Why does Eq. 59 include the factor 0.5? Is a corresponding normalization applied in the IMPRINT branch?
  4. What are the results with uniformly sampled forget classes? How were the learning rates and τ\tau selected, and which learning rate is used for BS+IMPRINT?
8 comments
AuthorsOpenReview ↗

Response to Reviewer MoT6 (1/4)

We sincerely thank the reviewer for the time, effort, and valuable feedback. We clarify the implementation and report additional controlled evaluations below.


W1 & Q1, Q3. Gradient propagation, loss coefficients, and Equation 59

Native IMPRINT recomputes margins and weights from the current model at every forget minibatch, with gradients propagating through the weights. We use this procedure in all native IMPRINT runs, regardless of the base unlearning method. The stop-gradient variant recomputes weights from its own evolving model using the same rule, then detaches them.

The implemented RL-based objective applies the same coefficient α\alpha to the forget loss and retain cross-entropy.

\alpha\mathcal L_{\mathrm{forget}}(\tilde\theta_u) + \alpha\mathcal L_{\mathrm{retain,CE}}(\tilde\theta_u) + \beta\mathcal L_{\mathrm{retain,MSE}}(\tilde\theta_u) \right].$$ The forget loss uses the corresponding weighting rule. The retain MSE term is included when the replay buffer stores logits (Appendix A.1). The shared coefficient $\alpha$ is inherited from the replay-buffer cross-entropy coefficient in the corresponding CIL setting and remains fixed during unlearning. In the mean-matched comparison, both IMPRINT and the control use $\alpha=0.5$. At each minibatch step, the control assigns every forget sample the mean weight $\bar{\omega}_t$ recorded from IMPRINT for that step. Both variants divide the weighted forget-loss sum by the minibatch size, giving the same minibatch-average forward forget coefficient, $0.5\bar{\omega}_t$. The mean-matched forget-loss term appears in Eq. (59) of the submitted manuscript and Eq. (63) of the revision (Appendix B.9). We have made the shared coefficient explicit in Section 3.4 and clarified weight recomputation and gradient flow in Section 3.5. --- ### W1 & Q1. Stop-gradient and mean-matched controls For the unscaled per-sample forget loss $\ell_i^f$, $$\nabla_\theta(\omega_i\ell_i^f) =\omega_i\nabla_\theta\ell_i^f +\frac{\ell_i^f}{\tau}\omega_i(1-\omega_i)\nabla_\theta m_i.$$ For a nonnegative forget loss, a gradient-descent step using the second term alone produces a nonpositive first-order change in the true-class margin wherever the margin is differentiable. It therefore contributes beyond scaling the base-loss gradient. We compare the RL controls on CIFAR-100, buffer size 2,000, using five paired 30-class deletion configurations as in Appendix B.9. Sigmoid variants use the default $\tau=0.5$. All variants share the deletion requests and optimization settings. **Table: Stop-gradient and mean-matched comparisons with RL.** |Method|FA↓|RA↑|MIA↑|FT↓|RT↑|Avg.Gap↓| |-|-:|-:|-:|-:|-:|-:| |Retrain|0.00|62.14|100.00|0.00|55.35|-| |RL|0.18|45.27|99.98|1.23|52.65|4.20±1.05| |RL+Stop-gradient IMPRINT|5.68|62.26|97.18|4.20|56.67|3.01±0.28| |**RL+IMPRINT**|**0.12**|61.36|99.66|**0.04**|56.40|**0.69±0.44**| |RL+Mean-matched uniform|43.56|63.15|68.88|39.55|56.45|23.46±1.28| Detached weighting improves retention over RL, while full IMPRINT achieves lower FA and FT and a smaller mean gap. The stop-gradient comparison identifies differentiation through the weights as a contributor to the improved forgetting–retention tradeoff. The mean-matched uniform control also yields a larger gap despite matching the average forget-loss coefficient. We have added the gradient derivation to Appendix B.1 and the stop-gradient and mean-matched comparisons to Appendix B.9 (Table 20). --- ### W1 & Q2. Hard correctness gating and RL temperature sensitivity Corollary 1 motivates margin as a local decision-level removal signal. The sigmoid is an empirical bounded, monotone, differentiable choice. Its derivative is largest near the decision boundary, so the benefit need not depend on finely separating strongly positive-margin examples. We compare the hard correctness gate, $\omega_i^{\mathrm{hard}}=\mathbb{1}[m_i>0]$, with IMPRINT for RL and SalUn, on CIFAR-100, buffer size 2,000, using five paired 30-class deletion configurations. The gate is recomputed at every forget minibatch. We additionally sweep $\tau\in\{0.25,0.5,1.0\}$ for RL; SalUn uses the default $\tau=0.5$. **Table: Hard correctness gating for RL and SalUn, and temperature sensitivity for RL.** |Method|FA↓|RA↑|MIA↑|FT↓|RT↑|Avg.Gap↓| |-|-:|-:|-:|-:|-:|-:| |Retrain|0.00|62.14|100.00|0.00|55.35|-| |RL|0.18|45.27|99.98|1.23|52.65|4.20±1.05| |RL+Hard correctness gate|4.86|62.65|96.74|3.40|57.03|2.86±0.28| |**RL+IMPRINT($\tau=0.25$)**|0.21|60.83|99.53|0.11|56.15|**0.66±0.42**| |RL+IMPRINT($\tau=0.50$)|0.12|61.36|99.66|0.04|56.40|0.69±0.44| |RL+IMPRINT($\tau=1.00$)|0.09|59.97|99.75|0.01|55.56|0.86±0.64| |SalUn|0.91|49.83|99.85|3.75|53.58|3.78±0.87| |SalUn+Hard correctness gate|6.00|62.92|96.14|4.13|57.14|3.38±0.34| |**SalUn+IMPRINT($\tau=0.50$)**|**0.25**|61.40|99.38|**0.09**|56.30|**0.86±0.78**|
AuthorsOpenReview ↗

Response to Reviewer MoT6 (2/4)

For both RL and SalUn, IMPRINT preserves most of the retention gains from hard correctness gating while achieving lower FA and FT and a smaller average gap to retraining. This pattern holds across all tested RL temperatures. The reported variability leaves the ranking among temperatures uncertain, and the default temperature performs well in this setting. The RL stop-gradient ablation shows that differentiating through the sigmoid weights contributes to this trade-off, reducing FA, FT, and the average gap at similar retention levels (Appendix B.9, Table 20). We have added the hard-gate comparison and expanded RL temperature sweep to Appendix B.7 (Table 15).


W2 & Q4: Forget-Class Selection and Baseline Tuning

1. Uniformly sampled forget classes.

We evaluate three 30-class requests sampled uniformly without replacement and independently of margins on one CIFAR-100 source model with ResNet-18, DER++, and buffer size 2,000. Unlearning uses learning rate 0.010.01, batch size 256, and ten epochs.

Table: Uniform class sampling across three deletion requests.

MethodFA↓RA↑MIA↑FT↓RT↑Avg.Gap↓
Retrain0.0064.36100.000.0055.62-
RL2.1557.3299.381.2150.693.19±0.35
RL+IMPRINT0.0262.4099.860.0155.510.48±0.01
SalUn5.9659.1997.643.6752.414.08±0.32
SalUn+IMPRINT0.0362.4399.780.0155.290.54±0.05

IMPRINT improves both forgetting and retention for RL and SalUn, extending the gains to class-deletion requests selected independently of margin. We have added the results for uniformly sampled forget classes to Appendix C.6 (Table 31).

2. Learning rates, loss coefficients, and sigmoid temperature.

During the initial experiments, we considered learning rates in {10−3,10−2,5×10−2,10−1}\{10^{-3},10^{-2},5\times10^{-2},10^{-1}\} and chose the learning rate that has lowest Avg.Gap, treating GA separately because of its sensitivity to higher rates. The remaining optimization settings followed those of the original methods, as described in Appendix A.1. In the CIFAR-100/ResNet-18/DER++ experiments reported above, RL, SalUn, and their IMPRINT variants use a learning rate of 0.010.01, while BS+IMPRINT uses 0.0010.001. As clarified above, α\alpha is inherited from the replay-buffer CE coefficient of the corresponding CIL setting and remains fixed during unlearning.

To further assess learning-rate sensitivity, we evaluate SCRUB, NPO, GA, and BE over {10−5,10−4,10−3,10−2,5×10−2}\{10^{-5},10^{-4},10^{-3},10^{-2},5\times10^{-2}\} using the five paired configurations and ten-epoch budget of Appendix B.9. For each method, we selected the rate that minimizes Avg.Gap.

Table: Comparison with learning-rate-tuned baselines.

MethodLRFA↓RA↑MIA↑FT↓RT↑Avg.Gap↓
Retrain-0.0062.14100.000.0055.35-
SCRUB10−410^{-4}0.3140.1799.470.3643.726.96±1.37
NPO10−410^{-4}3.2658.7994.982.6354.053.20±2.37
GA10−510^{-5}30.6451.9565.4026.3245.9722.23±2.46
BE10−410^{-4}0.8622.3194.370.3020.1316.37±0.58
RL10−210^{-2}0.1845.2799.981.2352.654.20±1.05
RL+IMPRINT10−210^{-2}0.1261.3699.660.0456.400.69±0.44

Smaller learning rates substantially improve SCRUB. RL+IMPRINT retains the smallest mean gap among the configurations shown. We have documented the learning-rate selection and the rate used for BS+IMPRINT in Appendix A.1 and added the tuned-baseline results to Appendix C.7 (Table 32).

We chose τ=0.5\tau=0.5 as the default based on the sensitivity analysis in Appendix C.1 (Table 21 of the revised manuscript). As discussed in our response to Q2, we extended the RL sweep to include τ=0.25\tau=0.25 (Appendix B.7, Table 15).


W2: Forgetting Quality, Low-Weight Samples, and MIA

1. Separating forgetting and retention.

We agree that forgetting and retention improvements should be distinguished. When RL already achieves near-zero FA, IMPRINT’s benefits are primarily reflected in retention. However, the additional evaluations also demonstrate improvements in forgetting itself. Table 4 shows that IMPRINT improves both forgetting and retention when applied to AMUN and BS. In the sequential Task 9 → Task 0 unlearning experiment in Appendix E.2 (Table 44), RL+IMPRINT improves forgetting while maintaining retained-data performance. The uniformly sampled class-deletion results above further demonstrate joint improvements when class selection is independent of margin.

Table 9 deliberately examines a challenging low-imprint regime by selecting the 30 lowest-margin classes. We include this stress test to investigate the limits of margin-based allocation when the deletion targets already have weak decision support. In this setting, IMPRINT improves retention while worsening FA, FT, and MIA, so the smaller aggregate gap reflects a forgetting–retention tradeoff.

AuthorsOpenReview ↗

Response to Reviewer MoT6 (3/4)

Together, these experiments characterize IMPRINT across different deletion regimes: retention gains when baseline forgetting is already near its evaluation floor, joint forgetting and retention improvements in multiple settings, and a tradeoff in the deliberately constructed low-imprint stress test. We have clarified the forgetting–retention trade-off for the homogeneous low-margin forget set in Section 4.3 (Table 9).

2. Initially low-weight samples versus retraining.

A nonpositive margin indicates weak current decision support, not complete removal of historical training influence. We therefore evaluate a subgroup fixed before unlearning:

ωi(0)=σ(mi(0)/0.5)≤0.01\omega_i^{(0)}=\sigma(m_i^{(0)}/0.5)\le0.01

This criterion selects strongly negative-margin examples. We use the same three uniform-class requests, each containing 13,500 evaluated forget examples; subgroup sizes are reported below. The subgroup is fixed using the source model and used only for evaluation. Unlearning uses the full forget set.

For each initially low-weight forget example, we compare each model’s predictions with those of the model retrained for the same deletion request. The absolute cross-entropy (CE) gap measures the absolute difference in loss on the true label, and the confidence gap measures the absolute difference in the probability assigned to that label, in percentage points. Total variation (TV) and Jensen–Shannon (JS) divergence compare the full vectors of predicted class probabilities. TV is half the sum of the absolute differences in class probabilities. JS measures how much the two distributions differ by comparing each with their average, using natural logarithms. We average each metric over examples within a request and then weight all requests equally.

Prediction discrepancies relative to matched retraining for initially low-weight samples

MethodAbsolute CE gap↓True-label confidence gap(pp)↓Total variation↓Jensen–Shannon divergence↓
Source model3.43±0.241.41±0.050.59±0.010.31±0.01
RL2.87±0.270.42±0.010.73±0.020.38±0.02
RL+IMPRINT1.91±0.100.20±0.020.59±0.030.30±0.02
SalUn3.12±0.270.58±0.030.71±0.010.36±0.01
SalUn+IMPRINT1.90±0.110.20±0.010.59±0.030.29±0.02

Adding IMPRINT to either RL or SalUn brings predictions closer to matched retraining on all four metrics, with residual differences on each. Both IMPRINT variants also have smaller CE and true-label confidence gaps than the source model, while total variation is similar to the source model’s.

3. MIA protocol and subgroup evaluation.

We use the confidence-based SVM membership-inference evaluator from SalUn’s official repository. For each evaluated model, we fit the SVM using true-label softmax probabilities from retained training examples labeled as members and test examples labeled as nonmembers. The reported MIA ratio is the percentage of forget-set training examples classified as nonmembers. This ratio can saturate, limiting its ability to distinguish methods that already achieve near-zero forget accuracy.

We evaluate how well true-label confidence distinguishes initially low-weight forget-set training examples (members) from official-test examples (nonmembers). We apply the same source-model low-weight criterion (ωi(0)≤0.01\omega_i^{(0)} \le 0.01) to both groups. Membership is the positive label, and higher true-label confidence indicates stronger evidence of membership.

Before filtering, we split the official test set into disjoint calibration and final-evaluation sets of 2,000 and 8,000 examples, respectively. For each 30-class deletion request, we restrict both splits to the requested classes and apply the source-model low-weight criterion. A class qualifies for the AUC analysis if it has at least one initially low-weight forget-set member and at least one initially low-weight nonmember in each split. We compute AUC using only confidence scores from forget-set members and final-evaluation nonmembers in qualifying classes. All 30 classes qualify in request 1. In each of requests 2 and 3, 28 classes qualify because two classes have zero initially low-weight forget-set members. The reported member counts therefore also apply to the prediction-discrepancy analysis.

Filtered sample counts for the AUC analysis by deletion request

RequestForget-set membersCalibration nonmembersFinal nonmembersClasses represented in the low-weight evaluation
12,68416059430
22,52016357528
32,39713857728

For each request, we pool the eligible classes and compute one weighted ROC AUC using raw true-label softmax probabilities for forget-set members and final-evaluation nonmembers. Within each group, we divide each class’s total weight equally among its examples, giving all classes equal total weight. We compute each request’s absolute AUC gap to matched retraining using unrounded values, then average the AUCs and gaps equally across the three requests.

Reviewer_VuXJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces IMPRINT, a plug-and-play framework for machine unlearning in class-incremental learning (CIL). IMPRINT estimates each sample's current imprint using its logit margin and transforms it into an adaptive forgetting weight via a sigmoid function. These weights scale the forget loss, assigning larger weights to samples whose ground-truth class retains stronger support. The method can be integrated into existing unlearning approaches, and the paper compares it against ten baselines, including random labeling, SalUn, SCRUB, and AMUN. The evaluation covers CIFAR-100, Tiny-ImageNet, and ImageNet-1k across several CIL algorithms.

Critical strengths

The method is straightforward to implement and demonstrates compatibility in integration with several existing unlearning methods. The evaluation results cover comparisons against several established unlearning baselines, involve multiple datasets and model architectures, as well as several CIL algorithms. The paper also examines alternative weighting signals, heterogeneous versus homogeneous forget sets, sensitivity to training budgets, and behavior after subsequent learning.

Critical weaknesses

(1) The authors should further justify their problem formulation for machine unlearning in class-incremental learning (CIL). While the forget set can theoretically comprise any previously observed data, actual access in CIL is often restricted to the subset of forget samples preserved in the replay buffer. Notably, multiple existing unlearning approaches operate effectively on retained data alone without requiring explicit forget samples. (2) Expanding on this limitation, what would be the performance of IMPRINT in standard (non-CIL) unlearning scenarios? The challenge of sample non-uniformity appears specific to CIL environments, and it becomes especially pronounced when assuming access to forget samples that are absent from the replay buffer. (3) In Line 184, the authors mention that class accuracy does not capture how strongly individual samples are encoded in the model. Furthermore, in Figure 5, they compare different influence proxies, including accuracy and logit margin. However, in Figure 3, the authors validate the logit margin by showing its strong positive correlation with class accuracy. This circular logic undermines the justification for using logit margins; the authors need to provide an independent metric to properly validate this choice. (4) In Table 12, the authors provide an interpretation for the imprint scale associated with each sample. However, it seems that IMPRINT simply assigns binary weights (0 or 1), meaning the framing that stronger updates are assigned to strongly imprinted samples does not reflect what happens in practice. In particular, the authors need to compare IMPRINT with a simple baseline where the sigmoid is replaced with a step function. That is, the model would assign a weight of one to forget samples that currently have a positive margin (i.e., those that are still correctly classified) and a weight of zero to samples that are already misclassified. This setup is equivalent to setting τ→0\tau \rightarrow 0 in the sigmoid.

Questions for authors

Please see the weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer VuXJ (1/3)

We sincerely thank the reviewer for the time, effort, and valuable feedback.


W1. Unlearning under buffer-restricted deletion

We appreciate the comment. We first clarify the rationale for our problem setup, which assumes full access to the forget set, and then present additional experiments under restricted access to forget data.

Our experimental setup reflects a practical scenario in which a service provider continually updates its model using data received from users. As new data arrive, the provider performs continual learning while retaining only a limited subset of past training samples in a replay buffer, rather than storing all previously received data. Users, however, may still retain the original data they provided. If a user later wishes to have some of these data unlearned, they can resupply the corresponding samples as part of their deletion request. The full forget set is therefore available at unlearning time, even if some or all of these samples are no longer stored in the replay buffer. In contrast, access to retain data remains restricted to the samples preserved in the replay buffer. In this setting, the replay buffer limits which past data the provider stores, but does not determine which samples users can request to have unlearned.

Following the reviewer's suggestion, we further investigate a more restrictive setting in which users request the deletion of their previously provided data without resupplying the original samples. For example, users may no longer possess their original data but still wish to have their contributions to the model unlearned. To examine this scenario, we conducted additional experiments in a buffer-restricted deletion setting, where only forget samples retained in the replay buffer are accessible during unlearning.

We evaluated three forgetting configurations on CIFAR-100 with a replay buffer of 2,000 samples, using a learning rate of 0.01 for RL, SalUn, and their IMPRINT-augmented variants. We additionally evaluated two retain-data-only unlearning approaches: fine-tuning (FT) and L1-sparse fine-tuning (FT+L1), following the sparsity-aware unlearning approach introduced in [1] (NeurIPS 2023). Both methods also used a learning rate of 0.01. For FT+L1, we selected λL1\lambda_{\mathrm{L1}} = 0.001 from {10⁻⁶, 10⁻⁵, 10⁻⁴, 10⁻³, 10⁻², 0.2} using validation performance and linearly decreased it to 0.0001 over ten epochs, where λL1\lambda_{\mathrm{L1}} controls the strength of the sparsity-inducing penalty on model parameters during fine-tuning. In the buffer-restricted setting, SalUn mask construction, margin computation, and forget updates use only forget samples retained in the replay buffer. FA and MIA, however, are evaluated on the complete original forget set, including samples no longer stored in the buffer.

Forget-data accessMethodFA ↓RA ↑MIA ↑FT ↓RT ↑AvgGap ↓
Retrain reference—0.0064.36100.000.0055.620.00
Buffer onlyRL6.5162.4294.894.2155.393.62 ± 1.30
Buffer onlyRL+IMPRINT0.1862.4199.600.1356.060.62 ± 0.18
Buffer onlySalUn10.5462.1491.027.4954.985.97 ± 2.23
Buffer onlySalUn+IMPRINT0.5461.1298.920.3654.551.26 ± 0.56
Full forget setRL2.1557.3299.381.2150.693.19 ± 0.35
Full forget setRL+IMPRINT0.0262.4099.860.0155.510.48 ± 0.01
Full forget setSalUn5.9659.1997.643.6752.414.08 ± 0.32
Full forget setSalUn+IMPRINT0.0362.4399.780.0155.290.54 ± 0.05
Buffer retain data onlyFT56.3158.4251.6748.1450.6232.74 ± 1.92
Buffer retain data onlyFT+L1 (λL1=0.001\lambda_{\mathrm{L1}}=0.001)0.0031.22100.000.0028.8611.98 ± 5.40

The results show that FT preserves relatively high retain accuracy but fails to sufficiently remove the influence of the forget samples, resulting in poor forgetting performance. In contrast, FT+L1 achieves near-zero FA and FT, but substantially degrades RA and RT compared with the retraining reference. These results highlight the difficulty of balancing forgetting and retention when unlearning relies solely on the limited retain data available in the replay buffer. In contrast, IMPRINT consistently achieves a substantially smaller gap to retraining in both data-access settings compared to existing baselines. These results demonstrate that IMPRINT remains effective even when only forget samples preserved in the replay buffer are accessible, supporting its applicability to the more restrictive setting suggested by the reviewer.

We have clarified the data-access assumptions in Sections 3.1 and 3.5 and added the buffer-restricted deletion results and retain-only baselines to Appendix F.1 (Table 45).

AuthorsOpenReview ↗

Response to Reviewer VuXJ (2/3)

W2. Performance of IMPRINT in standard unlearning scenarios

Following the reviewer’s suggestion to examine IMPRINT beyond CIL, we conducted additional experiments on a jointly trained CIFAR-100 model across five forgetting configurations, with full access to the forget and retain sets during unlearning. For the additional baselines, we searched method-specific learning rates and selected them based on validation performance. The results are summarized below.

MethodLRFA ↓RA ↑MIA ↑FT ↓RT ↑AvgGap ↓
Retrain—0.00099.979100.0000.00077.5610.000
RL0.019.55499.914100.0004.99276.4213.150 ± 0.600
RL+IMPRINT0.010.00099.975100.0000.00078.3180.153 ± 0.097
SalUn0.017.44499.895100.0003.90876.2182.556 ± 0.297
SalUn+IMPRINT0.010.00099.974100.0000.00078.3460.199 ± 0.108
NPO0.000016.62495.60599.0073.70871.4824.355 ± 0.865
SCRUB0.00112.83699.97199.2803.39278.0043.497 ± 1.019
BE0.00123.03094.69792.39610.78367.54311.344 ± 0.725

The results show that IMPRINT also provides benefits in standard unlearning. Adding IMPRINT to RL and SalUn reduces both FA and FT to zero while maintaining high retain accuracy, reducing the average gap to retraining from 3.150 to 0.153 and from 2.556 to 0.199, respectively.

These results suggest that sample-wise weighting can be useful beyond CIL: even in a jointly trained model, samples may not have identical logit margins and require different forgetting intensity. CIL introduces additional sources of heterogeneity through sequential learning and limited replay. Samples are introduced at different stages, and only a subset is revisited through the replay buffer, resulting in unequal learning and replay histories. As shown in our response to W1, IMPRINT provides consistent gains both when the full forget set is available and when access is restricted to forget samples in the replay buffer. Its benefits therefore do not depend on access to forget samples outside the buffer.

Our focus on CIL is motivated by the heterogeneous sample imprints induced by sequential learning and limited replay, together with restricted access to retain data during unlearning. The joint-training results demonstrate broader applicability, while CIL remains our primary setting for studying how these constraints affect sample-adaptive unlearning.

We have added the standard unlearning results on jointly trained CIFAR-100 to Appendix F.2 (Table 46) and summarized them in Section 4.3.


W3. Validation of logit margin as a sample-wise imprint proxy

We thank the reviewer for this comment. Our intention is not to dismiss accuracy as a meaningful indicator of retained knowledge, but to highlight its limited granularity for sample-wise weighting. For sample-wise weighting, using class accuracy assigns the same aggregate score to all samples within a class, whereas using individual-sample correctness provides only a binary signal (correct or incorrect). Neither distinguishes, for example, a barely correct prediction from a confidently correct one. Logit margin provides this finer-grained information by measuring the model’s preference for the ground-truth class over its strongest competitor.

From this perspective, Figure 3 examines whether average logit margin is consistent with accuracy at the class level, while logit margin additionally distinguishes samples within each class. These roles are compatible: a sample-level measure can agree with a meaningful aggregate indicator while providing greater granularity. However, we agree that this correlation alone does not establish its validity as a sample-wise imprint proxy.

For a separate sample-level assessment, Section 3.3 (Table 2) compares logit margin with IMU’s gradient-based influence estimate, computed using final-classifier gradients and a damped diagonal-Fisher approximation. Their strong rank correlation (Spearman’s ρ = 0.728) supports the relevance of logit margin to sample-level influence. Figure 5 further demonstrates its practical utility for weighting unlearning updates, yielding a smaller average gap to retraining than accuracy- or entropy-based weighting.

We have clarified the role of class accuracy in Section 3.3. Appendix B.1 reports a complementary analysis of the effort required to cross the decision boundary.

AuthorsOpenReview ↗

Response to Reviewer VuXJ (3/3)

W4. Comparison of IMPRINT with binary correctness gating

To directly address this concern, we conducted the comparison suggested by the reviewer, replacing the sigmoid weighting with a dynamically updated binary correctness gate while keeping all other unlearning settings unchanged. The binary gate is defined as

$ \omega_i^{\mathrm{binary}}=\mathbb{1}[m_i>0], $

where correctly classified forget samples receive a weight of one and misclassified samples receive a weight of zero.

Importantly, the binary correctness gate is also derived from the logit margin, as the sign of the margin determines whether a sample is correctly classified. As noted by the reviewer, binary gating can be viewed as the limiting case of the proposed sigmoid weighting for nonzero logit margins as τ→0+\tau \to 0^+, where forgetting intensity is determined solely by prediction correctness. In contrast, IMPRINT uses the magnitude as well as the sign of the logit margin, assigning intermediate weights to samples near the decision boundary rather than treating all positive-margin or all non-positive-margin samples identically.

We evaluated both weighting schemes across five CIFAR-100 forgetting configurations, using a replay buffer of 2,000 samples and a learning rate of 0.01, while keeping all other optimization settings identical.

MethodFA ↓RA ↑MIA ↑FT ↓RT ↑AvgGap ↓
Retrain0.00062.138100.0000.00055.3460.000
RL0.17945.27499.9781.23352.6494.199 ± 1.050
RL+IMPRINT0.12161.36399.6560.04056.4000.692 ± 0.444
RL+Binary correctness gate4.86262.64896.7413.40057.0262.856 ± 0.276
SalUn0.90849.83399.8493.74753.5803.775 ± 0.870
SalUn+IMPRINT0.24661.39999.3840.08756.3030.862 ± 0.779
SalUn+binary correctness gate5.99762.91596.1384.13357.1433.382 ± 0.344

The results show that binary correctness gating improves retain performance, while having worse forget performance compared to baselines. In contrast, IMPRINT achieves substantially more effective forgetting while improving retention performance close to the retrained model. This improvement is consistently observed when applying IMPRINT to both RL and SalUn, resulting in considerably smaller average gaps to retraining than their corresponding binary-gating variants. These results indicate that continuous sigmoid-based weighting offers a better empirical forgetting–retention trade-off than hard correctness gating.

We have added the hard correctness-gating comparison to Appendix B.7 (Table 15).

[1] Jia, Jinghan, et al. "Model sparsity can simplify machine unlearning." Advances in neural information processing systems 36 (2023): 51584-51605.


We appreciate your time and effort in reviewing our paper. We would be happy to address any remaining questions or concerns the reviewer might have.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comments to AC and all Reviewers

We thank all reviewers for their time and constructive feedback, which helped us improve the manuscript. We have addressed each reviewer’s questions and comments in our individual responses. Below, we summarize the major revisions made during the rebuttal period.

  1. Data access and standard unlearning (Reviewer VuXJ). We clarified the data-access assumptions in class-incremental learning and added experiments on replay-buffer-restricted deletion, including retain-only baselines, and standard unlearning on jointly trained CIFAR-100 (Sections 3.1, 3.5, and 4.3 and Appendix F.1–F.2).

  2. Logit margin and weighting ablations (Reviewers VuXJ and MoT6). We clarified how the comparison with a gradient-based influence estimate supports the use of logit margin at the sample level. We also clarified how weights are recomputed at each forget minibatch and how gradients flow through them, and added hard correctness-gating, stop-gradient, and mean-matched controls (Sections 3.3–3.5 and Appendices B.7 and B.9).

  3. Forgetting evaluation and baseline tuning (Reviewer MoT6). We added experiments with uniformly sampled forget classes, learning-rate-tuned baselines, and comparisons with matched retraining for initially low-weight samples. We also documented the membership-inference protocol and clarified the forgetting–retention trade-off for the homogeneous low-margin forget set (Section 4.3 and Appendices A.1 and C.6–C.8).

  4. Broader experimental coverage (Reviewer Gmkd). We moved the ImageNet-1k plug-and-play comparison into the main text and added experiments with ViT and DER++ on CIFAR-100 and random sample forgetting on Tiny-ImageNet (Section 4.2 and Appendix D.1–D.2).

  5. Robustness and runtime (Reviewers Gmkd and MoT6). We expanded the temperature sweeps across datasets and backbones, analyzed class-wise logit-scale variation, and reported retain-set reuse, runtime, and the effects of less frequent weight updates (Appendices B.7–B.8, C.1, and C.4–C.5).

Please refer to our individual responses for further details. We would be happy to address any remaining questions or concerns.

Best, Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.