acceptodds
Withdrawn from ICLR 2027

Learning with Unlearning: Synaptic Role Partitioning for Simultaneous Continual Learning and Machine Unlearning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual Learning (CL) asks a model to acquire new knowledge without forgetting what it already knows. Machine Unlearning (MU) asks it to destroy specific knowledge without damaging the rest. The challenge is having them in the same system simultaneously; the two objectives pull on the same weights, and an edit made for one harms the other. To address this, we propose Learning with Unlearning (LwU), a unified framework that resolves the conflict at the level where it occurs, the individual synapse. LwU scores every weight twice, once against the data to keep and once against the data to delete, and partitions the network into four zones with separate update rules. Weights that only require retention are frozen. Weights that only require deletion are cleared and returned to the network as free capacity. Contested weights take a gradient-ascent step projected orthogonally to the retention gradient. The remaining capacity learns the incoming task by a privileged teacher via self-distillation. Deletion is served post hoc, and no retraining is required. Across both CL and MU jobs, LwU outperforms or matches the strongest baseline. Its real-world practicality has been demonstrated on the Physical Artificial Intelligence (PAI) dataset we conducted. Additionally, in the image generation task and in removing dangerous content, it outperformed the closest baseline, as shown in Fig.1.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Machine Unlearning, Self Distillation, CLass Incremental Learning, Class deletion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YjDtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the joint problem of continual learning and machine unlearning, where a fixed-capacity model must learn new tasks, retain useful knowledge, and remove the influence of specified past training data. The proposed method, LwU, divides the network parameters into four regions based on their importance to the retain and forget sets, and uses parameter protection and resetting, gradient projection in the conflict region, and teacher distillation to support knowledge retention, deletion, and subsequent learning. Experiments evaluate continual learning and unlearning separately on image classification and haptic datasets. The method is also tested for content suppression in diffusion models, but this extension does not include continual learning experiments.

Critical strengths

The problem setting is relevant to the practical needs of long-term model deployment.

Critical weaknesses

My main concerns are the method’s technical novelty and the limited experimental evidence. Estimating parameter importance from gradients and using it to select and partition parameters is already common practice. The paper needs to explain what the four-region design adds to existing importance-based methods and how it improves the joint handling of continual learning and machine unlearning.

Most experiments evaluate the two separately, and the joint experiments are largely limited to a small-scale Task-IL setting on Split CIFAR-10. This is not enough to support the main claims of the framework. I would like to see comparisons with existing joint methods and reasonable combinations of CL and MU methods across more varied sequences of interleaved learning and deletion, to determine whether the unified design offers a practical advantage.

Questions for authors

Please refer to the weakness.

3 comments
AuthorsOpenReview ↗

Response (1/4)

We thank the reviewer. We respectfully disagree that the experimental evidence is limited.

LwU is a single unified method, and every experiment below evaluates that same model. We compare it against 15 baselines (EWC, SI, LwF, SpaceNet, WSN, NISPA, PEC, SNV, Amnesiac, UNSIR, SSD, SCRUB, Bad Teacher, SalUn, and the unified CL–MU method UniCLUN) and 5 reference models (joint training, SGD, Original, Fine-tune, Retrain), on CIFAR-10, CIFAR-100, Tiny-ImageNet, our real-world haptic dataset RTI, and Stable Diffusion v1.4. The unified baseline UniCLUN is evaluated on all four classification datasets. LwU is the only method that places on both overall podiums (Tab. 1 and Tab. 2).

The submission is 57 pages, most of it experiments:

BlockWhereScope
CL benchmarkTab. 1Seq-CIFAR-100, Seq-Tiny-ImageNet, RTI; Task-IL and Class-IL; ACC, BWT, FWT, PS; 9 baselines plus joint and SGD bounds; GTEP protocol, 3 task orderings
MU benchmarkTab. 2, 19–24CIFAR-10, CIFAR-100, RTI; CNN and ResNet-18; class, subclass and instance deletion; 7 baselines plus Original, Fine-tune, Retrain; 3 seeds
Per-example privacyTab. 2, Fig. 6–11U-LiRA, 16 shadow models per cell, AUC and TPR at 1% FPR, utility vs. leakage
Joint CL+MUTab. 5interleaved deletion, 10 component variants, 4 retain-subset sizes, 4 deletion schedules including two requests
Component ablationsTab. 6–10, 135 settings across 3 granularities, 2 backbones, 2 datasets
SensitivityTab. 11–12dominance margin, teacher masking, retain-subset size, Zone-C budget and step size
Importance estimatorTab. 14–15random, magnitude, single Fisher (retain, forget), dual Fisher
Mechanism checksTab. 16–17, Fig. 5measured orthogonality of the applied Zone-C step; test-time memory under 3 stream orders; final zone sizes
CostFig. 12, Tab. 25–28unlearning wall-clock; CL training time, GPU-hours, energy, memory, storage, inference
Generative transferTab. 3–4, Fig. 1, 3, 4Stable Diffusion v1.4 nudity suppression, held-out prompts, multiple seeds, paired bootstrap CIs
QualitativeFig. 13–22per-class removal on CIFAR-10
TuningTab. 18, 29–37validation-only grids for every method, reported in full

"Limited" is not a fair description of this. However, as reviewers asked, we are running the same ablation in the paper on CIFAR-100 and will publish it here before the discussion phase. We will add the following ASAP: the interleaved protocol on Seq-CIFAR-100 and RTI, Task-IL and Class-IL, against UniCLUN and composed pipelines (LwF + SSD, LwF + SalUn, EWC + SCRUB), with 1–3 deletion requests at different stream positions. Class-IL was left out of Tab. 5 because on 2-class Split CIFAR-10 old tasks fall to 0% without replay; on Seq-CIFAR-100 LwU keeps 33.2% Class-IL accuracy (Tab. 1), so deletion is measurable there.

AuthorsOpenReview ↗

Response (2/4)

"Estimating parameter importance from gradients is common practice"

We agree, and we use the simplest estimator, diagonal Fisher, on purpose. The contribution is what is built on it. Existing importance-based methods compute one score to make one binary decision: protect or free in CL (EWC, SI, WSN, NISPA), dampen or mask in MU (SSD, SalUn). LwU contributes:

  1. A synapse-level formulation of the joint problem. Retained and forgotten knowledge share parameters, so the CL–MU conflict must be resolved per weight, not per task with buffers (UniCLUN) or per module (CLPU).
  2. Dual scoring with dominance, giving four zones and four update rules: freeze (A), reset (B), retain-orthogonal ascent (C), plastic learning (D). One estimator, two scores, four actions.
  3. Coupled deletion and learning. Reset Zone B weights rejoin the plastic pool, so deletion frees capacity for the next task, and the next task overwrites what the reset leaves. A hard reset is too destructive for standalone MU and has no role in CL; only the joint setting makes it affordable.
  4. Retain-orthogonal ascent on contested weights, with a first-order guarantee on the retain loss and no forget–retain trade-off coefficient. The measured cosine to the retain gradient is at most 3e-7 (Tab. 16).
  5. One masked privileged teacher for both objectives. It preserves old classes during learning and repairs retained behavior after deletion, and because it is conditioned on retained classes it never supplies a target on the forget set.
  6. Buffer-free throughout. For unlearning this matters: a replay buffer can hold the very data a user later asks to delete.

Every component is ablated

  • Partition vs. estimator (Tab. 14–15). With everything else fixed, a single Fisher map leaves forget accuracy at 42.4% (CNN) and 78.9% (ResNet-18). Dual scoring with four zones reaches 0.0% on both; the gap to Retrain falls from 16.3 to 1.0 and from 28.5 to 0.8. Random and magnitude scores are worse still. The gain comes from the partition, not the estimator.
  • Against SSD, which also uses diagonal Fisher (Tab. 2): U-LiRA AUC of 50.9 vs. 75.9 (CIFAR-10, ResNet-18) and 51.6 vs. 87.9 (CIFAR-100, ResNet-18).
  • Zone B reset (Tab. 13): removing it raises the gap to Retrain from 1.0 to 13.6, 0.8 to 27.8, and 0.5 to 27.2.
  • Dominance (Tab. 5, 11–12): removing the margin drops retain accuracy to 31.8 (CNN) and 52.5 (ResNet-18); in the joint stream ACC falls from 95.9 to 87.1 and BWT from +0.4 to −9.5.
  • Teacher masking (Tab. 5): with an unmasked teacher the deleted class survives at 95.0% in the joint stream.
  • Components across 5 settings (Tab. 6–10, 13), sensitivity (Tab. 11–12), and test-time memory (Tab. 17).
  • Negative results (App. D): 14 alternatives we tried and rejected, including binary partitioning, forget-only scoring, permanent dampening, and a single weighted retain–forget objective. Each design choice in LwU replaces a failure we measured.

Soundness

The review rates soundness 1 ("significant concerns that do not appear easy to address") but does not identify an incorrect claim, derivation, or experimental flaw; both weaknesses concern novelty and scope. We would be grateful if the reviewer could point to the soundness issue, or reconsider that score.

A request

A Clear Rejection largely decides the outcome regardless of the other reviews. We therefore respectfully ask that it rest on concerns we can act on: (a) which existing importance-based method the reviewer considers equivalent to the four-zone design, and (b) which joint baselines, CL+MU compositions, or deletion schedules would change the assessment. We will run what is named within the discussion period. Please note that the paper already includes 15 baselines and a unified CL+MU baseline in its comparisons.

AuthorsOpenReview ↗

Response (3/4): new results on CIFAR-100

As promised, we ran the joint protocol on CIFAR-100. Setup: 5 tasks × 20 classes, ResNet-18, Task-IL, 3 seeds. One class c_f is deleted after task 3, then tasks 4–5 are learned. F_del is c_f accuracy right after deletion, F_end after task 5.

MethodACC ↑BWT ↑F_del ↓F_end ↓
Retrain (never sees c_f)71.75±0.73−2.95±0.11–0.00±0.00
LwU, no deletion72.65±1.20−2.78±0.32–71.92±13.65
CLPU-DER++65.81±1.25−15.63±1.820.00±0.000.00±0.00
LwU73.65±1.13−1.45±0.670.00±0.000.00±0.00

Varied interleaving. Same stream; the default retain subset is 10%.

SettingACC ↑BWT ↑F_del ↓F_end ↓
Delete after task 273.81±1.09−1.59±0.470.00±0.000.00±0.00
Delete after task 373.65±1.13−1.45±0.670.00±0.000.00±0.00
Delete after task 474.12±1.32−0.82±0.600.00±0.000.00±0.00
Two requests (after 2 and 4)74.28±1.11−1.14±0.880.00±0.000.00±0.00
Retain subset 1%71.80±1.52−3.90±0.980.00±0.000.00±0.00
Retain subset 5%73.35±1.43−1.69±0.720.00±0.000.00±0.00
Retain subset 25%73.75±1.02−1.17±0.870.00±0.000.00±0.00

what the four-zone design adds, measured in the joint stream

The reviewer asked what the four-zone design adds to existing importance-based methods. A protect-or-free method computes one score and makes one decision per weight. LwU adds a dominance margin between two scores, a reset zone, and a teacher that both preserves old classes and repairs retained behavior after deletion. Removing each of them on Split CIFAR-100 (same protocol as 1/3) shows what it contributes.

Continual learning only (no deletion)

VariantACC ↑BWT ↑
w/o self-distillation teacher45.78±1.52−36.72±0.55
w/o cumulative freeze mask72.00±1.48−5.75±1.43
LwU72.65±1.20−2.78±0.32

Joint stream (delete c_f after task 3)

VariantACC ↑BWT ↑F_del ↓F_end ↓
No dominance margin (ω = 1)34.15±0.65−45.08±0.590.00±0.000.00±0.00
w/o self-distillation teacher48.80±2.95−32.91±3.450.00±0.000.00±0.00
w/o teacher repair71.12±1.56−4.54±1.276.05±3.868.41±8.14
w/o Zone B reset73.84±1.05−1.31±0.587.69±10.102.87±4.31
w/o Zone-C repair pass73.01±0.81−2.15±0.560.00±0.000.00±0.00
w/o cumulative freeze mask73.97±0.12−3.30±0.130.00±0.000.00±0.00
LwU73.65±1.13−1.45±0.670.00±0.000.00±0.00

The dominance result matches Split CIFAR-10 (Tab. 5: ACC 95.9 → 87.1), now at 100 classes with 20 per task.

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

This paper presents Learning with Unlearning, a framework that combines continual learning and machine unlearning in a single fixed-capacity model. When processing a forgetting task, it uses two scores with diagonal empirical Fisher information: one is on the forget set, and one is on the retain subset. The two scores separate the weights into four zones: safe retain, pure forget, conflict, and plastic. To forget the task, the model re-initializes the safe retain weights and puts them in the plastic zone. When learning new tasks, only the weights in the plastic zone will be trained. In addition, the method employs masked self-distillation to preserve old knowledge and introduces test-time memory to improve adaptability during the inference stage.

Critical strengths
  1. The motivation is clear and practically important: in a deployed system, deletion requests and new tasks arise in an interleaved manner, and the parameters affected by a deletion may already be shared with tasks learned later.

  2. Compared with a single global retain–forget objective, the four-zone partition enables different update rules for different parameters. The Zone C projected-gradient mechanism is particularly interesting because it explicitly accounts for the interaction between the retain and forget gradients.

  3. If released, the RTI dataset would be a valuable step beyond CIFAR-style benchmarks. Moreover, the finding that no existing method achieves instance-level unlearning on RTI is both an honest and an interesting negative result.

Critical weaknesses
  1. The joint CL+MU setting, which is the paper’s main motivation, remains under-evaluated. The introduction promises comparisons with joint CL–MU frameworks over long sequences of interleaved learning and deletion. However, the only interleaved experiment (Table 5: Split CIFAR-10, 5 tasks, one or two deletions) reports only LwU variants. UniCLUN appears only in pure CL (Table 1, without deletions) and standalone MU (Table 2), and CLPU-DER++ is omitted from the comparison.

  2. The claim that “forgetting buys plasticity” is tested empirically: with Zone B accounting for only 0.06%, the recycled capacity appears negligible, and the two-deletion result in Table 5 is within one standard deviation of the single-deletion result.

  3. There is a tension with the “buffer-free” claim. The paper defines buffer-free as storing no exemplars from any previous task at any point, yet every deletion requires a labeled retain subset D’_r (10% of D_r by default), which is used for Fisher scoring, the Zone C projection, and teacher repair. In the CL+MU setting, D_k is discarded after each task, so D’_r must consist of previously seen data from earlier tasks. It is not stated which tasks D’_r is drawn from in Table 5, and the repair step on D’_r may effectively act as replay.

Questions for authors
  1. Which tasks is D’_r drawn from in the CL+MU experiments? How is D’_r obtained without storing previous data? And how much of Table 5’s post-deletion BWT improvement comes from training on D’_r during repair?

  2. Can you provide comparisons with other CL+MU baselines under the same interleaved learning-and-deletion setting, with a longer task sequence and multiple deletions?

  3. Can you provide any evidence supporting the capacity-recycling benefit or real-time applicability, given that Zone B contains only about 0.06% of the parameters and RTI inference exceeds the stated 1 ms control period?

5 comments
AuthorsOpenReview ↗

Response 1/2

We thank the reviewer for the careful reading and for questions we could answer with new experiments.

In short:

  • Q2: five joint baselines under interleaved deletion, with 1 and 3 requests, on a 5-task and a 20-task stream.
  • Q1: with D′_r drawn only from data available at the request (no image of earlier tasks), deletion still drives the class to 0 and costs no accuracy.
  • Q3: a sync-free, compiled test-time memory runs at 0.57 ms per sample at batch size 1, the backbone's own latency.
  • W2: we agree and narrow the capacity claim.

One correction to the summary: LwU re-initializes Zone B (pure forget) and returns it to the plastic pool. Zone A (safe retain) stays frozen.

Setup. Split CIFAR-100 (5 tasks × 20 classes, ResNet-18, Task-IL), 3 seeds. The first class of task 1, c_f, is deleted after task 3, then tasks 4–5 are learned. F_del is c_f accuracy right after deletion, F_end after task 5. In the composed pipelines, the CL method learns tasks 1–3, the MU method edits the model with D_f and the same D′_r that LwU receives, and the CL method resumes with LwF's teacher and EWC's anchor reset to the edited model, so they cannot restore c_f. Every unlearning step and every CLPU setting was selected on validation with LwU's criterion and a grid no larger than LwU's.

MethodStored images at the requestACC ↑BWT ↑F_del ↓F_end ↓
Retrain (never sees c_f)–71.75±0.73−2.95±0.11–0.00±0.00
LwF + SSD†D′_r (2,480)64.53±9.76−13.68±13.8230.89±36.5932.82±38.22
LwF + SalUnD′_r (2,480)69.94±2.45−7.24±0.834.42±3.842.26±2.68
EWC + SCRUBD′_r (2,480)66.90±0.890.10±0.9268.87±25.7565.54±29.65
UniCLUNbuffer (1,000) + D′_r49.62±10.16−26.19±6.2528.12±23.4525.74±43.23
CLPU-DER++‡buffer (600) + model copy65.81±1.25−15.63±1.820.00±0.000.00±0.00
LwUD′_r (2,480)73.65±1.13−1.45±0.670.00±0.000.00±0.00
LwU, request-time D′_rnone from tasks 1–272.55±1.51−2.73±0.740.00±0.000.00±0.00
LwU, D′_r = 1%D′_r (248)71.80±1.52−3.90±0.980.00±0.000.00±0.00

† SSD needs a Fisher over the full training set, which this protocol never keeps; we use D′_r ∪ D_f. ‡ CLPU-DER++ isolates c_f in its own model copy from task 1, so it must know in advance which data will be deleted.

Among methods that learn without knowing which class will be deleted, LwU is the only one that drives c_f accuracy to 0, matching Retrain, without losing accuracy. The composed pipelines receive the same 2,480 images and do worse: the best.

Longer stream (Q2). Split CIFAR-100, 20 tasks × 5 classes, requests after tasks 5, 10 and 15. No method was reselected for 20 tasks, LwU included; all reuse their 5-task validation settings.

MethodKnows requests in advanceACC ↑BWT ↑F_del r1 / r2 / r3 ↓F_end r1 / r2 / r3 ↓
LwF + SalUnno79.24±3.68−7.20±2.028.78 / 17.09 / 28.452.01 / 28.25 / 12.13
CLPU-DER++yes81.22±1.73−6.31±0.580.00 / 0.00 / 0.000.00 / 0.00 / 0.00
LwUno74.71±1.17−7.79±0.510.00 / 0.00 / 3.980.00 / 0.00 / 3.06

We report this table because it qualifies our claim. Deletion still works: LwU drives the class to 0 in 8 of 9 class–seed pairs, while LwF + SalUn, the other method in this table that learns without advance notice, leaves up to 28% of a deleted class. LwU is the only method here that handles requests it was not told about in advance and drives the class to 0.

W3 / Q1: where D′_r comes from. In the joint experiments, D′_r is a uniform random 10% of the retained training images of the tasks learned before the request, drawn when the request arrives (2,480 images for tasks 1–3). It excludes D_f and any previously deleted class. The reviewer is right that these are stored images from earlier tasks. LwU's learning steps store none, and its test-time memory keeps fixed-size weights, not examples. We will revise "buffer-free" to say exactly this, and report D′_r next to every baseline's buffer.

How much is replay? We redrew D′_r from data available at the request only: the task just learned, same size, no image of tasks 1–2.

LwU settingImages of tasks 1–2 storedACC ↑BWT ↑F_del / F_end ↓
No deletionnone72.65±1.20−2.78±0.32–
Deletion, request-time D′_rnone72.55±1.51−2.73±0.740.00 / 0.00
Deletion, D′_r from tasks 1–3 (default)yes73.65±1.13−1.45±0.670.00 / 0.00

Without any old-task image, deletion still drives c_f to 0 on every seed and costs nothing: ACC and BWT match the no-deletion run (72.55 vs 72.65, −2.73 vs −2.78). The extra +1.0 ACC and +1.3 BWT of the default setting are replay on stored images, as the reviewer suspected. We will make the request-time variant the main result and drop any claim that deletion improves BWT. Even this variant beats every composed pipeline that receives 2,480 stored images (72.55 against 69.94 for LwF + SalUn).

W2: capacity recycling. We agree, and we narrow the claim. In these runs Zone B holds 0.01–0.04% of the weights per request, and ACC with one, two and three requests (73.65, 74.28, 74.58) differs by less than one standard deviation, so recycled capacity does not measurably change plasticity at this scale. The reset earns its place through deletion instead. Without it, c_f is not driven to 0 in the joint stream (F_del 7.69 against 0.00, Update 1 to Reviewer YjDt), and 65.33% of the deleted class survives standalone unlearning on ResNet-18, against 0.00% with it (Update 2). We will replace "forgetting buys plasticity" with "deletion returns its weights to the plastic pool instead of consuming capacity."

Q3: real-time. The 1 ms period constrains inference; deletion runs outside the control loop. Batch size 1, RTX PRO 6000, ResNet-18, 32 × 32 inputs, every sample written to memory, GPU otherwise idle:

Settingms per sample
Backbone (= LwU with memory off)0.57
LwU + memory, submitted code1.07
LwU + memory, sync-free (identical outputs)0.77
LwU + memory, sync-free, compiled0.57

Most of the overhead in the submitted code is host–device synchronization from per-batch diagnostic calls. A sync-free version computes the identical update (outputs agree to within 1e-7) and, compiled, runs at the backbone's own latency, inside the 1 ms period. Turning memory off is also an option: it leaves joint-stream accuracy unchanged (73.69 vs 73.65 ACC, F = 0 either way). We will release the sync-free version.

Changes to the paper

  1. A joint CL+MU section with the five baselines above, a storage column, and the 20-task stream.
  2. "Buffer-free" redefined: no stored images for learning; deletion uses D′_r, reported next to every buffer.
  3. The request-time D′_r variant as the main joint result; no claim that deletion improves BWT.
  4. The capacity claim narrowed as above.
  5. A limitation on long-stream continual-learning accuracy.
  6. The latency table and the sync-free test-time memory.

We hope these results address the reviewer's main concerns, and we would be grateful if the reviewer would reconsider the rating.

Reviewer_wseGOpenReview ↗

Response to rebuttal

Thank you for the detailed rebuttal and additional experiments.

First, I apologize for the inaccuracy in my summary regarding Zone A—thank you for clarifying that Zone B (Pure Forget) is reinitialized and returned to the plastic pool.

While the new results address several of my initial concerns, three questions remain:

  1. Long-sequence performance: On the 20-task benchmark, LwU trails LwF+SalUn in ACC (74.71% vs. 79.24%) despite stronger forgetting. Does this indicate an inherent CL–MU trade-off, and what explains this gap?
  2. Class-IL evaluation: Since the new joint CL–MU experiments are limited to Task-IL, can LwU effectively handle interleaved learning and deletion requests in the more challenging Class-IL setting?
  3. Unlearning effectiveness: Beyond deleted-class accuracy, have the authors tested whether forgotten samples remain identifiable via membership inference after multiple learning and unlearning cycles?
Reviewer_rqLKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Learning with Unlearning (LwU), a framework for continual learning and post-hoc deletion within a fixed-capacity network. Separate retain/forget importance scores partition weights into safe-retain, pure-forget, conflicting, and plastic zones. The method resets forget-dominant parameters, applies retain-gradient-orthogonal ascent to conflicting parameters, and uses masked self-distillation to repair retained behavior and guide subsequent learning. A causal test-time memory is also introduced. Experiments evaluate standalone continual learning, class/subclass/instance deletion, a small interleaved learning-deletion protocol, and a haptic tele-manipulation dataset with 20 participants. The appendix additionally studies image-generation concept suppression, component ablations, and privacy attacks.

Critical strengths
  • Supporting learning and selective deletion within the same model is a useful problem, and the four-zone construction gives a concrete implementation of retention, removal, and capacity reuse. The paper includes substantive ablations rather than only full-system results, including retain-subset sensitivity and deletion timing, importance estimation, projection behavior, and test-time memory.
  • The evaluation recognizes that forget-set accuracy alone does not establish unlearning. Retraining references, per-example privacy auditing, and explicit discussion of subclass and RTI failures are valuable. The haptic dataset also exposes a more distinctive setting than standard image classification, although its evidence and governance need clarification.
Critical weaknesses
  • The central privacy evidence is internally inconsistent in ways that materially affect the conclusion. For CIFAR-10 instance deletion with ResNet-18, Table 2 reports LwU U-LiRA AUC 50.90 and TPR@1% FPR of 0.00, whereas Figure 7 reports AUC 77.72 versus 80.08 for the original model and states that instance-deletion performance remains close to the original model. Figure 8 likewise reports substantial remaining CIFAR-100 instance leakage, unlike the near-chance AUCs in Table 2. These may be different attack protocols or checkpoints, but the manuscript does not make their relationship clear. Until the discrepancy is reconciled, I cannot determine whether the claimed privacy advantage is supported.
  • The evidence for simultaneous, sustained CL and MU is substantially weaker than the breadth of the framing. The most relevant interleaved experiment is Table 5 on five Split-CIFAR-10 tasks in Task-IL; its caption explicitly omits Class-IL because old-task accuracy drops to zero without replay. Two deletion requests are considered, but this is not broad validation over long interleaved lifetimes or the motivating participant-level scenario. Standalone CL and MU results cannot by themselves establish that a joint lifecycle preserves both capabilities. This matters especially because the pre-edit teacher can retain deleted-data influence, as Section 2.4 acknowledges, and class masking is unavailable for instance/subclass deletion.
Questions for authors

How does repair avoid transferring deleted-data influence back from the pre-edit teacher? Section 2.4 correctly notes that masking forgotten classes removes their output probability, not their influence on retained-class predictions; for instance and subclass deletion, there is no new class mask at all. Distillation on retained inputs may therefore reintroduce dependencies learned from deleted data even without querying the teacher on the forget set. What empirical evidence shows that repair restores permitted behavior without undoing deletion, especially after subsequent tasks? Comparing the privacy/utility trajectory immediately after the destructive edit, after teacher repair, and after later learning would directly address this concern. I am not treating teacher contamination as a proven failure, but it is an unresolved mechanism question central to the proposed combination.

3 comments
AuthorsOpenReview ↗

Response 1/2

We thank the reviewer for reading the privacy evidence this closely. The reviewer is right about the discrepancy, and we correct it here.

W1. Figs. 7–8 come from our U-LiRA protocol (16 retrained shadow models, seed 42; 4,500 deleted images for instance deletion). Table 2's LwU instance-deletion entries do not match that protocol, and we will replace every U-LiRA entry in Table 2 with numbers from this one protocol:

Instance deletion

SettingOriginal AUCLwU AUCLwU TPR@1% FPR
CIFAR-10, CNN66.7862.528.80
CIFAR-10, ResNet-1880.0877.7220.29
CIFAR-100, CNN81.7876.4815.76
CIFAR-100, ResNet-1893.0584.2338.11

Class deletion, standalone

SettingOriginal AUCLwU AUCLwU TPR@1% FPR
CIFAR-10, CNN100.0073.2831.96
CIFAR-10, ResNet-18100.0053.380.31
CIFAR-100, CNN100.0061.634.67
CIFAR-100, ResNet-18100.0060.640.22

LwU lowers membership leakage relative to the original model in all eight cells, but reaches Retrain level (AUC 50) only in part. We will report all of these numbers and limit the privacy claim to what they show.

Does repair bring deleted-data influence back?

We ran the comparison the reviewer proposes. The same destructive edit is followed by repair with (a) LwU's pre-edit teacher, (b) an oracle teacher that never saw the deleted data (the Retrain model, a diagnostic control only), (c) no teacher (cross-entropy on D′_r), or (d) edit-anchored repair, a new variant described in 3/3. We measure U-LiRA (16 shadow streams per seed, 3 seeds) after the edit, after repair, and after tasks 4–5. AUC 50 means the attack cannot tell the model from Retrain.

With a class mask: class deletion, Split CIFAR-100, the joint protocol of our replies to Reviewers YjDt and wseG.

RepairStageU-LiRA AUCDeleted accRetained acc
nonebefore request83.61±15.9770.1971.61
noneafter edit62.09±6.0712.4769.51
LwU (pre-edit teacher)after repair48.36±5.440.0073.14
LwU (pre-edit teacher)after tasks 4–547.75±5.470.0072.57
Oracle teacherafter repair / end48.27 / 48.540.0073.49 / 72.52
No teacherafter repair / end59.27 / 59.280.00 / 0.5173.33 / 73.19
Edit-anchored repairafter repair / end56.60 / 50.860.0071.64 / 71.61

The pre-edit teacher matches the oracle at both stages, so it does not carry the deleted class back, and later learning does not undo the deletion. The masked teacher is what removes the membership signal: without it, AUC stays at 59. It also blocks fast recovery: one relearning epoch at the end brings the class back to 0% with either teacher, 19% without a teacher, and 54% without repair. Edit-anchored repair (3/3) is meant for deletions without a class mask; with a mask it is slightly worse than standard repair (50.86 vs 47.75 at the end), so we use it only when no class can be masked.

Without a class mask: participant deletion. As a proxy for the participant-level scenario, a "participant" is one fine class inside a kept CIFAR-20 superclass, deleted after task 3, so no output can be masked.

RepairStageU-LiRA AUCParticipant accRetained acc
nonebefore request60.35±6.3175.3782.98
noneafter edit48.06±4.8934.6578.82
LwU (pre-edit teacher)after repair / end61.42 / 58.7175.66 / 73.7683.47 / 83.49
Oracle teacherafter repair / end60.65 / 60.9171.88 / 74.8583.45 / 83.52
No teacherafter repair / end60.87 / 57.4078.26 / 76.6482.99 / 83.66
Edit-anchored repairafter repair / end54.12 / 50.4747.53 / 41.0082.52 / 83.33

The reviewer's concern is real, but the source is not the teacher. Standard repair undoes the deletion: AUC returns to about 60, the pre-request level, and it does so with the oracle teacher too. The cause is repair on retained data that shares the participant's label. Edit-anchored repair ends at AUC 50.5 with retained accuracy unchanged (83.3 vs 83.5).

Standalone instance and subclass deletion (8 cells, seed 42, 16 shadows). The oracle teacher's AUC is within 2 points of the pre-edit teacher's in 6 of 8 cells. In the other two, both ResNet-18 instance deletion methods are 8.3 points lower on CIFAR-10 (70.3 vs 78.6; both recomputed with the current code, against 77.7 in 1/3) and 4.1 points lower on CIFAR-100 (80.1 vs 84.2). Teacher influence is measurable but small; most of the remaining leakage returns with any repair, including the oracle.

AuthorsOpenReview ↗

Response 2/2

Sustained CL+MU, and changes to the paper

W2: CL+MU. Since submission we have added, in our replies to Reviewers YjDt and wseG:

  • five joint baselines (three composed CL→MU pipelines, UniCLUN, CLPU-DER++) on Split CIFAR-100 under interleaved deletion (wseG 1/3);
  • one, two and three requests at different stream positions (YjDt Update 1, wseG 1/3);
  • a 20-task stream with three requests, which also exposes a limitation: there, LwU's continual-learning accuracy trails replay-based methods (wseG 2/3);
  • D′_r drawn only from data available at the request, with no image of earlier tasks (wseG 2/3);
  • the U-LiRA trajectories through later learning in 2/3.

Participant-level deletion, 10 seeds (the CIFAR-20 proxy from 2/3). The target is Retrain's participant accuracy, not 0.

MethodACC ↑BWT ↑Participant acc after task 5Distance to Retrain ↓
Retrain (never sees the participant)83.34±1.01−1.4856.39±29.14–
LwU (pre-edit teacher)83.91±0.62−1.3878.76±14.4222.4
LwU, edit-anchored repair83.12±1.38−2.3441.74±22.8714.7
LwU, edit only80.23±2.69−6.0029.45±26.2626.9
LwF + SalUn83.64±1.15−4.3546.58±19.949.8
CLPU-DER++77.71±1.63−12.1034.63±26.5221.8

Standalone (seed 42), edit-anchored repair lowers U-LiRA AUC relative to LwU in 7 of 8 cells.

Changes to the paper

  1. Table 2 corrected; every privacy number under one U-LiRA protocol; the privacy claim limited to what those numbers show.
  2. The U-LiRA trajectories (edit, repair, later learning) with all four repair variants.
  3. Section 2.4 updated with the measured size of teacher influence.
  4. Participant-level deletion as a stated limitation, with edit-anchored repair as a proposed fix for deletions without a class mask.
  5. A joint CL+MU section with the baselines, 1–3 requests, and the 20-task stream, with its Task-IL scope stated.

We hope this resolves the inconsistency and answers the mechanism question, and we would be grateful if the reviewer would reconsider the rating.

Thank the authors for the detailed response and additional experiments. The clarification of the U-LiRA protocol resolves my concern about the inconsistent privacy reporting. I also appreciate the new joint CL+MU evaluations and repair analysis. However, the corrected results still show substantial privacy leakage in instance deletion, and the new experiments confirm that the original repair procedure can reverse unlearning without class masking. While edit-anchored repair provides a promising mitigation, its advantages over existing baselines remain unclear, and the evidence for sustained CL+MU is still limited. Therefore, I will maintain my original rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response (4/4)

Unlearning ablation on CIFAR-100

We also reran the unlearning ablation on CIFAR-100, on both backbones, over 3 seeds. Gap is the mean absolute difference to Retrain over Forget, Retain and MIA. Relearn@1 is forget accuracy after one epoch of fine-tuning on D_f plus an equal number of retain images.

Class deletion, CNN

VariantForget ↓Retain ↑Test ↑MIAGap ↓Relearn@1 ↓
Original83.00±3.6166.15±0.3166.32±0.2856.96±1.6829.86±1.1795.67±1.15
Retrain0.00±0.0066.82±0.4766.15±0.4651.05±2.460.000.00±0.00
w/o Zone B reset6.33±4.5166.54±0.2065.94±0.2555.48±3.273.74±1.3296.00±1.73
w/o teacher repair20.67±26.1061.02±2.7060.62±2.4451.34±1.039.64±9.6989.33±13.28
LwU0.00±0.0066.29±0.1765.63±0.1748.60±4.061.38±0.560.00±0.00

Class deletion, ResNet-18

VariantForget ↓Retain ↑Test ↑MIAGap ↓Relearn@1 ↓
Original85.67±3.0670.16±0.4470.31±0.4263.61±1.9933.60±1.4792.00±1.00
Retrain0.00±0.0070.23±0.4869.53±0.4748.61±0.900.000.00±0.00
w/o Zone B reset65.33±4.1670.30±0.4170.25±0.3763.38±3.8726.77±0.5394.33±1.53
w/o teacher repair1.00±1.7364.05±3.8763.42±3.8550.48±6.363.80±0.8381.33±16.77
LwU0.00±0.0069.68±0.7368.99±0.7250.86±2.330.96±0.760.00±0.00

Subclass deletion (apple), CNN

VariantForgetRetain ↑MIAGap ↓
Original95.00±1.0077.82±0.5151.25±5.444.22±0.87
Retrain90.67±2.3178.04±0.5746.90±4.550.00
w/o Zone B reset94.67±0.5877.94±0.4951.27±5.314.07±1.05
w/o teacher repair32.00±55.4371.47±6.4848.98±3.3124.46±18.50
w/o Zone-C step92.67±3.0677.22±0.4147.28±1.512.76±0.55
w/o C projection91.00±4.5877.10±0.4047.56±2.852.74±1.08
w/o Zone-C repair pass87.33±9.0776.75±0.6446.18±1.284.09±0.34
w/o test-time memory92.33±2.3177.12±0.4347.39±2.402.60±0.79
+ weight alignment93.00±2.0077.08±0.3447.71±1.692.68±0.36
LwU92.33±2.3177.16±0.3847.52±2.152.55±0.75

The composed pipelines (LwF + SSD, LwF + SalUn, EWC + SCRUB), UniCLUN, Class-IL and RTI under the joint protocol will follow in a separate comment. All tables have been added to the revised appendix (Now we are 63 pages).

We hope these results address the concern that the joint evidence is limited to Split CIFAR-10, and we would be grateful if the reviewer would reconsider the rating.

Code for these ablations will also be added to an anonymous link.

Thank you for the responses and the additional experiments.

However, I do not think the responses have well addressed my main concerns.

In the response, the authors restate how gradient-based parameter importance can be utilized for unified continual learning and model unlearning, which I already understand. However, this seems to be a relatively straightforward extension of existing ideas, without any new insights.

Although the authors have added joint experiments on CIFAR-100, the experimental validation of the proposed method remains limited. The paper aims to jointly address continual learning and model unlearning, but the majority of the experiments are not conducted under the unified seting.

Based on these considerations, I will maintain my original score.

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

Related papers

Discussion (0)

Sign in to comment.