acceptodds
Withdrawn from ICLR 2027

What Makes Learned Knowledge Last in Continual Learning?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Catastrophic forgetting limits the ability of neural networks to learn multiple tasks sequentially. Continual learning methods address this problem by reducing the interference introduced by later tasks. However, once a sample is classified correctly, existing methods treat it as securely learned, without considering how easily that prediction may later be reversed. We show that predictions that are equally correct at task completion can differ substantially in their future retention. To capture this difference, we introduce acquisition stability (AS), which measures how often a correct prediction remains correct under mild, label-preserving input changes. In our experiments, low-AS predictions are forgotten up to 7.1× more than high-AS ones. Since AS can be measured while the current-task data are still available, fragile predictions can be identified and strengthened before future interference arrives. We therefore propose AS-CL, which (i) strengthens low-AS predictions before leaving a task, and (ii) gives greater replay weight to predictions that remain fragile during later tasks. Across three benchmarks, AS-CL consistently reduces forgetting, with the largest gains on low-AS predictions.

Primary area:
learning theory
Keywords:
Continual Learning, Prediction Stability, Knowledge Retention

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_SAvrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper solves the catastrophic forgetting problem in continual learning from the perspective of stable and fragile learning. They argue that sequential learning can change all correct samples after new task parameters are introduced. They introduce acquisition stability (AS) to measure how correct predictions remain correct.

Critical strengths
  1. The idea is interesting and well motivated.
  2. The flow of the paper is understandable and logical.
Critical weaknesses
  1. The structure of the paper can be improved with greater precision. For example: a. The notation table should appear before defining the notation, not at the end. Also, for ICLR with a 9-page cap, please consider removing the notation table and using space more efficiently. b. Lines 192 to 202 are not part of the method section. I do not understand why they are appearing there. They should be in the experiment section. c. Some acronyms, like Class-IL in line 269, appear without a full definition. Or CL and Continual learning have been used interchangeably, while you have already defined Continual Learning as CL. d. Fig. 2 should be more abstracted and informative. It reads like an AI-generated summary of the paper. The main point should be conceptually informative so a first reader, like me, can grasp the paper's idea, not a flowchart with text-based rectangles. e. Most of the paper reads as future-referencing; I see a lot of referring to upcoming sections and the appendix, which breaks storytelling consistency. f. Section 3.2 says methods addressing fragility (ii) “through parameter perturbations” include Bhat et al. 2022 and Hacohen & Tuytelaars 2025. Related Work describes Bhat as input-view consistency and Hacohen as learning-speed-based memory construction. Neither perturbs parameters, and Related Work itself contrasts AS-CL with weight perturbation (STAR, C-Flat).

  2. The paper idea is interesting; however, I found some Algebraic errors, such as: a. lines 152- 153: the lowercase z notation is undefined on first appearance. b. The Sec. 3.2 coupling argument does not follow. ‖∇{W₁}m‖ = ‖g‖‖x‖ and ‖∇ₓm‖ = ‖W₁ᵀg‖ scale differently, so the orderings of examples can differ. Δ{t→s} moves all layers, not just W₁. Eq. 5 linearizes around a(x) − x, but a 10° rotation or 2-pixel shift is a large, non-gradient-aligned pixel change. The argument also assumes a linear first layer, while the model is a ResNet-18 with BatchNorm. This one breaks your logical motivation. Unfortunately, I couldn't find the code for the paper in your submission to see how you implemented this part. c. Alg. 1 and Eq. 8 normalize v(x) differently.

  3. Some inconsistencies in your paper make me wonder if you have double-checked everything before submission or not. For example: a. Stage 2 is described two ways. The abstract and intro say replay weight goes to predictions that “remain fragile during later tasks” and that AS is “recomputed for the stored examples”. Sec. 3.5, Table 1 and Fig. 2 say q_x is computed once at θ_t and never updated. Recomputation at each replay only appears as an ablation (App. E, Table 8). How do you explain this inconsistency? b. Surprisingly, the base-epoch objective differs between text and algorithm. Sec. 3.5 says base epochs “use L_CL”. Alg. 1 line 1 says L_CL with the replay term replaced by L^AS_replay. It is never stated whether the stabilization epochs also use weighted replay. This matters for the Table 4a “Stage 1 only” row. I don't think this can be well justified in the rebuttal phase. c. Prop. 8 contradicts Alg. 1. The proof says that with λ_s = 0 “no view passes are computed”. Alg. 1 computes K = 32 views for Q̃ (lines 3–5) and for q_x (lines 10–12) unconditionally, so per-step cost does not reduce to the backbone’s. d. Some minro errors are: A(t,s) (accuracy matrix) and A_t (acquired set) share a symbol. T denotes both the number of tasks and the transformation family, and also the horizon in the Fig. 6 caption. C_t is used in Sec. 3.5 but is missing from Table 1.

  4. I found some writing quality issues that prevent me from understanding the paper clearly. a. . Fragile, brittle, vulnerable, susceptible, unstable and low-AS are used interchangeably. Specifically, none of them are interchangeable. They have their own different meanings; be precise about this. b. Mixed spelling conventions. The text uses “stabilization” and “analyzed”, but Fig. 7 has “stabilisation budget”, Sec. 4.3 has “analyse”, and App. E.2 has “recognisable”. Choose US or UK spelling throughout, including figure text. c. Garbled Fig. 1 caption. d. The discussion paper is heavily dense and full of broken sentences; it needs to be written more clearly and consistently. I cannot point to each sentence one by one, but please read it carefully; it is fragmented, dense, and not easy to follow.

  5. I expect to see a reproducibility statement at the end of the paper and the code of the paper attached as supplementary material. This can not be addressed during the rebuttal phase, and I cannot confirm that the numbers appear in the paper.

  6. For a fairer comparison, please consider benchmarking also with the WSN and SNV papers, as those papers are based on mask freezing; it shows how freezing some part of the model can affect the AS introduced by you. Also, when comparing with related work, emphasize how your method differs from theirs and, in the buffer-free setting, how your method can perform.

WSN: Forget-free Continual Learning with Winning Subnetworks, Haeyong Kang, ICML 2022 SNV: Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?, Vahedifar, ICML 2026

Questions for authors

Please refer to Weaknesses section.

Reviewer_4xXMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether predictions that are correct at the end of a task are equally robust to future interference in continual learning. The authors introduce acquisition stability (AS), which measures the proportion of mildly transformed inputs that remain correctly classified after the current task has been learned. Based on this measure, they propose AS-CL, a two-stage continual learning framework. The first stage strengthens predictions with low acquisition stability while the current-task data are still available, whereas the second stage assigns larger replay weights to stored examples that remain fragile after stabilization.

Critical strengths
  • The paper investigates example-level differences in susceptibility to catastrophic forgetting, an important problem in continual learning. Distinguishing predictions that are equally correct at the end of a task but differ in their subsequent retention could provide a useful perspective on forgetting.
  • The method is evaluated on three image-classification benchmarks with comparisons to established baselines and comprehensive ablation studies.
Critical weaknesses
  • Although the proposed method is simple and effective, its core components essentially combine robustness training with a strategically designed replay mechanism to mitigate forgetting. To better establish the advantages of the proposed approach, the authors should compare it against combinations of standard data augmentation techniques (or +contrastive learning) applied during training and various replay strategies.
  • The paper relates sensitivity to input perturbations to sensitivity to parameter updates. However, these two types of perturbations are fundamentally different, and future parameter updates are determined by the characteristics of subsequent tasks rather than arbitrary local perturbations. The authors acknowledge some of these limitations, but the theoretical discussion does not fully explain when or why input stability should reliably predict future forgetting. For example, the authors could discuss when AS is uneffective to predict forgetting (e.g., considering the relationship between consecutive tasks.).
  • the and DER+id baseline is missing.
Questions for authors

see weaknesses

Reviewer_J8ehOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

the work introduces the idea of acquisition stability, which measures the confounded stability of image classification labels in a class-incremental content learning task.

Critical strengths

the idea of studying gradual forgetting and its confounding factors in backpropagated deep neural networks is an interesting problem, especially when dealing with the large number of methodologies of continual learning.

Critical weaknesses

the title and the abstract make or allude to ideas which are not presented in the paper. The paper deals with a very narrow interpretation of continual learning with respect to image classification on a very specific architecture and models for very specific datasets. However, the title and the abstract make it seem like continual learning in general is being spoken about. Please stick to scientifically accurate descriptions. The description of label-preserving input changes to abstract and uninformative. It's unclear what is mild here. The changes in brightness, translations, and rotations are all domain differences, which have been studied to different degrees in earlier papers. Without a proper understanding of what mild label-preserving changes mean, it's unclear what is being referred to here. The sentence "The first prediction is correct in a neighborhood of the input. The second is correct at the point" is unclear. The paper is incorrect in its understanding of "stability" of earlier predictions has not been studied. Many concepts related to it, including gradual forgetting, forgetting of previous tasks (which is a metric in itself), as well as catastrophic remembering, are all related to it. It kind of points towards the fact that the authors have not reviewed the area adequately. Further, the setup of this work confounds adversarial robustness or robustness to domain differences with that of forgetting. For a proper analysis, there are many unclear questions which have not been studied in this work and which should have been studied. For example, do the data statistics affect the type of classes or data which is more affected in terms of gradual forgetting? Are any kind of input data augmentations applied, and do they help with respect to this type of forgetting, which will deal with the compounding factor of input domain differences that are being tested on? How does this affect different kinds of content learning methods in image classification? Our data replay method or generative replay methods were affected more, as was catastrophic remembering. Are methods which do strict content learning affected less? task subspace joining methods, or task subspace separation methods. Which one of them is affected, and how? Further, I find the comparative methods as well as the benchmarks inadequate. The field has moved on beyond image classification, Cifar100, and tiny ImageNet. Please refer to recent work on continual learning from recently published conferences.

Questions for authors

please refer to the weakness section. The current work requires major rewriting and much more significant analysis rework.

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

Related papers

Discussion (0)

Sign in to comment.