- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Summary
The paper investigates, through input/output interpolation and task sampling, the role in loss of plasticity for continual learning that abrupt task transitions have.
The motivation is that many natural distribution changes appear progressively rather than abruptly or instantaneously.
Thus, transition abruptness is an experimental variable rather than a property.
The main research question asks if a neural network still loses plasticity when the underlying problem changes gradually.
This question may have a positive impact on the plasticity literature if the results survive to stronger controls, but currently the main empirical result is smaller task-transition steps lead to better preservation of plasticity.
Comment to authors
Overall, I see potential in this work and I consider it close to the acceptance boundary. The main question is well motivated and Fig. 5 is compelling but under-explored. My main concerns revolve around whether the 'gradulness' of the abrupt changes has been properly isolated from other effects, and in the reliance on synthetic benchmarks. These concerns seem addressable rather than fundamental to the main research question. Therefore, I would be happy to raise my score if the rebuttal provides convincing evidence or clarification on some of the major points I raised.
Critical strengths
Overall, the core empirical observation, specially Figure 5, is interesting and potentially relevant to the plasticity literature and to how plasticity benchmarks should be designed and evaluated. Specifically:
- Good research question. Transition abruptness is under-examined as an experimental variable, and is central to the continual learning discussion.
- Main experiment is clean. Controlling for transition granularity aims to separate the problem from the cause. Figure 5 is strong evidence. However, it might need to be isolated better as the causal factor (See Critical Weaknesses).
- Trainability+Generalizability. Correctly identifies them as two different aspects of plasticity, and studies them independently. Seems to me particularly important to the question of abruptness.
- Multiple architecture/domains. MLPs, ResNet-18 and Transformer-based model. It checks all boxes for this paper, but the main claims will require scaling it (See Critical Weaknesses).
- Mechanistic diagnostic. This paper is more about a deeper phenomenon and its mechanistic consequences than purely a SOTA-style method. However, I consider that the study of such mechanisms (dormancy, representation change, curvature analysis) should play a larger role in this paper (See Critical Weaknesses).
Critical weaknesses
Major
-
The largest issue is whether "gradualness" has actually been isolated as the cause of improvement. Random-label experiment undergoes label smoothing. The measured improvement could come from the smooth transition but also could come from that intermediate uniform-label phase acting as regularizer. On the other hand, task-sampling mixes both old and new data during the transition and that could act as a sort of replay or curriculum learning.
-
With only five seeds and min/max bands, it is hard to make robust claims about monotonic relationships between step size and plasticity. Please, consider adding confidence intervals, aggregate statistical comparisons or increment the number of seeds to make this easier to assess.
-
The empirical evidence is synthetic despite the paper's motivation being around real-world gradual changes. I understand that most benchmarks in the CL community, and specially around loss of plasticity, tend to be like that (random labels, pixel permutations, etc), however, the language experiments are also synthetic. A natural drifting dataset or task-free stream would substantially strengthen the overall validity. One example could be: CLEAR [1].
-
The mechanistic analysis is under-studied. It is unclear if gradual transitions preserve plasticity because they prevent dormancy or curvature problems or if all those quantities chance due to making the optimization easier. Digging into these diagnostics to explain the mechanisms more strongly would strengthen the experiments.
Minor
-
Small typos. Unresolved references such as “Figs. ??” in the generalization section (lines 374 and 399). The AI-use statement repeats the same sentence twice: "We did not use them to formulate the research ideas, develop the theory, or generate data." (line 489)
-
The generalizability evidence is uneven. in permuted EMNIST, the unregularized network only loses 10% of test accuracy after 100 tasks (Fig.3; line 325), and that looks relatively little plasticity loss to fix. On the other hand, while Bigram Cipher seems more compelling, but it is synthetic, making the generalizability claim weaker than the trainability one.
-
Terminology inconsistency in the loss-landscape analysis. Section 4.3 indicates that \textit{maximum eigenvalue of the weight matrices} measures loss-landscape sharpness, while Appendix C computes the maximum Hessian eigenvalue. These are different quantities: weight-matrix eigenvalues characterize the network's linear transformations, while Hessian eigenvalues characterize curvature of the loss with respect to the parameters. In CL, and loss of plasticity literature, the latter is the quantity associated with local sharpness. Could the authors clarify this?
References
[1] Lin, Zhiqiu, et al. "The clear benchmark: Continual learning on real-world imagery." Thirty-fifth conference on neural information processing systems datasets and benchmarks track (round 2). 2021.
Questions for authors
-
What information is assumed available to the learner? Task sampling appears to require simultaneous access to and , but it is mentioned in line 80 that the experiments operate under a task-unaware setting. How would the method work under a real task-unaware setting?
-
Do the authors consider that the abruptness effect could persist on continuously drifting data rather than interpolations constructed from discrete tasks? How would the authors go about testing that? I understand time constraints, and the complexity of adding extra benchmarks, however, a real or procedurally generated continuous-drift benchmark would substantially strengthen the main claim.
-
In line 217, the authors state that: "The key observation is that after each micro-task, GD converges to the new basin’s minimizer, restoring the preconditions needed for the next micro-task", and then Theorem 3.7 relies on that observation, however, in the experiments each of the N=50 micro-tasks only get 1/N of the training budget. Have the authors verified that the optimization reaches the neighborhood claimed by Lemma 3.6 before each next transition?
-
Could the authors clarify whether Definition 3.2 is intended to be the standard local strong convexity? It seems to be a mismatch between Definition 3.2 and the properties used in Lemma 3.3.
4 comments
-
Q1 (information available to the learner). Our setting is not the task-free online setting. As is common in plasticity research, the learner makes multiple passes over each task’s data. Our gradual transitions (interpolation and task sampling) construct the data stream from , and the known task boundary. The learner only receives batched data and is given no task identity or boundary signal. We have clarified this in Section 2 (page 2, lines 86-89) and Section 3 (page 3, line 134). In an online setting where the stream cannot be controlled or stored, our protocols cannot be applied as a mitigation method. Our result only aims to show that a stream which already changes gradually should cause less plasticity loss than an abrupt one.
-
Q3 (convergence within each micro-task). Yes, we checked this. For each benchmark we set the per-task budget and learning rate so that every micro-task converges, i.e. the training loss plateaus before the next transition. We cannot measure the exact distance to the minimizer required by Lemma 3.6 in a deep network, and we mention this in the Limitations (Appendix G, lines 1480-1482).
-
Q4 (Definition 3.2). Thank you for catching this. We intend standard local strong convexity, but the definition was stated only relative to the minimizer, which is weaker than what the proof of Lemma 3.3 uses. We have restated it as the standard inequality for all pairs of points in the neighbourhood of the minimizer (page 4, line 182), and corrected the corresponding step in the proof of Lemma 3.3 (Appendix A.1, from line 861).
References
[1] Li, Ziyan, and Naoki Hiratani. "Optimal task order for continual learning of multiple tasks." Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:34578-34603, 2025.
Summary. We added controlled experiments that separate gradualness from the regularizing effect of uniform labels and from replay. We also added experiments on the CLEAR benchmark, as suggested by the reviewer. In addition, the figures now report the mean and standard deviation over runs. Please let us know if you have any further questions. We look forward to your support of our paper.
Thank you to the authors for your rebuttal. Most of my concerns have been addressed and I believe the paper provides interesting and relevant results to the continual learning community. I am increasing my score from Weak Reject (2) to Weak Accept (3).
Thank you for your careful review and constructive feedback, which helped us improve the manuscript. We are glad that our responses addressed your concerns, and we appreciate your decision to raise the score.
We sincerely thank the reviewer for acknowledging the potential in this work and for the detailed and constructive review. All changes are highlighted in blue in the revised PDF.
Major 1 (isolating gradualness). Thank you. We have added experiments for the alternative explanations you raised using an abrupt probe task (Appendix D).
Major 2 (seeds and statistics). We have updated the figures in the main paper to show the mean and standard deviation across seeds, in place of min/max bands, and revised the text accordingly (page 5, line 257). The new CLEAR-10 results use 10 seeds and report standard errors.
Major 3 and Q2 (natural data, continuous drift). Thank you for the suggestion. We added CLEAR-10 experiments on natural images as suggested (Appendix E). We order tasks by similarity in image appearance. With similar consecutive tasks, new-task probe test accuracy is 0.336 vs. 0.256 for the dissimilar order, at the cost of more forgetting.
Major 4 (mechanism related to plasticity). We have reported three diagnostics across transition granularities in Appendix C, including the maximum Hessian eigenvalue, relative representation change, and the fraction of dormant units. All three improve as transitions become finer. We have also added dormant-unit measurements on natural images (Appendix E, page 23), where models trained on similar consecutive tasks keep a lower proportion of dormant units (about 11%, compared with 53% for dissimilar tasks). We agree these diagnostics are correlational. They show that sharpness and dormancy change together with plasticity as transitions become finer, but they cannot tell us whether lower sharpness and dormancy are the reason plasticity is preserved or just a side effect of it.
Minor 1 The typos, unresolved references and the repeated sentence are fixed.
Minor 2 We agree that the generalizability evidence on permuted EMNIST is mild and have explained in the caption of Figure 3. We have added experiments with natural dataset in Appendix E.
Minor 3. Sorry for the mix-up. The terminology was our error, and we have changed it to the Hessian. We measure the sharpness of the loss landscape, i.e. the largest eigenvalue of the Hessian. We also recomputed Figure 6 with the Lanczos method on the full data of each task, since our earlier power-iteration estimate returned the eigenvalue of largest magnitude and could be negative (page 8, lines 425-431, and Figure 6 on page 9).