acceptodds
Under review as a conference paper at ICLR 2027

Do Neural Networks Lose Plasticity in a Gradually Changing World?

Abstract

Continual learning has become a trending topic in machine learning. Recent studies have discovered an interesting phenomenon called loss of plasticity, referring to neural networks gradually losing the ability to learn new tasks. However, existing plasticity research largely relies on benchmarks with abrupt task transitions, without examining whether the abruptness itself contributes to the observed plasticity loss. In this paper, we investigate the role of transition abruptness by simulating gradually changing environments through input/output interpolation and task sampling. We perform theoretical and empirical analysis, showing that the severity of plasticity loss is closely tied to the abruptness of task transitions, and can be substantially reduced when the environment changes gradually.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Loss of Plasticity, Continual Learning, Supervised Learning, Deep Learning

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_GNF2OpenReview ↗
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

  1. 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.

  2. 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.

  3. 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].

  4. 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

  1. 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)

  2. 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.

  3. 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
  1. What information is assumed available to the learner? Task sampling appears to require simultaneous access to DtD_t and Dt+1D_{t+1}, 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?

  2. 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.

  3. 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?

  4. 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

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).

    • Uniform distribution as regularizer. We used a variant of output interpolation whose target never becomes uniform (the labeled class always keeps the largest share). It preserves plasticity nearly as well as the original output interpolation that goes through a uniform target distribution (probe AUC 0.93 vs. 0.94, where abrupt change falls from 0.91 to 0.32 by task 30). The uniform phase is therefore not what preserves plasticity.
    • Replay. We compare task sampling with sparse replay, which switches abruptly to the new labels and inserts previous-task samples at a fixed rate. Sparse replay preserves less trainability than task sampling at every proportion of previous-task samples, with a probe AUC of 0.77, 0.82 and 0.87 at 10%, 25% and 47.5%, against 0.94 for task sampling. The 47.5% setting matches the proportion in task sampling, so replay alone does not explain its benefit.
    • Curriculum. We agree that task sampling can be viewed as a curriculum over the transition, ordered by similarity to the previous task instead of by difficulty. To our knowledge, task ordering has been studied for catastrophic forgetting [1] but not as a factor in plasticity loss.
  • 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).

  • 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 DtD_t, Dt+1D_{t+1} 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.

Reviewer_d3DQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates how the abruptness of task transitions affects plasticity loss in neural networks. It proposes input/output interpolation and task sampling to simulate gradually changing environments. Experiments on image and synthetic language tasks show that gradual transitions substantially reduce plasticity loss, often outperforming existing mitigation methods. The paper also provides theoretical analysis suggesting that gradual changes allow optimization to track local minima.

Critical strengths
  • Interesting perspective on plasticity loss: The paper questions the abrupt task transitions commonly assumed in existing plasticity benchmarks. The observation that transition abruptness substantially affects plasticity is interesting and suggests that benchmark design may influence the severity of observed plasticity loss.
  • Reasonably thorough experiments: The paper evaluates both trainability and generalizability across image and synthetic language tasks, with comparisons against several existing plasticity mitigation methods. The analysis of transition granularity is also useful.
Critical weaknesses
  • Unclear significance as a plasticity mitigation method: The proposed approach is more of a transition-level remedy than a fundamental solution to plasticity loss. Rather than directly improving the model's ability to learn new tasks, it makes task transitions smoother through interpolation or sampling. While this is an interesting finding, I am not convinced that it demonstrates preservation of the model's underlying plasticity, as opposed to making the optimization problem easier. I think this distinction should be discussed more explicitly in the Introduction and Conclusion, with a clearer statement of what the results do and do not establish.
  • Limited applicability to online continual learning: The proposed methods require access to both the previous and upcoming task data to construct intermediate transitions. However, in many online continual learning settings, future task distributions are unknown and environmental changes cannot be controlled. This limits the practical applicability of the approach, particularly for genuinely abrupt distribution shifts. The paper should better acknowledge this limitation rather than presenting the approach as a general plasticity mitigation method.
  • Limited connection between theory and experiments: The theoretical analysis relies on near-identity linear transformations and locally strongly convex basins. It is unclear how these assumptions relate to the interpolation and task-sampling procedures used in the experiments.
Questions for authors

How could the proposed approach be applied when future task data is not available, as in online continual learning?

Overall, I find the empirical observations interesting, and I would lean toward weak acceptance if the authors adequately address the concerns above, particularly by clarifying the scope and significance of their contribution in the Introduction and Conclusion.

2 comments

We are pleased that the reviewer appreciates the presentation of our paper and found our perspective interesting. We address each concern below and have revised the Introduction and Conclusion accordingly. All changes are highlighted in blue in the revised PDF.

  • W1 (significance as a mitigation method). We agree that our method is not a fundamental or universal solution, and we have clarified our scope in the Introduction (page 2, lines 54-56 and 63-70) and the Conclusion (from line 485). However, our additional experiments show that the underlying plasticity is preserved. We added an abrupt probe task that checks the ability to learn at intermediate checkpoints (Appendix D). Networks trained with gradual transitions keep learning the probe while abruptly trained ones degrade, with probe AUC at about 0.94 against a fall from 0.91 to 0.32 by task 30. The same trend holds on natural images, where tasks are not made easier by construction. The network trained on similar consecutive tasks learns a new task better, with test accuracy 0.336 against 0.256 (Appendix E).

  • W2 and Question (online continual learning). Thank you for your suggestion. The paper now states that our protocols apply only to supervised settings where the training schedule can be controlled and data from both tasks are available during a transition. They do not apply to online settings where old data cannot be stored or the learner must switch immediately, and they do not transfer directly to reinforcement learning (Appendix F). We have revised the paper accordingly in Section 2 (page 2, lines 86-89), the Conclusion and the Limitations (Appendix G).

  • W3 (connection between theory and experiments). Our theory is meant to give a formal picture of the intuition, not an exact model of our training procedures. To prove that gradient descent can follow a minimizer as the loss changes, some assumptions are needed. We model each small transition step as a near-identity linear change of the loss, and assume smoothness and local strong convexity around each minimizer. Interpolation and task sampling in deep networks do not satisfy these assumptions exactly. The experiments support the intuition the theory formalizes. Smaller transition steps preserve more plasticity (Figure 5, page 8), and gradual transitions keep the largest Hessian eigenvalue lower for most of the run (Figure 6, page 9). We now acknowledge this gap between the theoretical model and the experimental procedures in the Limitations (Appendix G, lines 1476-1482).

Summary. We have shown with probe tests that gradual transitions preserve the ability to learn, and not only make each transition easier. We have also limited our scope to supervised settings where the training schedule can be controlled, and present the work as a report of this phenomenon. In addition, we state the gap between the theory and the experiments as a limitation. Please let us know if you have any further questions. If you also feel this work is valuable, we look forward to your support for our paper.

Dear Reviewer d3DQ,

Thank you again for reviewing our submission. We understand this is a busy period, and since the discussion phase ends on November 18, we wanted to follow up on our responses. We have addressed your comments and incorporated your suggestions into the revised paper, with changes highlighted in blue. The other two reviewers have indicated that their concerns were addressed and that they are raising their scores to Weak Accept. We would appreciate hearing whether our revisions address your concerns as well, and we remain available for any further questions.

Reviewer_uV2MOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors investigate tasks that gradually change, contrary to tasks with abrupt changes usually evaluated in the literature, in continual learning. The main claim is that plasticity can be preserved when environments are assumed to be gradually changing. The analysis by the authors involves four environments where they show that interpolation and sampling variations on task changes preserve plasticity.

Critical strengths
  1. The authors study gradually changing environments, a topic very important given its relation to human learning.
  2. The results show that plasticity seems to be preserved in gradually changing environments.
  3. Two new language generation tasks are proposed to study continual learning. The relation to language evolution over time is interesting.
Critical weaknesses
  1. The results for plasticity preservation seem very dependent on transition smoothness. A deeper discussion of which tasks this smoothness holds is needed.
  2. No experiments evaluate plasticity loss in continual reinforcement learning, where gradually changing environments seem more natural than the artificially created tasks.
  3. The paper is redundant in places, and the writing and presentation could be improved.
Questions for authors

Comments:

  1. Both language generation environments seem to be proposed by the authors. Readers would benefit from figures to better understand these tasks. Especially for the Bigram Cipher task, the explanation is very unclear, and visuals would help to explain it.
  2. Additional discussion on which cases it can be assumed that environments and tasks gradually change is needed. For which cases only sampling and revisiting previous versions of the task would be enough? For which cases are these gradual transitions related to complete forgetting?
  3. The theorems presented seem to assume convexity. Is this a plausible assumption for these continual learning settings?
  4. If applicable to these environments, adding soft shrink and perturb would be helpful.
  5. Why do the authors choose to use only a pre-trained Transformer architecture for the language generation tasks? From a plasticity investigation perspective, it seems crucial to also train a Transformer architecture from scratch.
  6. Figures can be improved by clarifying in the labels which interpolation and sampling techniques are proposed by the authors (making it clearer from the baselines). Figure 6 can also be improved.
  7. The paper would be substantially improved if a continual reinforcement learning experiment were conducted where the state and/or the reward function gradually change over time. Additionally, a clear discussion on how this research relates to the literature on environments with dynamics that change over time seems to be needed.

Minor Comments:

  1. (lines 69-70) Typo: “learn new from new data”
  2. (line 126) Typo: “environments”
  3. (line 136) Typo: “Image” -> “Pixel”
  4. (line 152) and (lines 241-256) Redundant Writing
  5. (line 320) Typo: “regularizated”
  6. (line 374) and line (374) "Fig ??"
  7. Related Work section is highly redundant with other parts of the manuscript.
3 comments

We appreciate that you recognize the importance of the problem we study and thank you for the helpful comments. We address your comments below. All changes are highlighted in blue in the revised PDF.

  • W1 and Q2 (when smoothness holds). We agree that the result depends on transition smoothness, and the paper now discusses where it can be assumed or constructed. Smoothness holds naturally when consecutive tasks are similar. On natural images, a stream ordered by similarity keeps fitting every task, with a training accuracy of at least 99.7%, whereas the same tasks in a dissimilar order lose trainability after task 55 (Appendix E). Smoothness can be constructed in supervised settings where the training schedule can be controlled and data from both tasks are available, which is the setting of our main experiments. Revisiting previous data alone is not enough, since sparse replay with the same proportion of previous-task samples reaches a probe AUC of 0.87 against 0.94 for task sampling (Appendix D, page 22).

  • W2 and Q7 (continual reinforcement learning). We added preliminary RL experiments with PPO and SAC (Appendix F), including a case where the observation changes gradually, as you suggest. The return drops after abrupt changes, but in both agents the drop is explained by factors specific to reinforcement learning. In PPO it comes from collapsed exploration and is removed by resetting only the policy's standard deviation. In SAC it comes from the critic's overestimation and is removed by twin critics without any reset. A gradual change helps SAC only when the replay buffer is kept consistent with the current observation. We therefore limit our claims to supervised continual learning. For Q7, we added a paragraph relating our work to the literature on environments that change over time (Section 5, page 9, lines 471-481).

  • W3, Minor 4 and 7 (redundancy). Thank you. We removed the repeated intuition in Section 3 (page 4, line 165) and now define trainability and generalizability only once in Section 2. We also moved the model details of Section 4.1 to Appendix B and removed the overlap between the Introduction and Related Work.

  • Q1 (definition of Bigram Cipher). We rewrote the definition of it (page 3, lines 111-120). We also added illustrations of both language tasks (Figures 7 and 8, page 19).

  • Q3 (convexity assumption). We assume strong convexity only locally, in a neighbourhood of each minimizer, and the loss may be non-convex elsewhere. This holds when the Hessian at the minimizer is positive definite and is an idealization for overparameterized networks, which we state as a limitation (Appendix G, page 28, lines 1476-1482). The theory formalizes our intuition, and the experiments are the evidence for deep networks.

  • Q4 (soft Shrink & Perturb). We added it as a baseline on Random MNIST (Figure 1, page 6). It is one of the strongest baseline there, with a final training accuracy of about 0.99, well above the original Shrink & Perturb at 0.64. This is consistent with our observation that small, gradual changes preserve plasticity better than abrupt ones. Given the time constraints of the rebuttal, we will add the remaining benchmarks in the final version.

  • Q5 (pre-trained Transformer). We believe the initialization does not affect our conclusions. After the first task, the model is always warm-started from the previous task's solution, so plasticity loss builds up over the task sequence regardless of the starting weights. We also keep only the first block of T5-small, and our synthetic tasks are far from the pre-training data, so little pre-trained knowledge carries over. Our main claim compares abrupt and gradual transitions under the same initialization, and the pre-trained model already shows severe plasticity loss (e.g., test BLEU2 falls to 20% after 50 tasks on Bigram Cipher).

  • Q6 (figures can be improved). The legends now mark which methods are ours. We recomputed Figure 6 with the largest Hessian eigenvalue, obtained by the Lanczos method instead of power iteration on the full data of each task, and redrew it with our methods marked in the legend (page 9).

  • Minor 1, 2, 3, 5 and 6. All fixed. Thank you for the careful reading.

Summary. We have explained in which cases transition smoothness holds. We have also added preliminary results in reinforcement learning and clarified the scope of our claims. Finally, we fixed the inconsistencies and redundancies in the paper. Please let us know if you have any further questions. We look forward to your response.

Reviewer_uV2MOpenReview ↗

Answer to Authors

Thank you to the authors for your rebuttal.

The rebuttal addresses many of my concerns. A few additional points:

  1. W2, Q7 (continual reinforcement learning): checking these results, I am reflecting if a title like: "Do Neural Networks Lose Plasticity in Gradually Changing Tasks?" would be better for this manuscript. The use of the word "World" might mislead many readers.
  2. Q4: Can you clarify what the good results for the soft S&P mean? From your rebuttal explanation, is it also related to gradual changes? That part is not clear to me.

I will increase my score from Weak Reject to Weak Accept.

Thank you for taking the time to read and respond to our rebuttal. We really appreciate your willingness to raise your score, and we would be grateful if you could update it in the system when convenient.

1. We agree that “World” may suggest a broader scope than the one we study. We have changed the title to “Do Neural Networks Lose Plasticity in Gradually Changing Tasks?” and revised similar wording throughout the paper to refer to gradually changing tasks.

2. We apologize for the unclear wording in our earlier reply. The original S&P shrinks and perturbs the weights once at each task boundary, which is one large and abrupt change to the network. Soft S&P applies a much smaller shrink and perturbation after every gradient step, so the intervention is spread gradually over training. On the random-label MNIST benchmark, Soft S&P preserves plasticity much better, with final accuracy at about 0.99 against 0.64 for the original S&P. Prior work reports a similar pattern. Lee et al. [1] observe periodic drops in accuracy when a network is reinitialized at intervals, and Juliani and Ash [2] find that continual regularizers mitigate plasticity loss better than intermittent interventions. We see this as analogous to our finding, in that a change spread over many small steps preserves plasticity better than one abrupt change. The difference is that Soft S&P spreads a change to the weights, while our work spreads the change of the task.

References

[1] Lee, Hojoon, Hyeonseo Cho, Hyunseung Kim, Donghu Kim, Dukgi Min, Jaegul Choo, and Clare Lyle. "Slow and steady wins the race: Maintaining plasticity with Hare and Tortoise networks." Proceedings of the 41st International Conference on Machine Learning, PMLR 235:26416-26438, 2024.

[2] Juliani, Arthur, and Jordan T. Ash. "A study of plasticity loss in on-policy deep reinforcement learning." Advances in Neural Information Processing Systems 37, 2024.

Comments on the paper

1 comment
Area_Chair_doUYOpenReview ↗

Discussion deadline approaching

Dear Reviewers and Authors,

Thank you for the detailed reviews and rebuttal.

To Reviewers uV2M and d3DQ: As the author–reviewer discussion deadline approaches, could you please review the authors’ responses and indicate whether your concerns have been addressed?

Kind regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.