acceptodds
Under review as a conference paper at ICLR 2027

Layer-Selective Utility Gating for Continual Learning

Abstract

Neural networks in continual learning suffer from plasticity loss. Utility-based perturbed gradient descent (UPGD-W) mitigates this problem through utility-dependent modulation of parameter updates. We investigate how this modulation should be allocated across network layers. We first analyze an idealized continuous-time model and show that, under regularity and stationarity assumptions, its expected displacement-based utility rate relates to gated local curvature with an explicit correction for non-equilibrium probability currents. We then derive a distributional tracking bound that separates the contributions of representation and head drift. These results provide a theoretical motivation for comparing output-only, hidden-only, and full gating across different adaptation settings. On three supervised label-permutation benchmarks, output-only gating applies utility modulation to only 0.16%–2.5% of parameters while achieving 114%–285% higher mean accuracy than Shrink and Perturb and up to 13% higher than full gating. All layers continue to update. The gating variants perform similarly on Input-Permuted MNIST. In Gridworld with controlled dynamics shifts, hidden-only gating achieves higher mean return than output-only gating. Together, these findings show that the benefits of utility gating depend on its placement and the adaptation demands of the learning setting.

Primary area:
optimization
Keywords:
Continual learning, loss of plasticity, utility-gated optimization, layer-selective gating, distributional tracking, continual reinforcement 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_eL9nOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

The paper explores the question of where in a neural network should a utility-dependent gradient modulation be applied to preserve plasticity. It compares three kinds of gating: output-only, hidden-only and full-network. The motivation behind this question is the observation that most plasticity-preserving methods are generally applied across the whole network, but different distribution shifts might require different parts to adapt. They use an already established method UPGD-W and provide both theoretical analysis and empirical results in a variety of benchmarks. Making gating location an experimental design variable has a potential positive impact in the literature.


Comment to authors

Overall, I found this paper to be strong and thoughtful. The research question is clear and the ablations and main empirical results are sufficient to support the main claim. My main concerns are about robustness and scope rather than the main idea. I think that the discussion should better incorporate recent work, and better place itself around those. If these points are addressed satisfactorily, I would be comfortable increasing my score.

Critical strengths

Overall, I believe the paper has potential for impact in the community, and that the strengths outweigh the weaknesses.


  • Clear research question. And clear experimental ablation construction with the isolation of gating placement from other types of parameter freezing.
  • Strong empirical results on label-perm. The main empirical finding is clear rather than marginal.
  • Extensive and qualitatively different shift benchmarks. The results are more nuance than just a single simple claim. There is a difference between label permutation results (favoring output-only) and Gridworld (favoring hidden-only), and that points to the idea of properly placing gating based on the adaptation dynamics.
  • Proper and deeper mechanistic analysis. I appreciate the authors going deeper into the diagnostics and checking whether the performance differences are consistent with what the actual mechanism proposes rather than just an artifact of some other obscure interaction.
  • Honesty about limitations. I appreciate the intellectual honesty about the theory and claims made in the paper such as the checkpoint displacement being only a heuristic proxy or that theory does not provide a universal gating-placement rule.
Critical weaknesses

Major

  1. The strongest result for supervised learning are biased for output-layer. The 3 label-permutation tasks provide an input that is fixed and only change the mapping classes->output labels, thus it might logically follow that output-only gating performs much better. Including an end-to-end representation shift benchmark like domain shifts or purely class-incremental learning with new classes would strengthen the claims of the paper.

  2. The statistical results might be biased. The main comparisons should be re-run with matched independent seeds. If this was the case, please indicate so.

  3. Gridworld results should be verified under a fixed nº of env interactions or same nº of optimizer steps. If this was the case, please indicate so and/or consider it a necessary ablation to make robust claims about such results.


Minor

  1. Incomplete coverage of recent plasticity related literature. There are related-work omissions that are directly relevant to the mechanisms studied while interpreting layer-wise plasticity. First of all, Lyle et al. [1] argues that plasticity loss arises from many mechanisms interacting at the same time rather than a single property, and that needs to be take into account for the papers interpretation. Also, Lillo & Cheney [2] showed that the activation function of choice strongly affects plasticity in supervised CL and RL, and that is relevant to the authors work as they fix ReLU while drawing conclusions about hidden-layer adaptability. Park et al. [3], and Rohani et al. [4] also studied activation dynamics and increased linearity as mechanisms for maintaining plasticity, and that ties into the study of dormant units and representation plasticity. Finally, Rosseau et al. [5] introduces AdamO, which is an Adam-style optimizer that decouples isometry-promoting regularization from the adaptive gradient update, thus this should also be discussed as a recent optimizer-based approach to maintaining plasticity.

  2. The gated parameter fraction (0.16%-2.5%) might be misleading. While output-gating acts only on those small %s, the utility still tracks over all parameters. The main text should make this distinction clear.

  3. A more careful statement about GridWorld dynamics-shift. Adam achieves basically the same return and that should be explicitly stated and discussed.

  4. The replay-buffer ablation in App. I.1 weakens generalization of GridWorld results. I'd consider discussing this in the main text and exploring its consequences as the results become closer in absolute value.


References

[1] Lyle et al., “Disentangling the Causes of Plasticity Loss in Neural Networks” (CoLLAs 2025)
[2] Lillo & Cheney, “Activation Function Design Sustains Plasticity in Continual Learning” (ICLR 2026)
[3] Park et al., “Activation by Interval-wise Dropout: A Simple Way to Prevent Neural Networks from Plasticity Loss” (ICML 2025)
[4] Rohani et al., “Preserving Plasticity in Continual Learning with Adaptive Linearity Injection” (CoLLAs 2026)
[5] Rosseau et al., “Preserving Plasticity in Continual Learning via Dynamical Isometry” (ICML 2026)

Questions for authors
  • Based on Theorem 3.2, could the authors empirically measure displacement-based utility and examine its relationship with gated curvature?

  • Do the authors recognize if the layer-placement effect might persist in other shifts? Like in domain shifts, class-incremental shifts, or end-to-end CNNs (see Critical Weaknesses Major 1).

  • Could GridWorld be matched by env transitions or optimizer updates rather than episodes? (See Critical Weaknesses Major 3).

  • Could the authors clarify if there is a multiplier for non-gated parameters in the RL experiments? If so, what made that specific value to be chosen? There should be exploring those dynamics.

4 comments
AuthorsOpenReview ↗

Gridworld and matched comparisons (Part 1 of 2)

Gridworld budgets and adaptation (Major W3, Minor W6, Q3, Q4)

We thank the reviewer for these suggestions. Rebuttal Table 2 in our response to Reviewer qNrp shows the new Gridworld results. The submitted runs used equal episode counts. The new experiments instead fix both interactions and optimizer updates, giving every method 500,000 interactions and 499,936 updates per run across all three shifts, with ten paired seeds per condition. For the question about nongated parameters, nongated layers multiply the normalized gradient plus injected noise by 0.5, with weight decay applied separately. We use a fixed multiplier of 0.5 because the sigmoid-based gating function returns 0.5 at zero normalized utility.

We agree that the original dynamics-shift results did not establish an advantage of Hidden over Adam. In the new experiment, mean early success was 90.31% for Hidden, 90.05% for Constant-0.5, and 87.78% for Adam. Together with the other shift conditions, these results support our finding that the usefulness of gating placement depends on the learning setting.

Replay and the scope of the finding (Minor W7)

We thank the reviewer for raising this concern. We repeated all three shifts with replay, using five methods and ten paired seeds per condition. Batch size remained one, with 500,000 interactions and 499,936 updates per run. The 10,000-transition buffer was retained across shifts.

MethodGoal, finalDynamics, earlyDynamics, finalJoint, final
Output-only gating72.05 / 10.6681.88 / 74.4696.63 / 96.2976.96 / 12.43
Full gating62.55 / 15.9382.07 / 75.8096.88 / 96.7392.17 / 18.10
Hidden-only gating62.46 / 24.3890.31 / 80.6699.98 / 99.9885.23 / 53.21
Constant-0.569.38 / 15.5490.05 / 80.6699.98 / 99.9973.21 / 28.07
Adam36.38 / 13.7887.78 / 79.0399.83 / 99.8642.28 / 21.64

Rebuttal Table 3. Success (%) without / with replay, averaged over ten seeds and nine post-shift phases. Success means reaching the goal within 200 steps. Early/final cover the first/last 10K interactions of each 50K phase. Bold marks the highest averages separately without and with replay; comparisons are descriptive.

Under dynamics shifts, Hidden and Constant-0.5 both reached 80.66% mean early success, compared with 74.46% for Output. Replay reduced final success for every method under goal and joint shifts and changed the placement ordering. We will discuss this dependence on replay in the main text. Our interpretation is that replaying rewards from the old goal interferes with adaptation to the new goal. This early interference may hinder exploration and learning even after the old experiences leave the buffer.

Independent seeds and matched comparisons (Major W2)

We checked all 28 entries in Table 1 and recalculated the statistics across distinct seeds. Counting the same seed three times caused the zero SE for Full and Output on CIFAR-10. The corrected accuracies are 59.96 ± 0.43% and 61.71 ± 0.51%, respectively, across ten seeds. Input-MNIST S&P now uses five seeds, giving 78.29 ± 0.15%. The corrected Table 1 lists all seed counts and optimizer settings. To address your request for matched independent seeds, we ran a new Mini-ImageNet comparison with Full, Output, and Constant-0.5 across five independent seeds. Within each seed, initialization, example order, label permutations, noise streams, learning rate, and decay were matched across methods. Output exceeded Full by 7.51 percentage points, with a 95% confidence interval of 6.73–8.29. The new Gridworld comparisons use ten paired seeds per condition. Table 1 was corrected from existing runs; the new matched experiments cover Mini-ImageNet and Gridworld.

AuthorsOpenReview ↗

Task design, efficiency and related work (Part 2 of 2)

Task design and generalization (Major W1, Q2)

We follow the streaming input- and label-permutation benchmarks of Elsayed and Mahmood (ICLR 2024), Sections 4.2–4.4 and Figures 3 and 6. Dohare et al. (Nature 2024), Methods, “Robust loss of plasticity in permuted MNIST,” and Extended Data Figure 3, study input-permuted MNIST one example at a time. Lyle et al. (2025), Section 3.1 and Figure 1, study plasticity loss under repeated label randomization. We keep the architecture and training conditions fixed across gating variants to compare placement directly.

For label permutations, both Full and Output gate the output layer. Their comparison tests whether excluding hidden layers from adaptive gating improves adaptation when useful features can be reused. Hidden layers continue to learn even when they are not adaptively gated. Neither label nor input permutation mathematically determines the best gating placement. Output has higher mean accuracy than Full on the three label-permutation benchmarks, while the observed means are close on Input-Permuted MNIST. These results support studying gating placement across different learning settings.

Efficiency potential and current cost (Minor W5)

The gated-parameter fraction does not by itself measure computational savings. We benchmarked the same Mini-ImageNet MLP and optimizer settings on one CPU thread in FP32, using batch size one and five repetitions of 1,000 training updates after warm-up. Output took 3.58 ms per update compared with 4.39 ms for Full, about 18.5% less time. Both retained the same 2.7 MB utility state because utility tracking still covers every parameter. Removing unused utility tracking from Constant-0.5 reduced update time from 3.56 to 3.07 ms without changing its training trajectory. These measurements support a runtime benefit for output-only gating in this implementation, while reducing utility storage remains future work. We will revise the abstract to distinguish the placement finding from computational savings.

Displacement utility and gated curvature (Q1)

Appendix E.2, Figure 4, and Table 6 report Spearman correlations of utility magnitude with diagonal Hessian magnitude and squared gradients across supervised learning, Gridworld, and MuJoCo. Across the reported settings and methods, mean correlations range from approximately 0.22 to 0.98 for curvature and 0.76 to 0.99 for squared gradients. These measurements show that the utility used to set our gates is positively associated with local curvature, with stronger correlations to squared gradients. They support the utility–curvature connection for the implemented optimizer, although they do not directly test the displacement-based relationship in Theorem 3.2.

Related work and the interpretation of plasticity (Minor W4)

We thank the reviewer for this suggestion. We will discuss how the papers you suggested relate to gating placement in the related-work section of the main text. We keep the architecture and update rule fixed to examine where utility gating is useful. Future work could test whether activation design changes which layers benefit from utility gating.

We will share all experiment and analysis code along with the full hyperparameter settings.

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.

However, I do not see the new revised version as a the pdf submission. Could the authors update the pdf submission of the paper with the new revised version before the rebuttal process ends?

If that is correctly done I will be keeping my Weak Accept (3) rating and I will argue for acceptance of this paper.

Thank you very much for reviewing our responses and for your support. We appreciate the reminder to update the submission PDF. We are incorporating the corrections, new experiments, and clarifications into the manuscript and will upload the revised version before the discussion deadline. We will let you know here as soon as it is available.

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

The paper studies where to apply the utility gating of UPGD-W, i.e., at the output layer only, at the hidden layers only, or at all layers, and motivates this with an SDE analysis and a tracking bound with separate representation and head terms. The authors report that output-only gating is best on label-permuted Mini-ImageNet, EMNIST, and CIFAR-10, with 114%-285% higher accuracy than Shrink and Perturb, and that hidden-only gating is better than output-only gating under dynamics shifts in a Gridworld DQN.

Critical strengths

The question where to apply utility gating is practical, and on label permutation the effect is large, e.g., 0.693 vs. 0.615 for output-only vs. full gating on Mini-ImageNet (Table 1).

The paper is open about its limits and also reports results in the appendix that go against its own story, e.g., the replay ablation (Table 19).

Critical weaknesses

My main problem is that the supervised results cannot be checked. The cached runs do not retain seed identities (lines 1421-1423), Table 1 gives an SE of 0.000 for full and output-only gating on CIFAR-10 although the runs in Figure 1c spread from roughly 56% to 65%, and apart from one learning rate for Input-MNIST S&P (caption of Figure 1), I could not find any optimizer settings for the supervised methods. Also, there is no explicit control without utility at the same step size. Almost all utility scores, and so the gates, are close to 0.5 (Figure 5), and the closest control, clamping the scores to [0.48, 0.52], still gives 30.29% on Mini-ImageNet vs. 18.0% for S&P (Table 9 and Table 1), so part of the gain over S&P may come from the smaller step size and not from the gating.

Second, the evidence that gating the hidden layers helps under a shift seems weak. Under the Gridworld dynamics shift, plain Adam reaches 0.556 ± 0.004, about the same as hidden-only gating with 0.555 ± 0.004 (Table 2), and with a replay buffer hidden-only gating is the lowest of all four methods there (Table 19). The theory does not decide the placement either, as Appendix B.4 says itself (lines 1131-1132), although Takeaway 2 calls it a "theoretical justification for layer-selective gating".

Questions for authors
  1. How many independent seeds and which hyperparameters (learning rate, noise, weight decay, nongated scale) were used for each method in Table 1, and why is the CIFAR-10 SE 0.000?
  2. How does output-only gating compare with a fixed scale of 0.5 on all layers without utility, in the supervised tasks and in Gridworld?
6 comments
AuthorsOpenReview ↗

Supervised results and the fixed-scale control (Part 1 of 2)

Table 1 corrections and reproducibility (W1, Q1)

We thank the reviewer for pointing out this inconsistency. In Table 1, we mistakenly counted the same seed’s result three times for both Full and Output on CIFAR-10. This gave an incorrect SE of zero. We checked the seed identities and recalculated all 28 cells. The corrected CIFAR-10 accuracies are 59.96 ± 0.43% for Full, 61.71 ± 0.51% for Output, and 46.35 ± 0.59% for Hidden, reported as mean ± SE with ten distinct seeds per method. Input-MNIST S&P now uses the five-seed configuration reported in the appendix, giving 78.29 ± 0.15%.

Table 1 gives the corrected results, seeds, and optimizer settings. We report mean online accuracy over the final 40 tasks and SE across independent seeds. Output-only gating retains the highest mean accuracy on all three label-permutation benchmarks. We will share all experiment and analysis code along with the full hyperparameter settings.

Full UPGD applies adaptive gains to every layer. Output-only gating applies them to the output layer, while hidden-only gating applies them to the hidden layers. Ungated layers use S&P updates with a fixed factor of 0.5 on the gradient and noise, leaving shrinkage unchanged. All gating variants use the same global utility-normalization rule across all layers. We keep this rule fixed to compare gating placements under the same optimizer design. The S&P learning rate is 0.005, compared with 0.01 for the gated methods. The table also lists the noise and decay settings for each method.

Matched fixed-scale control (W1, Q2)

The sigmoid gives a score of 0.5 at zero normalized utility. Scores near 0.5 therefore motivate testing an exact fixed-gain control. S&P provides a non-utility baseline, but its settings differ from the gated methods. We therefore ran a new matched comparison on Mini-ImageNet using the requested Constant-0.5 control, matching initial weights, data and noise streams, learning rate, and decay. Constant-0.5 uses S&P updates with a fixed gain of 0.5 on the gradient and noise in every layer, with shrinkage unchanged. Across five paired seeds, final online accuracy was 69.40% for Output, 61.89% for Full, and 4.54% for Constant. Output exceeded Constant by 64.9 percentage points [59.2, 70.6], a 95% confidence interval. Uniform 0.5 scaling alone therefore does not explain the advantage in this configuration.

AuthorsOpenReview ↗

Gridworld adaptation, replay and the theory (Part 2 of 2)

Gridworld budgets and adaptation (W2, Q2)

The submitted Gridworld experiments used a fixed number of episodes. Since episode lengths depend on the policy, this did not fix the number of interactions or optimizer updates. We therefore repeated all three conditions, giving every method 500,000 environment transitions across ten phases of 50,000 transitions. After a 64-transition warm-up, each method performed 499,936 optimizer updates. We tested all three Gridworld conditions using Adam, Constant-0.5, Output, Hidden, and Full over ten paired seeds per condition. The four UPGD methods share initialization, optimizer settings, and random streams. Nongated layers use a fixed multiplier of 0.5 on the normalized gradient and injected noise; Constant-0.5 applies this multiplier to every layer, with weight decay unchanged. All layers remain trainable, and different policies can produce different trajectories.

MethodReward (goal) shift, finalDynamics shift, earlyDynamics shift, finalJoint shift, final
Output-only gating72.0581.8896.6376.96
Full gating62.5582.0796.8892.17
Hidden-only gating62.4690.3199.9885.23
Constant-0.569.3890.0599.9873.21
Adam36.3887.7899.8342.28

Rebuttal Table 2. Gridworld success rates (%), higher is better. Success requires reaching the goal within 200 steps. Early/final use the first/last 10K interactions per phase, averaged over nine post-shift phases. Bold marks the highest averages; comparisons are descriptive.

In the new Gridworld comparison, the largest observed gain over Constant-0.5 was under joint shifts. Full reached 92.17% average final success, compared with 73.21% for Constant-0.5, a difference of 18.96 percentage points. Hidden reached 85.23%. Under goal shifts, Output had the highest observed average final success at 72.05%, compared with 69.38% for Constant-0.5.

The original dynamics-shift results showed similar returns for Hidden and Adam. In the new experiment, Hidden had lower early regret than Output, supporting our motivation for studying gating beyond the output layer. Hidden and Constant-0.5 both reached 99.98% final success, leaving little room for improvement on this measure.

Replay and the scope of the finding (W2)

We repeated all three shifts with replay at batch size one and the same interaction and update budgets as the main experiments. With replay, Hidden and Constant-0.5 both reached 80.66% early success under dynamics shifts, compared with 74.46% for Output. Replay reduced final success under goal and joint shifts for every method and changed the placement ordering. Rebuttal Table 3 in our response to Reviewer eL9n gives the comparison with and without replay.

Theory and clarifications (W2, Q1, Q2)

The corrected Table 1 provides the seed counts and optimizer settings and resolves the zero SE. The matched Mini-ImageNet and Gridworld (Rebuttal Table 2 above) experiments compare adaptive gating with a fixed multiplier of 0.5. Our tracking bound separates representation and output-layer contributions. Under the allocation assumptions in Appendix B.4, larger gains are assigned to layer groups whose target parameter distributions change more. This motivates adapting layers differently, while the experiments test which gating placements help. We will clarify this distinction consistently in Takeaway 2 and Appendix B.4.

Thank you for the answer and the additional experiments, which address most of my first concern. The supervised results can now be checked, and the matched Mini-ImageNet control (69.40% for output-only gating vs. 4.54% for Constant-0.5) suggests that the gain does not come from the step size, which was my concern in the review. The planned clarification of Takeaway 2 is also fine for me. So I raise my soundness score from 1 to 2.

In the new Gridworld experiment under the dynamics shift, hidden-only gating reaches about the same success rate as Constant-0.5 (90.31% vs. 90.05% early and 99.98% for both at the end, Rebuttal Table 2), while output-only gating is lower (81.88% and 96.63%). So hidden-only gating seems to be better than output-only gating there, but not better than the same update with a fixed factor of 0.5 instead of the utility gate. The clearest gain from gating hidden layers is now under the joint shift (92.17% for full gating vs. 73.21% for Constant-0.5 and 76.96% for output-only gating), which I think is interesting, but Rebuttal Table 2 gives no SE. So I think the conclusion (lines 482-483) should rest on the joint shift with SEs and say that under the dynamics shift hidden-only gating does not improve on Constant-0.5.

Reviewer_7tvhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies where utility-based modulation of parameter updates should be applied, rather than introducing a new utility estimator. It compares output-only, hidden-only, and full-network gating within UPGD-W. The results show that output-only gating works best on the tested label-permutation tasks, while hidden-layer gating can help under Gridworld dynamics shifts. Continuous-time analyses provide conceptual motivation through connections to curvature and representation-versus-head drift.

Critical strengths
  • The paper is well-motivated.
  • Simple output-only gating produces meaningful improvements over full gating, which confirmed the claim that applying utility modulation everywhere is not necessarily beneficial.
  • The experiments demonstrate that the preferred placement is context-dependent, which may be helpful for future work.
Critical weaknesses
  • The connection between the theory and the implementation is somehow unclear.
  • The main findings are made from limited architecture and task scope. The experiments use two-hidden-layer MLPs, batch size one, and abrupt permutations. Mini-ImageNet uses fixed pretrained features, instead of end-to-end feature learning.
  • Practical benefits are limited. Gating only 0.16%–2.5% of parameters does not provide significant computational savings as the implementation still tracks utility for every parameter, requiring O(N) computation and storage.
Questions for authors

Please refer to the weakness part.

2 comments
AuthorsOpenReview ↗

Experimental design, new evidence and clarifications

Research question and theory (W1, Q1)

Our main finding is that gating placement matters for continued learning as tasks change. Selective gating can outperform full gating, and the best observed placement depends on the learning setting.

Both the theory and implementation apply parameter-wise gains to the gradient and injected noise. Our analysis in Theorem 3.2 relates displacement-based utility to curvature, motivating different update scales across parameters. The tracking bound in Theorem 3.5 and Corollary 3.6 separates representation and output-layer contributions. Under the allocation assumptions in Appendix B.4, layer groups whose target parameter distributions change more receive larger gains. This motivates comparing how different layers adapt. For label permutations, the output mapping changes while existing features may remain useful. We therefore test whether excluding hidden layers from adaptive gating improves adaptation compared with full gating.

Additional experimental evidence

We have also strengthened the empirical comparisons. We repeated all three Gridworld shift conditions with equal interaction and update budgets, including Constant-0.5. The Gridworld results (Rebuttal Table 2 in our response to Reviewer qNrp) show that the observed placement ordering depends on the shift condition. We also ran a matched Mini-ImageNet comparison with Constant-0.5 and checked all 28 entries in the submitted paper's Table 1. The corrected table reports the results, seeds, and optimizer settings. We will share all experiment and analysis code along with the full hyperparameter settings.

Experimental design and generalization (W2, Q1)

Our supervised comparisons use 300/150-unit ReLU MLPs. We keep the architecture and training conditions fixed across gating variants to examine placement directly. The streaming input- and label-permutation tasks follow Elsayed and Mahmood, ICLR 2024, Sections 4.2–4.4, Figures 3 and 6, and Appendix I.2. Dohare et al., Nature 2024, Methods, “Robust loss of plasticity in permuted MNIST,” and Extended Data Figure 3, study input-permuted MNIST one example at a time. Lyle et al. (2025), Section 3.1 and Figure 1, study plasticity loss under repeated label randomization. Batch size one supports immediate prediction and updating on incoming examples. Abrupt permutations give controlled, repeated changes whose timing is not supplied to the learner. For Mini-ImageNet, we also follow Elsayed and Mahmood's design in Section 4.4 of training an MLP on fixed ImageNet-pretrained ResNet-50 features. We reuse these established task designs to compare gating placements under the same learning conditions. The Mini-ImageNet comparison tests placement in the downstream MLP; extending the findings to end-to-end CNN training is a separate generalization question.

Efficiency potential and current cost (W3, Q1)

Our main finding concerns where gating helps continued learning. The fraction of gated parameters does not by itself measure computational savings. To measure computational cost, we benchmarked the same Mini-ImageNet MLP and optimizer settings on one CPU thread, using batch size one and five repetitions of 1,000 training updates after warm-up. In this Mini-ImageNet CPU benchmark, Output took 3.58 ms per update compared with 4.39 ms for Full, about 18.5% less time. Both retained the same 2.7 MB utility state because utility tracking still covers every parameter.

MethodTraining time (ms/update)Peak process memory (MiB)Utility state (MB)
Full gating4.39 ± 0.071247 ± 252.70
Output-only gating3.58 ± 0.081219 ± 162.70
Constant-0.53.56 ± 0.041216 ± 112.70
Constant-0.5 without utility tracking3.07 ± 0.041210 ± 240

Rebuttal Table 1. Mini-ImageNet CPU cost (FP32). Time and peak memory are mean ± SD across five repetitions. Peak process memory includes loaded features; utility state is counted separately.

Removing unused utility tracking from Constant-0.5 reduced update time from 3.56 to 3.07 ms without changing its training trajectory. Utility tracking therefore has a measurable cost. These measurements support a runtime benefit for output-only gating in this implementation, while reducing utility storage remains future work. We will revise the abstract to distinguish the placement finding from computational savings.

Thank you for your comments. We have uploaded the revised manuscript, clarifying how the theoretical results motivate the gating comparisons and why we use the established streaming benchmarks. We also added runtime and memory measurements and revised the abstract to distinguish the placement findings from computational savings. The revision includes the corrected supervised results and additional control experiments. We appreciate your time in considering these changes.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised manuscript and summary of updates

We thank the reviewers for their helpful comments and patience while we incorporated the feedback from the discussion into the revised manuscript. We have now uploaded the revision, including the corrected supervised results, the control experiments presented in our responses, and updated Gridworld and replay results with SEs. We also clarified the interpretation of the experimental results, the theoretical findings, and computational cost. We appreciate your time in considering these updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.