acceptodds
Under review as a conference paper at ICLR 2027

Why Subspace Protection Needs a Fixed Rank and Neuron Recycling in Task-Free Continual Learning

Abstract

Neural networks trained on long task-free streams suffer from two failure modes known as loss of plasticity and catastrophic forgetting. In practice both failures can co-occur and a method commonly used against one can worsen the other, yet few studies address them together. Here we design task-free streams for supervised and reinforcement learning in which tasks recur after long and variable gaps. On these streams we study how subspace protection against forgetting interacts with neuron recycling against loss of plasticity. We find that protection whose subspace continues to grow over the stream eventually saturates and reduces plasticity even when recycling keeps almost every neuron active. A fixed rank avoids saturation, but our analysis identifies a second cost. Protection aligned with the mean input direction of a layer can suppress one of the main pathways through which dormant neurons reactivate. Removing this direction from the protected subspace restores reactivation in our experiments as the analysis predicts. Recycling therefore remains necessary even at a fixed rank. We further show that protecting layers whose input directions are less shared across tasks lowers plasticity on both streams. These results suggest that continually learning neural networks may need a bound on the protected rank along with neuron recycling and a careful choice of the protected layers. They also indicate that studying forgetting and loss of plasticity together may be important for practical continual learning algorithms.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
task-free continual learning, loss of plasticity, catastrophic forgetting, subspace protection, dormant neurons, neuron recycling

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

This paper analyzed the subspace protection and neuron recycling in task-free continual learning. This work combines fixed-rank subspace protection with neuron recycling and designed the CN-ImageNet and CN-Procgen pipelines for long-term task reproduction at non-fixed intervals. This paper is clearly written and the experiments are thorough. Experiments show that this method performs well on the ConvNet task in CN-ImageNet.

Critical strengths
  1. This paper analyzed the subspace protection and neuron recycling in task-free continual learning. The author investigates the impact between forgetting and plasticity; this research has practical implications for long-term non-stationary learning.
  2. Compared with existing methods, this paper identifies two mechanisms about subspace saturation and persistent neuron dormancy, and explain why different countermeasures are required, it’s meaningful. What's more, the experimental results provide strong evidence.
Critical weaknesses
  1. In development stream author choose rank=2 as optimal protection rank, but in main experiments use rank=4. Although the author explains it in Appendix, but this choice still reduces the independence between hyperparameter select and final evaluation. It’s recommended the author confirm the main conclusions still apply when using rank = 2.
  2. In Table I for CN-ImageNet, FRSP+ReDo achieves an R0 of 0.053, which is lower than AdamW's 0.066 and much lower than other stronger preservation methods such as OWM and GPM; the authors note that "memory is not guaranteed." Authors are suggested to jointly analyze plasticity and memory retention.
Questions for authors
  1. Why does the headline CN-ImageNe not use k=2? If k=2 was exclusively used, would all key paired claims, factorial conclusions, and layer-selection conclusions remain intact?
  2. For CN-Procgen, is it possible to design a layer-selection rule that ignores the main schedules altogether (e.g., one that selects the first block based on dev-stream overlap or early online stats only) and evaluates it all at once on the new set of third-group schedules?
  3. CN-Procgen runs only one initialization per schedule. If we add at least 3–5 seeds per schedule, will the paired/clustered CIs for whole-trunk, first-block, and ReDo continue to support the cross-domain design principles given in the main text?
6 comments

We thank the reviewer for the questions, which made the paper more rigorous. The rank and the rule that selects the protected block now come from the development stream alone, and the main conclusions hold.

Q1 and W1.

Quantity (pairs)Submission (k=4)Revision (k=2)
P of FRSP+ReDo (Table 1)1.4101.411
R0 of FRSP+ReDo (Table 1)0.0530.050
P, FRSP+ReDo − ReDo (30, Table 11)+0.027 (30/30)+0.028 (30/30)
R0, FRSP+ReDo − ReDo (30, Table 11)+0.009 (29/30)+0.006 (26/30)
P, FRSP+ReDo − self-normalized resets (30, Table 11)+0.003 (28/30)+0.003 (29/30)
Factorial, recycling on P (30, Tab. 2, 11)+0.371 (30/30)+0.362 (30/30)
Factorial, protection on P (30, Tab. 2, 11)−0.012 (5/30)−0.003 (12/30)
Factorial, protection on R0 (30, Tab. 2, 11)+0.011 (30/30)+0.006 (29/30)
Factorial, interaction on P (30, Tab. 2, 11)+0.079 (30/30)+0.062 (30/30)
P, range over five other sets of protected layers, each − conv1 to conv3 (15, Figure 7a)−0.056 to −0.021 (0/15 each)−0.045 to −0.021 (0/15 each)

We agree that a k=4 headline setting weakened the independence of selection and evaluation, so the headline setting now uses k=2, the rank chosen by the development rule. The protection main effect on P and the gain in P of protection added to self-normalized resets (+0.004, was +0.007) move into the tie margin (Section 6.2). No other row changes its sign or verdict. Protection still raises P only with recycling, since the interaction stays positive on 10 of 10 schedules.

W2.

LearnerPR0End-of-visit accuracy (denominator)First-batch accuracy (numerator)
FRSP+ReDo (conv1 to conv3, k=2)1.4110.0500.9860.239
FRSP+ReDo, all layers, k=21.3870.0620.9740.248
FRSP+ReDo, all layers, k=321.2570.0830.9040.258
ReDo1.3830.0430.9710.233
AdamW1.1140.0660.8350.241
OWM1.0190.1450.8080.288
GPM0.3010.1060.4420.226

As you point out, the R0 of FRSP+ReDo is below those of AdamW, OWM and GPM, so the table splits it into denominator and numerator (R0 subtracts chance, 0.2, from each). Against AdamW the first-batch accuracy is level (higher on 5 of 15 pairs, interval including 0) and the end-of-visit accuracy is higher, so the lower R0 is a denominator effect of learning more, not a loss of memory. AdamW also lies inside the Pareto front of P and R0 (Figure 8). OWM does remember more (higher first-batch accuracy on all 15 pairs) at the cost of end-of-visit accuracy and P, a genuine trade-off that the amount of protection sets within FRSP+ReDo. Protecting all layers at a higher rank raises R0, with the first-batch accuracy itself rising, and lowers P. The revision adds this analysis as Appendix F.1 and states the joint result in the conclusion.

Q2.

New schedules 10 to 14, 15 pairs (95% interval, positive pairs)PNew game
FRSP+ReDo (first block only) − ReDo+0.064 (−0.043 to +0.191, 10/15)−0.046 (−0.100 to +0.025, 4/15)
FRSP+ReDo (whole trunk) − ReDo+0.049 (−0.063 to +0.156, 8/15)−0.084 (−0.135 to −0.032, 5/15)

Following your suggestion, the rule protects the IMPALA block with the highest overlap on neural networks trained on the development stream, the first block regardless of the learner. We evaluated it once on schedules 10 to 14, unused in the submission, with three initializations. As on schedules 0 to 9, the first block it selects can be called neither better nor worse than ReDo, since both intervals include 0. Whole-trunk protection, the setting selected on the development stream, is again below ReDo in New game with an interval excluding 0, so it appears to learn unseen games less well. The revision adds the rule to Section 6.3 and the table as Table 14 (Appendix F.2).

Q3.

Schedules 0 to 9 (30 pairs)PNew game
FRSP+ReDo (whole trunk) − ReDo−0.038 (−0.064 to −0.010, 9/30)−0.054 (−0.101 to −0.012, 12/30)
FRSP+ReDo (first block only) − ReDo+0.000 (−0.028 to +0.032, 17/30)+0.035 (−0.010 to +0.086, 17/30)
FRSP+ReDo: first block only − whole trunk+0.039 (+0.003 to +0.075, 19/30)+0.089 (+0.048 to +0.134, 24/30)

As requested, each schedule now has three initializations. For learning new games the principle holds. In New game the whole trunk is below both ReDo and the first block with intervals excluding 0, so protecting less shared layers appears to lower the learning of unseen games, and restricting protection to the most shared first block appears to reduce this deficit. In overall return (P) the whole trunk is below ReDo on schedules 0 to 9 but not on 10 to 14, where the interval includes 0. The revision therefore updates Table 3 with three initializations and narrows the statement of Section 6.3 on CN-Procgen to the deficit in learning new games.

Dear Reviewer hDon,

Thank you again for the time you have given to our submission. To make the changes easier to find, we have re-uploaded the revision with the added and revised text marked in green. Among the changes, your concerns led us to confirm the main conclusions at k=2, the rank selected on the development stream, and to analyze plasticity and memory jointly (Appendix F.1).

If time allows before the discussion closes on November 18 (AoE), we would be grateful to hear whether our response addresses your concerns and whether you have any further comments that could improve the paper. If anything remains unclear, we would be glad to clarify it.

Best regards,
The authors

Dear Reviewer hDon,

Following your suggestion, we analyzed plasticity and memory jointly, and we would like to point to one result. As you noted, some baselines have a higher first-batch recall than FRSP+ReDo. Figure 8 of the revision shows that each of them has a lower plasticity. With all layers protected at rank 32, FRSP+ReDo is above AdamW in both quantities.

We hope this is helpful and would be happy to address any further questions or comments. We sincerely thank you for your time and review.

Best regards,
The authors

Reviewer_Zq9bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper fouces on the conflict between forgetting and the plasticity in continual learning. The authors show that traditional subspace protection causes saturation and persistent dormancy. To address this, they propose FRSP+ReDo and evaluate it on supervised and reinforcement learning streams.

Critical strengths
  1. Effective and practical joint solution of FRSP+ReDo.

  2. The paper writing of this study is easy to understand.

Critical weaknesses
  1. Fixed rank kk significantly impacts performance (Figure 7). How can kk be determined for unknown tasks?

  2. Why does protection not benefit from recycling in MLPs (Table 8), while severely affecting ConvNets?

  3. When adding GroupNorm/LayerNorm, ReDo must reset normalization parameters to work (Table 9).

  4. The cost of streaming power iteration in large-scale networks like LLMs, ViTs, must be reported.

  5. Proposition 2 relies on the subspace containing the mean input μ\mu. Can feature centralization bypass this without modifying projections?

  6. FRSP+ReDo exhibits low first-batch recall. Does this trade away immediate memory retrieval for plasticity?

  7. Certain settings were selected on schedules 0-4 but degraded on held-out schedules 5-9. Does this suggest overfitting?

Questions for authors

Please check the weaknesses part.

3 comments

We thank the reviewer for the points, which let us test the cause of dormancy directly and widen the evidence to normalized architectures and to the cost at scale.

W1.

The table below is reconstructed from Figure 7b and Tables 8 and 10 to show P on the main stream against the rank. "Selected" marks the rank chosen on the development stream.

P on the main streamk=2k=4k=8k=16k=32ReDo
ConvNet, conv1 to conv31.411 (selected)1.4101.4061.3931.3681.383
ConvNet, all layers1.3871.3781.3591.2861.2571.383
MLP, all layers1.1411.1501.151 (selected)1.1201.0810.937

As you point out, P falls substantially as the rank grows when all layers of the ConvNet are protected. On conv1 to conv3, the layers selected on the development stream, P stays within 0.005 from rank 2 to rank 8 and falls below ReDo only at rank 32.

For unknown tasks k is selected once on a development stream built from classes that do not appear in the main stream (Appendix C). On the ConvNet and the MLP the rank selected in this way has the highest P on the main stream (by 0.001).

W2.

On the MLP protection alone leaves almost no neuron dormant (Table 8), so recycling has little to reinitialize. Both the MLP and the ConvNet protect the mean input direction, and the difference lies in the inputs of the protected hidden layers. In conv3 the mean direction holds most of the input energy, and protection fixes the weight component along it, so a neuron whose mean pre-activation has fallen stays dormant. In the second layer of the MLP it holds about a third, and the variation of the input still activates the neuron.

W3.

As you note, on the normalized ConvNet recycling must also reset the normalization parameters of a recycled unit, and the revision states this as the default rule of recycling in Section 5. The reset raises P of ReDo (1.142 to 1.195, Table 9) and not that of FRSP+ReDo, so it helps the baseline and not our method, and FRSP+ReDo stays above ReDo with the reset on all 15 pairs. On ResNet-20 with GroupNorm ReDo has the same P with and without the reset (1.220 and 1.222), and FRSP+ReDo is above it on all 15 pairs (Appendix F.3).

W4.

As requested, the revision reports the cost on large neural networks in Appendix F.5 (Table 17). With every linear layer of ViT-B/16, GPT-2 small and GPT-2 medium protected at k=2, a training step takes 0.9% to 6.4% longer than a step of AdamW alone in fp32 and 4.3% to 52% longer with the forward pass in bf16, where the step of AdamW is shorter. After the warm-up the extra state is 0.5 to 1.3 MB per model, and peak memory rises by at most 19.1%.

W5.

We tested this directly (Appendix F.4) by centering the inputs of the protected convolutional layers with a running mean and repeating the two-by-two comparison of Table 2. In the protected layers centering removes most of the dormancy that protection adds, without changing the projection. This agrees with Proposition 2, where the alignment with the mean direction is the cause. With centering, protection alone still leaves about half of the units dormant, most of them in the linear layers that are neither centered nor protected, and has a lower P than AdamW with centering. The large gain from recycling and the positive interaction remain.

W6.

On the ConvNet the Pareto front of P and first-batch recall R0 (Figure 8) consists of FRSP+ReDo, variants of FRSP with more protection, OWM and OWM + ReDo. ReDo and AdamW lie inside it, so against these two nothing is traded between P and R0. FRSP+ReDo is above ReDo in both P and R0 (30 and 26 of 30 pairs, Table 11), and with all layers protected (k=32) it is above AdamW in both (Table 13). The lower R0 of the conv1 to conv3 setting against AdamW is a denominator effect, since its first-batch accuracy on return is level and its accuracy at the end of the previous visit is higher. The genuine trade-off runs along the front, where more protection raises R0 and lowers P, from FRSP+ReDo on conv1 to conv3 at the plasticity end to OWM at the memory end. The first-batch accuracy of FRSP+ReDo is not below that of ReDo either (Table 13), so immediate retrieval is not given up. What is given up against ReDo is the speed of relearning over the first few hundred samples after a return (A_320 in Table 11).

W7.

The drop you point out is that of first-block protection on CN-Procgen with one initialization per schedule. With three initializations (Appendix E) neither sign of overfitting remains. Schedules 5 to 9 are harder for every learner, and the first block drops least. Its order against ReDo also reverses, slightly below on schedules 0 to 4 and above on 5 to 9, so the gap in the submission appears to reflect variation across initializations. A rule that ignores the main schedules selects the same block, which on new schedules can be called neither better nor worse than ReDo (Appendix F.2).

Dear Reviewer Zq9b,

Thank you again for the time you have given to our submission. To make the changes easier to find, we have re-uploaded the revision with the added and revised text marked in green. Among the changes, your points led us to test Proposition 2 directly by centering the inputs of the protected layers (Appendix F.4) and to report the cost on ViT-B/16 and GPT-2 (Appendix F.5).

If time allows before the discussion closes on November 18 (AoE), we would be grateful to hear whether our response addresses your concerns and whether you have any further comments that could improve the paper. If anything remains unclear, we would be glad to clarify it.

Best regards,
The authors

Dear Reviewer Zq9b,

Regarding your question on first-batch recall, we would like to point to one result from our response. In Figure 8 of the revision, the accuracy of FRSP+ReDo on a returning task starts at the level of AdamW and then rises above it (panel c). FRSP+ReDo is also above ReDo in both plasticity and first-batch recall.

We hope this is helpful and would be happy to address any further questions or comments. We sincerely thank you for your time and review.

Best regards,
The authors

Reviewer_f1nAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the interaction between catastrophic forgetting and loss of plasticity in task-free continual learning. It introduces recurring-task streams to evaluate both and identifies two limitations of subspace protection: accumulating protected directions can exhaust learning capacity, while mean-aligned protection can slow dormant-neuron reactivation.

Fixed-rank protection combined with neuron recycling is evaluated as a mitigation. Experiments support the identified effects, although the benefit of the combination depends on the architecture and protected layers.

Critical strengths

A relevant interaction studied through a suitable protocol. Recurring tasks with long, variable gaps make it possible to examine both continued learning and retention. The finding that recycling can keep neurons active without preventing subspace saturation is particularly useful: dormant fraction and effective rank alone can miss an important learning limitation.

Plausible mechanisms with supporting controls. The factorial comparison, mean-excluded projection, random-subspace control, and matched update shrinkage provide meaningful evidence. The persistent-dormancy mechanism is more distinctive than the already recognized saturation problem, and fixed-rank protection with recycling is a plausible response.

Critical weaknesses

The practical improvement remains modest and mixed. On CN-ImageNet, FRSP+ReDo improves plasticity from 1.383 to 1.410 over ReDo, but first-batch recall remains low at 0.053 and below AdamW’s 0.066. On CN-Procgen, first-block protection does not improve overall normalized return over ReDo, while whole-trunk protection performs worse.

The mitigation therefore avoids certain failure modes, but does not establish a clearly preferable balance between learning and retention. The paper would benefit from a sharper explanation of when adding protection is worthwhile and which outcome should guide that choice.

The evidence supports a narrower recommendation than the title suggests. Complementarity between protection and recycling is established mainly on the supervised ConvNet. Recycling adds no benefit on the MLP, and the Procgen results depend strongly on which layers are protected. Normalized residual architectures remain untested in the supervised setting.

Furthermore, layer overlap is measured after training and does not predict individual transition gains. It helps describe the results, but provides limited guidance for choosing protected layers in advance. These limitations leave the general necessity of the proposed combination unresolved.

Questions for authors

Practical clarification: Which reported result best demonstrates a useful improvement in the joint plasticity–retention trade-off over recycling alone?

Scope clarification: What conditions distinguish settings that need both components from those where protection or recycling alone is sufficient?

Selection clarification: Is there evidence available that can guide layer selection before the main training run, given that the Procgen development stream selected the worse-performing whole-trunk configuration?

4 comments

We thank the reviewer for the questions, which led to a sharper statement of when each component is needed.

Q1, W1 and W2. Gain over recycling alone and the criterion for protection

The most direct result is the paired comparison with ReDo on the CN-ImageNet ConvNet (Table 11). P is higher on all 30 pairs (+0.028) and first-batch recall (R0) on 26 (+0.006), and both are higher together on 26, so one is not traded for the other. In the plane of P and R0 ReDo also lies inside the Pareto front and FRSP+ReDo on it (Figure 8, Appendix F.1).

As you point out, the gain in P on the ConvNet is small and the results differ across settings. The gain is larger on the two normalized architectures (Tables 9 and 15), but on ResNet-20 it shrinks over the stream and R0 is below that of ReDo (−0.010, Appendix F.3). On CN-Procgen first-block protection has a higher R0 (+0.037) and a P that can be called neither better nor worse than that of ReDo (Appendix F.2). The lower R0 against AdamW comes from a larger denominator and not from less memory. The first-batch accuracy on return is level with AdamW (0.239 against 0.241), and the accuracy at the end of the previous visit is higher (0.986 against 0.835, Table 13).

Protecting all layers at a higher rank moves FRSP+ReDo along the Pareto front, with P falling and R0 rising (from 1.411 and 0.050 on conv1 to conv3 to 1.257 and 0.083 with all layers at k=32, Table 13). The outcome that should guide the choice is therefore the level of memory that is needed, and one selects the amount of protection with the highest P at that level.

Q2 and W3. When both components are needed, and the scope of the recommendation

The table below is reconstructed from Tables 8, 15, 9 (with Appendix D.4), 2 and 12 (in row order) to compare P with and without protection and recycling, and the values and their aggregation are those of the cited tables. The first four rows are on CN-ImageNet.

PNeitherProtection onlyRecycling onlyBothDormant, protection only
MLP0.7261.1520.9371.1510.002
ResNet-20 with GroupNorm1.2091.3521.2201.3760.043
ConvNet, normalized1.1601.2111.1951.2590.382
ConvNet1.0541.0201.3851.4130.569
CN-Procgen, whole trunk (Table 12)0.8140.8790.9490.8910.194

On CN-ImageNet recycling is needed where protection alone leaves neurons dormant. On the four architectures the P that recycling adds to protection (Both minus Protection only) grows with the dormant fraction in the last column. The normalized residual architecture that you mention is now tested with ResNet-20 and GroupNorm (Appendix F.3). There protection alone leaves few neurons dormant and accounts for most of the gain in P, and recycling adds 0.024.

Whether protection is worth adding to recycling depends on whether the protected layers receive input directions shared across tasks. On CN-Procgen this sharing is highest in the first block and decreases in the later ones, and protecting the whole trunk is below ReDo on new games (Table 3).

As you point out, the evidence supports a narrower recommendation. The conclusion of the revision now states that both components were needed only on the ConvNet and that recycling added little on the MLP and on ResNet-20. On every CN-ImageNet row Both still has the highest P or is level with it, so there the combination is a safe default when one cannot tell in advance which component is needed.

Q3 and W4. Choosing the protected layers before the main run

There is such evidence on CN-Procgen. The rule "protect the block with the highest overlap on weights learned on the development stream" selects the first block on all eight development neural networks. The rule was fixed before any training on five new schedules and then evaluated on them only once (Section 6.3, Appendix F.2). The first block selected by the rule could be called neither better nor worse than ReDo, whereas the whole trunk selected by the development performance was below ReDo on new games. The twelve development segments did not show this deficit (Appendix C).

The limitation you point out remains on CN-ImageNet. Before training the overlap is highest in the first layer, so the same rule selects the first layer alone, and protecting only this layer is not the best choice. We see two reasons. First, the gain of protection grows as more convolutional layers are protected. The margin in P over ReDo rises from +0.007 for conv1 to +0.028 for conv1 to conv3 on the ConvNet (Table 10), and from −0.001 for the first layer to +0.156 for the whole trunk on ResNet-20 (Appendix F.3). Second, before training the overlap is nearly the same in every layer after the first, so it does not indicate where to stop (Appendix D.3). We therefore select the layers on CN-ImageNet by the development performance, and on both architectures the selected layers have the highest P on the main stream.

Dear Reviewer f1nA,

Thank you again for the time you have given to our submission. To make the changes easier to find, we have re-uploaded the revision with the added and revised text marked in green. Among the changes, your comments led us to give a narrower recommendation in the conclusion on when each component is needed and to add ResNet-20 with GroupNorm as a normalized residual architecture (Appendix F.3).

If time allows before the discussion closes on November 18 (AoE), we would be grateful to hear whether our response addresses your concerns and whether you have any further comments that could improve the paper. If anything remains unclear, we would be glad to clarify it.

Best regards,
The authors

Thank you for the detailed response and the revision. The paired comparison with ReDo clarifies the joint gain in plasticity and first-batch recall on the ConvNet, and the new GroupNorm ResNet-20 results make the limits of that finding clearer. Layer selection remains an open question across benchmarks, but the revised conclusion reflects the scope of the evidence. I’ll update my rating to 3 (Weak Accept).

Dear Reviewer f1nA,

We sincerely thank you for your time and careful consideration of our response and the revision. We would be happy to address any further questions or comments.

Best regards,
The authors

Comments on the paper

2 comments

We thank the reviewers for their careful reading and constructive comments. The points raised were specific, and we have addressed them with new experiments and analyses in the revised PDF. New results are collected in a new Appendix F (Tables 13 to 17 and Figure 8), so the numbers of the existing tables and figures are unchanged. The main text and the existing appendices are revised where a point raised by the reviewers changes a value or a statement, and the individual responses give the reasons together with the supporting results. This comment summarizes what the reviews changed.

  • Selection independent of the evaluation (hDon, Zq9b, f1nA). The rank of the headline setting and the rule that selects the protected block on CN-Procgen now come from the development stream alone. The headline setting uses the selected rank k=2, and the rule is evaluated on five new schedules with three initializations per schedule. The main conclusions hold. At k=2 three statements of Section 6.2 change, two effects that move into the tie margin and one match that becomes a near match (recycling only the linear layers, Appendix D.2). The statement on CN-Procgen is narrowed to the deficit in learning new games.
  • Plasticity and memory together (hDon, Zq9b, f1nA). Appendix F.1 separates first-batch recall into its two accuracies and places every method on the plane of plasticity and recall. The lower recall against AdamW is not a loss of memory, no baseline matches FRSP+ReDo in both quantities, and more protection trades plasticity for recall.
  • When each component is needed (Zq9b, f1nA). On CN-ImageNet recycling is needed where protection alone leaves neurons dormant. Protection adds persistent dormancy where the mean input direction holds most of the input energy of a protected layer (Section 4, Appendix D.4), and centering the inputs of the protected layers removes most of this added dormancy (Appendix F.4). The condition holds on the ConvNet but not on the MLP, and protection alone leaves almost no neuron dormant on the MLP and few on ResNet-20 with GroupNorm. Accordingly the conclusion now states that both components were needed only on the ConvNet and that recycling added little on the MLP and on ResNet-20 with GroupNorm.
  • Wider evidence (Zq9b, f1nA). The revision adds a normalized residual neural network, ResNet-20 with GroupNorm (Appendix F.3), and the cost on ViT-B/16 and GPT-2 (Appendix F.5).

Through these changes we believe that the paper has become more rigorous in how its settings are selected, broader in its evidence and more precise about when protection and recycling are each needed. We are grateful to the reviewers for the comments that made this possible.

We once again thank the reviewers and the area chair for taking the time to review our paper. We briefly summarize the discussion below.

Our paper. We study why subspace protection costs plasticity on task-free streams in which tasks recur after long and variable gaps. We identify two mechanisms, saturation of accumulating protection and persistent dormancy under protection, and evaluate fixed-rank subspace protection (FRSP) with neuron recycling (FRSP+ReDo) as a mitigation.

Strengths noted by the reviewers.

  • "A relevant interaction studied through a suitable protocol" (f1nA)
  • "Plausible mechanisms with supporting controls" (f1nA)
  • "Clearly written and the experiments are thorough" (hDon)
  • "Effective and practical joint solution" (Zq9b)

Main points of the discussion.

  • Plasticity and memory (f1nA, hDon, Zq9b). The reviewers pointed out that the first-batch recall of FRSP+ReDo is below those of some baselines. A joint analysis shows that FRSP+ReDo lies beyond the Pareto front of the baselines in plasticity and first-batch recall (Appendix F.1), and the second revision states this below Table 1, marked in purple.
  • When FRSP and recycling are both needed (f1nA, Zq9b). They also asked how far the complementarity of FRSP and recycling extends, given the result on the MLP. Experiments on the cause of dormancy and on normalized networks show that FRSP and recycling were both needed where FRSP alone left many neurons dormant, as on the ConvNet (Section 6.2, Appendices F.3 and F.4).
  • How the settings are chosen (hDon, Zq9b, f1nA). Another question was whether the conclusions depend on how the rank and the protected layers were chosen. With both chosen on the development stream, the main conclusions hold, also on five new schedules of CN-Procgen (Section 6.3, Appendix F.2).

Open question. As Reviewer f1nA notes, layer selection across benchmarks remains open, and the revised conclusion leaves it to future work.

We are sincerely grateful for these reviews, which have improved the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.