acceptodds
Under review as a conference paper at ICLR 2027

Loss of Plasticity and Its Recovery via Replay-Driven Compression in a Spiking Model of Hippocampal CA3

Abstract

Artificial neural networks trained sequentially on new information are prone to interference: new learning can degrade or overwrite what was previously acquired. Research into this failure mode has recently led to a distinction between catastrophic forgetting, akin to overwriting, and loss of plasticity, in which both older and newer memories are degraded. Here we ask whether similar limitations arise in biologically-grounded spiking neural networks and study how nature might address them. In a biologically detailed recurrent circuit that models the mammalian hippocampus, multiple memory traces (modeling spatial environments) are learned sequentially by a common network and evaluated through their spontaneous and cue-evoked replay. We find that sequential learning progressively saturates recurrent connectivity, increasing interference among memories and eventually eliminating the network's ability to replay stored representations. This collapse resembles loss of plasticity more closely than classical catastrophic forgetting. We then show that alternating online learning with offline and spontaneous replay-driven synaptic compression counteracts this saturation: it reorganizes the recurrent weight distribution while preserving recall, allowing the network to store additional memory traces. Together, these results identify synaptic compression and plasticity schemes organized around spontaneous replay as complementary and neurobiologically-grounded strategies for mitigating interference in recurrent spiking memory networks. This identification connects memory capacity limits observed in biology to an active distinction in the machine learning literature between loss of plasticity and catastrophic forgetting.

Primary area:
applications to neuroscience & cognitive science
Keywords:
spiking neural networks, continual learning, catastrophic forgetting, loss of plasticity, hippocampus, memory replay, synaptic plasticity, homeostatic plasticity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uw7XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies memory interference in a recurrent CA3 model with 8,000 pyramidal cells and 150 basket cells. Sequentially remapped environments share one synaptic network, and memory is assessed through spontaneous and cued replay. After five environments, moderate resetting of weak synapses restores replay, whereas stronger resetting destroys it. Interleaving learning with anti-Hebbian offline plasticity and resets redistributes weights while preserving replay across five environments. At higher loads, cued recall of Environment 1 persists longer than spontaneous replay under different evaluation protocols. The authors interpret compression as counteracting synaptic saturation and supporting continued memory function.

Critical strengths
  1. The shared recurrent circuit makes the memory-interference problem concrete and easy to follow. Different environments modify the same connections, while known place-cell orderings allow generated sequences to be associated with particular memories. This links the circuit description to a functional measure of memory expression and makes the consequences of synaptic modification interpretable beyond changes in average activity.

  2. The evaluation combines reset-threshold sweeps, recall across five environments, and weight-distribution analyses. Moderate resetting restores replay, whereas stronger resetting abolishes it, revealing an informative dependence on intervention strength. The combined protocol also redistributes weights above the direct reset threshold. Together, these observations provide a useful empirical basis for studying how a recurrent network can recover access to stored sequences after substantial synaptic modification.

Critical weaknesses
  1. The evidence does not fully separate retention of existing memories, successful acquisition of new traces at higher loads, and the incremental functional benefit of anti-hSTDP. The five-environment tests assess multiple memories, but Table 1 reports higher-load cued recall only for Environment 1, leaving successful retrieval of the additional traces unverified. Its disclosed before-reset spontaneous and after-reset cued measurements also compare complete protocols, limiting conclusions about retrieval mode itself. Although connectivity reorganization with preserved recall is demonstrated, the additional recall or capacity benefit of anti-hSTDP beyond resetting alone remains unclear without a matched comparison.

  2. Two reporting inconsistencies prevent a precise assessment of uncertainty and model specification. Figure 1b labels spontaneous replay n=5, whereas Section 2.2 states two simulations per threshold; the error-bar definition and treatment of cued trials sharing a network remain unclear. Appendix Equations (1)-(3) also combine additive currents with signs that appear inconsistent with their intended physiological effects. These may be reporting errors, but require checking against the actual runs and implementation before interpreting their consequences.

Questions for authors
  1. For the additional environments in Table 1, are environment-specific recall outcomes available for the newest and intermediate traces? Please report success counts and independent-run counts under a common recall protocol, where available. If only Environment 1 was tested, distinguish demonstrated retention of that trace from the claim that compression enables successful acquisition of additional memories.

  2. Table 1 already specifies the different reset stages. Is there an existing comparison of spontaneous and cued retrieval at the same reset stage and memory load, with detector criteria made comparable? If not, please attribute the observed capacity difference to the two complete evaluation protocols rather than to retrieval mode alone.

  3. With the reset schedule, thresholds, memory load, and evaluation protocol held fixed, does enabling anti-hSTDP improve replay or recall compared with resetting alone? Please report any available matched comparison. This would clarify its incremental functional benefit beyond the demonstrated connectivity reorganization with preserved recall.

  4. Please reconcile Figure 1b's n=5 with Section 2.2's two spontaneous simulations per threshold, and define the error bars. For the ten cued trials, explain how uncertainty accounts for the five environments sharing each learned network. Please identify the independent unit used in the reported summaries.

  5. Equation (1) adds synaptic and adaptation currents, Equation (3) uses g(V−E), and Equation (2) gives a positive spike-triggered adaptation increment. Please state the signs used in the implemented voltage equation and reconcile them with the printed expressions. Clarify whether any discrepancy is confined to manuscript notation or affects the simulations.

Responding to my questions will cover the points I‘ve addressed in the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer uw7X

Thank you again. The revised PDF addresses your questions as follows.

Q1. Fig. 1a reports which environment each spontaneous replay expressed (newest, intermediate, oldest), and Table S3 reports cued recall of every learned environment in the control. In the combined protocol only the first environment was cued, and §2.4 limits the claim at those loads to retention of that trace.

Q2. We removed spontaneous replay from Table 1. Fig. 1b now provides the matched comparison you asked for: both readouts are measured in the same simulations after the same reset (§2.2).

Q3. The control without offline plasticity (§2.4, Tables 1 and S3) reached comparable cued-recall capacity, so anti-hSTDP did not improve recall. The control also differs in reset schedule and in plasticity during recall, which we state explicitly. Anti-hSTDP’s contribution is different: comparable recall with 23% fewer potentiated synapses at seven environments (Table 2), stable dynamics at eight environments, and continued replay of newly learned environments (Fig. 1a). The title, abstract and §2.4 now make this claim rather than a capacity gain.

Q4. Fig. 1b now uses 10 independent simulations per threshold. Cue trials within a simulation share its weight matrix, so they are averaged within each simulation, and error bars give the standard error of the mean across simulations (§2.2, Appendix A.10).

Q5. Corrected as described in our earlier reply (Appendix A.3, Eqs. 1–3).

Thank you for the detailed response and additional analyses. The corrected equations, clearer separation of experimental conditions, and expanded repetitions improve the reliability of the presentation. The new controls also clarify that the increase in cued-recall capacity is largely attributable to the homeostatic reset, while anti-hSTDP mainly contributes to compression and replay stability. The revised framing is therefore more consistent with the evidence.

My main concern remains the connection to loss of plasticity and its recovery. Preserving spontaneous expression of the newest memory does not directly establish an improved ability to acquire subsequent memories, and fewer potentiated synapses alone do not establish that the resulting compression preserves future learning capacity. The revised results support a useful but narrower contribution; the central link between compression and recovered plasticity remains indirect. I am therefore maintaining my rating of 2.

Thank you for the careful follow-up. We agree with your summary, with one update from the final simulations. A schedule-matched control (reset after every environment, no offline anti-hSTDP) shows that the reduction in potentiated synapses also comes from the reset schedule, not from anti-hSTDP. With the reset matched, anti-hSTDP weakens the strongest synapses but leaves recall, capacity and spontaneous replay unchanged. The revised paper attributes both capacity and synaptic cost to the homeostatic reset.

On loss of plasticity, we agree that the link is indirect, and the revised Discussion says so. What the results do show is narrower. Without the reset, the network stops replaying any environment after five have been learned. With the reset applied after every environment, environments learned sixth and seventh are still stored well enough to be recalled from a cue (the newest environment in 5 of 6 simulations at seven environments). So the reset preserves the network’s ability to store new environments in retrievable form. It does not show that the rate or efficiency of learning is restored. As we note in the Discussion, a direct test would compare how quickly a loaded network and a fresh network learn the same new environment. We see that as the next step and have framed the contribution accordingly.

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

The paper studies sequential learning of one-dimensional spatial environments in a recurrent spiking model of hippocampal CA3, and uses the temporal structure of spontaneous and cue-evoked replay as the memory readout rather than classification accuracy. The authors report that under symmetric potentiation (st-LTP), increasing the number of stored environments progressively saturates recurrent connectivity, so that replay degrades and then disappears; resetting sub-threshold weak synapses to a 0.1 nS baseline together with offline anti-Hebbian plasticity (replay-driven synaptic compression) reorganizes connectivity and extends the range over which cue-evoked recall of the first environment remains possible. They interpret this as a replay-driven account of loss and recovery of plasticity, and argue that capacity under this protocol is retrieval-dependent. Validation includes a load sweep (twelve replicates per load), a five-environment homeostatic-reset experiment, synaptic weight-distribution analyses, and the extended-capacity experiment covering five to nine environments (Table 1, Table S1).

Critical strengths
  1. Replay sequences as the memory readout. Compared with downstream task accuracy, this readout can distinguish gradual degradation from outright collapse, and it directly matches the authors' question of whether a memory can still be read out. This is the paper's most valuable methodological difference.
  2. Transparent reporting. The load sweep has twelve replicates per load, Table S1 gives the per-threshold breakdown, and the Limitations section states plainly that standard continual-learning benchmarks are not used and that the sparsity manipulation is confounded with a raised weight cap. This makes the remaining claims easier to assess.
Critical weaknesses
  1. Recovery of acquisition is not isolated from improvements in retention or retrieval. The paper shows that replay can recover, and Section 2.2 tests cue-evoked recall for each of the five learned environments. However, the experiments do not independently establish whether the ability to acquire a new environment is restored, as opposed to already-encoded traces becoming easier to retain or retrieve. This limits the evidence for the plasticity-recovery interpretation, which is the paper's central claim. A loaded-versus-fresh acquisition comparison with matched readout conditions could help distinguish these explanations. This most affects my assessment.
  2. The extended-capacity endpoint has limited coverage and replication. The Table 1 cued-recall readout covers only the first environment and is a binary endpoint; its denominators pool eight thresholds, and at the critical loads of seven to nine environments most per-threshold success probabilities come from a single run. The capacity statement therefore lacks precision.
Questions for authors
  1. Under identical learning and retrieval conditions, does a network that has already stored several environments learn a new environment less effectively than a fresh network, and does the reset plus anti-hSTDP intervention restore that ability? Can recall be reported separately for old, intermediate and newly added environments?
  2. Why is spontaneous replay measured before that environment's homeostatic reset, while cued recall is measured after compression of all learned environments? Does this arrangement affect the interpretation that capacity depends on the retrieval protocol?
  3. Can independent replicates be added at a few informative loads, with per-threshold denominators and uncertainty reported?
  4. At five environments, which place-cell rate filter was actually used for Figure 1 and Table 1, given that the methods section says "five or more" and Appendix A.4 says "more than five"? Do the replay and capacity conclusions persist under a consistently specified detection rule?
3 comments
AuthorsOpenReview ↗

Response to Reviewer HeGq

Thank you again. The revised PDF addresses your points as follows.

W1/Q1 (acquisition vs. retention). We revised the claim to match the evidence. The title now reads “Mitigation”, and throughout the paper we describe the recovery as a recovery of replay. The loss-of-plasticity interpretation rests on new evidence: without offline plasticity, newly learned environments are crowded out of replay by older ones even though they were encoded (Fig. 1a), and at seven environments control simulations still recall their newest environment about as often as the first (Table S3). The Discussion and Limitations state that we do not show restored acquisition, and describe the fresh-vs-loaded comparison you suggested as the direct test.

W2/Q3 (coverage and replication). The combined protocol now has 25 simulations (11 at 1.75 nS, 7 at 2.0 nS), with per-threshold counts and 95% Wilson intervals (Table S2). Fig. 1b uses 10 simulations per threshold.

Q2 (spontaneous before reset, cued after). We removed spontaneous replay from the extended-capacity evaluation, so Table 1 reports cued recall only. The matched comparison of the two readouts is now in Fig. 1b, where both are measured in the same simulations after the same reset: above 2.0 nS, spontaneous replay is absent while a cue still evokes a full sequence (§2.2).

Q4 (place-cell filter). The 70 Hz filter is used at all loads, and §4.3 and Appendix A.4 state the rule consistently. Validated counts are nearly identical under a 10 Hz filter at every load (Table S1).

Thank you for the substantial revision. Several of my earlier concerns are addressed, but two points still affect my assessment:

1.Because the two protocols differ in reset timing and recall plasticity, Table 2 compares full protocols rather than isolating anti-hSTDP alone. Do you have a schedule-matched control for the 23% synapse reduction? 2.High-load recall across all environments (Table S3) is reported only for the control, while the combined protocol still tests Env.1 only. Are corresponding multi-environment results available for the combined protocol?

For now, I will keep my current score unchanged.

AuthorsOpenReview ↗

Response to follow-up: schedule-matched control and multi-environment recall

Thank you for these questions. Both led to new simulations, and the first changed one of our main claims. We have revised the paper accordingly.

  1. Schedule-matched control for Table 2. You were right that Table 2 compared protocols that differed in when the reset was applied. We ran a schedule-matched control at 1.75 nS (6 simulations). The reset followed every environment, as in the combined protocol, but there were no offline anti-hSTDP phases. Median potentiated synapses (≥ 1.75 nS, thousands):

Environments Combined Schedule-matched Single reset 6 1,072.9 1,087.7 1,031.0 7 1,232.7 1,246.8 1,591.8 8 1,387.3 1,400.7 2,044.7

The reduction relative to the single-reset control comes from the reset schedule. When the reset follows every environment, small increments are removed before they can accumulate. Anti-hSTDP adds only about 1% (Mann–Whitney, p < 0.001 at each load). We have removed the “quarter fewer potentiated synapses” claim from the abstract, introduction, §2.4 and Discussion, and Table 2 now includes the schedule-matched column.

Within-simulation comparisons show what anti-hSTDP does to the weights (new appendix, “Weight changes during the offline anti-hSTDP phase”). In every offline phase, about half of the synapses in the 4.0–4.5 nS bin moved below 4.0 nS (41–77% across loads), while the number of synapses above 2.0 nS changed by less than 1%. Anti-hSTDP therefore weakens the strongest synapses without reducing the number of potentiated ones.

  1. Multi-environment recall under the combined protocol. We ran the combined protocol with every environment cued and a 10 s spontaneous simulation after each load (7 simulations at six environments, 6 at seven). Recall was similar to the schedule-matched control: 5.0 of 6 environments per simulation against 4.2 at six environments (p = 0.12), and 4.2 of 7 against 4.0 at seven (p = 0.72). At seven environments, automated detection missed many cue-evoked replays, so we scored recall at this load from raster plots in both protocols.

Spontaneous replay was also similar. At six environments, it was confined to the first two environments in both protocols, and the newest environment was never replayed spontaneously (0/7 and 0/6), although it could still be recalled from a cue. The advantage of anti-hSTDP for recent memories in Fig. 1a therefore holds only without a reset. The results for each environment are in the appendix table “Schedule-matched control.”

Together, these results show that the homeostatic reset determines capacity and synaptic cost in our model, while anti-Hebbian replay weakens the strongest synapses but leaves recall and capacity unchanged. We have revised the abstract and Discussion to state this.

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

This paper investigates how sequential learning affects memory storage in a biologically detailed spiking neural network of the hippocampus. The authors find that learning more spatial environments gradually saturates synaptic connections, eventually preventing the network from replaying stored memories. To address this, they combine offline replay with anti-Hebbian plasticity and synaptic compression, which weakens or resets certain connections while preserving important memory traces.

Critical strengths
  1. The paper addresses an interesting question about how neural networks lose their ability to support additional memories during sequential learning, connecting biological memory mechanisms with continual learning in machine learning.

  2. Evaluating memory through spontaneous and cue-evoked replay is a useful approach, as it allows the authors to study whether learned activity sequences can still be retrieved.

  3. The analysis of synaptic weight distributions provides useful evidence that repeated learning changes recurrent connectivity and that these changes are associated with the loss of replay.

Critical weaknesses
  1. The paper does not provide a controlled comparison of anti-Hebbian replay alone, synaptic compression alone, and their combination under the same extended-learning setup, making it difficult to determine how much each mechanism contributes to the increased memory capacity.

  2. The experiments change the replay detection threshold and use stronger cues when evaluating higher memory loads, which makes direct comparisons between different memory capacities less straightforward.

  3. Some experiments have very few independent simulations, particularly the compression-threshold analysis and extended-capacity experiments, making it difficult to assess the statistical reliability of the results.

  4. The extended-capacity experiments mainly evaluate cue-evoked recall of the first learned environment, rather than systematically testing whether all previously learned environments remain accessible.

Questions for authors
  1. Could the authors compare anti-Hebbian replay alone, synaptic compression alone, and their combination using the same learning and evaluation settings? This would help clarify which mechanism is responsible for the improvement in memory capacity.

  2. How can the authors distinguish between memories that have actually been lost and memories that remain stored but can no longer be retrieved through spontaneous replay?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Nsej

Thank you again. The revised PDF addresses your points as follows.

W1/Q1. We added a control with no offline plasticity at any stage (§2.4, Tables 1 and S3). It differs from the combined protocol in three respects (no offline phases, a single reset after the last environment, no plasticity during recall), so it is not a single-factor ablation, and we state this in §2.4 and Limitations. Cued-recall capacity was comparable under both protocols, so the capacity extension is attributable mainly to the homeostatic reset. Anti-hSTDP instead reduced the number of potentiated synapses needed (23% fewer at seven environments, Table 2), kept the network stable at eight environments, and kept newly learned environments in spontaneous replay (Fig. 1a).

W2. One detection rule (70 Hz filter) is now used at every load, and a single cue protocol for all cued experiments (§4.3). Validated replay counts are nearly identical under a 10 Hz filter (Table S1).

W3. Fig. 1b now uses 10 simulations per threshold, and the combined protocol 25 simulations, with per-threshold counts and 95% Wilson intervals (Table S2).

W4. Fig. 1a reports which environment each spontaneous replay expressed, and Table S3 reports cued recall of every learned environment in the control. In the combined protocol, cued recall was tested for the first environment, so our claims at those loads are limited to that trace.

Q2. The Discussion now defines a memory trace as retained if it can be recovered spontaneously or after cueing. Spontaneous replay is the stricter probe: cues recover environments the network no longer replays on its own (Fig. 1b, Table S3). Testing whether “lost” traces speed relearning is noted as future work.

Presentation. Appendix A.10 (Table S6) summarizes the protocol, readout and sample size for every experiment.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of responses and revision plan (revised PDF uploaded)

We thank the reviewers for careful and constructive reviews. This comment summarizes our responses and the changes in the revised PDF,

The main changes are: Fig. 1 and §2.1–2.2 (new analyses; 10 simulations per threshold in Fig. 1b); §2.4 and Tables 1–2 (re-analysis with 25 simulations, a control without offline plasticity, and a comparison of weight distributions); §4.3 and Appendix A.4 (one detection rule at all loads; Table S1 compares filters); Appendix A.3 (corrected Eqs. 1–3); new Appendices A.8 (duration of the offline phase) and A.10 (summary of all protocols, Table S6); and revised title, abstract, Discussion and Limitations.

One correction to note: re-analyzing Table 1 with the corrected detection pipeline and 25 simulations removed the few successful recalls previously reported at eight environments. Under both protocols, cued recall now ends at eight environments (Table 1, Table S2).

A detailed change report comparing the submitted and revised PDFs is included in the supplementary material ([Compare Report] ICLR_2027_Submission vs Rebuttal.pdf). The same comparison is also available through OpenReview’s “Show Revisions” → “Compare Revisions”.

AuthorsOpenReview ↗

Final revision: summary of changes since our earlier comment

We thank the reviewers again. Two new controls, prompted by Reviewer HeGq, changed one of our conclusions, and we have revised the paper accordingly.

Schedule-matched control. We applied the reset after every environment without offline anti-hSTDP. The reduction in potentiated synapses we reported (“about a quarter fewer”) comes from the reset schedule, not from anti-hSTDP, which adds only about 1%. We have removed this claim from the abstract, introduction, §2.4 and Discussion. Table 2 now includes the schedule-matched column. Combined protocol with every environment cued. Recall (5.0 vs 4.2 of 6 environments; 4.2 vs 4.0 of 7) and spontaneous replay were similar with and without anti-hSTDP. The advantage of anti-hSTDP for recent memories in Fig. 1a holds only without a reset. What anti-hSTDP does. In every offline phase, about half of the synapses at the weight ceiling moved below it, while the number of potentiated synapses was unchanged (new appendix).

The revised paper therefore attributes capacity and synaptic cost to the homeostatic reset. It describes anti-hSTDP as weakening the strongest synapses without changing recall or capacity, and states that the link to recovered plasticity remains indirect. The schedule-matched results, per-environment recall and the weight analysis are in the appendix. We also clarified the cue-onset exclusion rule (Appendix, replay detection) and updated the AI-use statement.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.