acceptodds
Under review as a conference paper at ICLR 2027

Intrinsic timescale placement shapes temporal computation and learning in recurrent networks

Abstract

Neural systems comprise units with varying intrinsic timescales, but how this heterogeneity contributes to temporal computation and interacts with recurrent dynamics is not fully resolved. We show that intrinsic heterogeneity provides a richer temporal basis for a readout; however, in our temporal tasks, recurrence can largely reduce this advantage as diverse effective dynamics emerge. By fixing the recurrent graph, its weights, and the set of intrinsic timescales while varying only their assignment across nodes, we find that placement systematically changes the effective temporal repertoire. While degree and strength effects are topology dependent, alignment with slow recurrent modes shows more consistent effects across the tested topology families, indicating slow recurrent structure as a more general organizing coordinate. These dynamical effects also influence function: under fixed recurrent connectivity, slow mode alignment improves temporal computation in the tested tasks and topology families, with generally larger benefits at greater temporal demand, and serves as an inductive bias when recurrence is trainable. Finally, we test whether this principle is reflected in human cortical organization. Empirical magnetoencephalography (MEG) neural timescales are positively associated with slow modes of the structural connectome, and their observed spatial ordering produces a broader effective temporal repertoire than spatially rotated alternatives. Taken together, these results suggest that placement relative to slow recurrent structure is an organizing factor in how intrinsic timescale heterogeneity contributes to network dynamics, computation, and learning.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Recurrent Neural Networks, Neural Timescales, Network Dynamics, Temporal Computation

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_692HOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper investigates how heterogeneous time constants across units in a recurrent neural network impact computational performance. Even if time constants, τi\tau_i's, are set to be uniform the recurrent connectivity, WrecW_{\text{rec}} generates a range of time constants. The authors show that one can amplify the diversity of timescales by setting the time constants to be non-uniform and "aligning" them with the timescales implied by WrecW_{\text{rec}}. This might have computational benefits. A preliminary investigation of MEG data is undertaken to search for biological evidence of this "alignment".

Critical strengths
  • I like that the paper points out that the "benefits" of having heterogeneous time constants can be achieved by training WrecW_{\text{rec}} directly. My sense is that previous papers have potentially oversold the benefits of baking in this heterogeneity. This paper does a decent

  • The authors show that early learning is faster when W_rec is trainable and the time constants are aligned to amplify fast and slow modes. They also show in a reservoir

Critical weaknesses
  • I found this paper extremely difficult to read and honestly I did not understand the motivation behind the project or its significance. The series of experiments are introduced in section 2 without much explanation and I found the composition of the paper disjointed—When reading section 3, I had to refer back to section 2 to understand what was going on.

  • The tasks studied in this paper are very simple, toy problems. I am not against studying toy problems, but they need to be better motivated. For an ML audience, there is no strong connection to genuinely challenging modern tasks. For a neuroscience audience, the mapping to biology isn't made clear. Furthermore, these tasks have been extensively studied by Yang et al. (2019) and Driscoll et al. (2024) and its not clear what new insights the paper here provides.

  • The analysis of the human MEG dataset is potentially a new link to biology, but it is hard to interpret the results. The authors themselves say it provides "supportive but limited evidence". If I understand it correctly, the authors only have access to the effective timescales measured by MEG. The effective timescales reflect a mixture of recurrent connectivity and intrinsic time constants for each region, and it is hard to disentangle the two with this dataset. Since the rest of the paper deals with the intrinsic timescales, there is not direct evidence for or against any of the key points in the paper.

Questions for authors
  • Who is the target audience for this paper?
  • What are the take home messages for that audience?
  • What future directions of research are enabled by your contribution?
2 comments

We appreciate the reviewer's recognition that recurrent training can compensate for the benefits of intrinsic τ\tau heterogeneity, and that aligned placement improves early learning. We also thank the reviewer for highlighting the need to improve the clarity of the manuscript's motivation, significance, and organization. These comments helped us clarify the central question and make the importance of our work more accessible to a broader audience. We have accordingly revised the Introduction and the presentation of our experiments.

 

Weakness 1 — motivation and presentation

We appreciate the reviewer's point on the narrative and clarity. The central question of our work is not simply whether heterogeneous intrinsic timescales are beneficial for recurrent computation; recurrent connectivity itself can generate a diverse set of effective timescales, even when the intrinsic time constants of the neurons are identical. Thus, the question we address is a more specific one in computational neuroscience, namely whether changing the assignment of the same set of heterogeneous intrinsic timescales across neurons in a network changes the network's dynamics and computation.

In our experiments, we first investigate the contribution of heterogeneity to computation with and without recurrence. We then isolate the effect of intrinsic timescale placement on effective dynamics and task performance, and investigate whether the initial placement affects learning when recurrent weights are trained. Finally, we investigate whether cortical activity timescales are associated with slow modes of the human structural connectome. We have clarified the motivating question in the Introduction and at the beginning of the relevant Results subsections to make the experimental progression easier to follow.

Changes in the revised manuscript: Introduction, lines 60–76; Results, lines 296–299, 344–345, 407–408, and 455–456.

 

Weakness 2 & Questions 1–3 — task complexity, significance, target audience, and future directions

We thank the reviewer for raising the questions of significance, intended audience, and future directions for our work.

Aim Instead of evaluating performance on complex ML tasks, our aim is to establish the computational role of intrinsic timescale placement with controlled recurrent network experiments.

Contribution We have shown that the contribution of intrinsic timescale heterogeneity to network computation depends not only on which time constants are present, but also on where they are located in the recurrent network. Alignment with slow recurrent modes improves early learning, while training recurrent weights partly compensates for unfavourable initial placement.

Audience Our work is intended for computational neuroscience researchers interested in studying neural timescales and recurrent dynamics, as well as machine learning researchers interested in studying temporal computation and learning in recurrent networks.

The tasks we use in our experiments are established paradigms in computational neuroscience designed to investigate principles of neural computation under controlled conditions. We studied the placement principle across four different types of memory and decision tasks, three different types of network topology, and network sizes of 64, 128, and 256 units. In addition, we clarified the contributions and scope of our findings in both the Introduction and Discussion. Possible future directions include extending the study to more complex temporal tasks and larger networks; studying learning effects beyond MemoryPro; and investigating how intrinsic neuronal properties interact with recurrent dynamics in biological circuits.

Changes in the revised manuscript: Introduction, lines 60–76 and 98–106; Discussion, lines 533–537.

Weakness 3 — interpretation of the human MEG analysis

We appreciate the reviewer's comment about the distinction between the measured timescales of cortical activity and the intrinsic time constants of individual neurons. MEG timescales reflect effective population dynamics shaped by intrinsic neuronal properties, recurrent interactions, and their combined effects. In an additional analysis, we used the empirical 400-region structural connectome as a recurrent model in order to investigate whether connectivity can generate the observed cortical timescale hierarchy. To this end, we set the intrinsic time constants of all regions to the same value (τ0=8\tau_0=8, g=0.90g=0.90).

When we analyse the effective timescales generated by the recurrent model, we find a positive correlation with empirical MEG timescales (Spearman ρ=0.523\rho=0.523, spatial rotation p=.063p=.063). This positive association suggests that recurrent connectivity may contribute to the empirical spatial ordering. Including the slow-mode participation in the recurrent model as an additional predictor increased the explained variance by ΔR2=0.1069\Delta R^2=0.1069 (spatial rotation p=.020p=.020). Importantly, assigning the empirical MEG timescale ordering to intrinsic time constants in the connectome model remains a modelling assumption. Our controlled recurrent network experiments demonstrate the computational consequences of intrinsic timescale placement, and the cortical analysis examines the association between measured activity timescales and slow recurrent organisation. The cortical findings therefore provide a complementary biological perspective on the role of slow recurrent organization.

Changes in the revised manuscript: Sec. 3.5, lines 466–474; Discussion, lines 522–531; Appendix H, lines 1150–1159.

Reviewer_bqr3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates how the placement of intrinsic timescales in a recurrent network shapes its dynamics, computation and learning. The authors first show that heterogeneous time constants enrich the temporal basis available to a linear readout, but that this advantage largely disappears once recurrence is added in their tasks, because recurrence alone generates diverse effective timescales. They then keep the graph, the weights and the set of time constants fixed and vary only which unit gets which time constant. They compare local placement coordinates (degree, in-strength) with a network-level coordinate: participation in slow eigenmodes of a homogeneous reference Jacobian. Aligning slow intrinsic timescales with slow-mode units broadens the spread of effective timescales more consistently than local coordinates across ER, heavy-tailed and modular reservoirs. It also improves four cognitive tasks, with larger benefits at longer delays, and speeds early learning when recurrence is trainable. Finally, they relate group-level MEG timescales across 400 Schaefer parcels to slow modes of the human structural connectome. They report a positive association, and that the empirical ordering yields a broader effective timescale spread than spatially rotated orderings.

Critical strengths
  1. The paper does a very good job of isolating the effect of timescale placement. Permuting a fixed set of time constants on a fixed network separates placement from timescale diversity and connectivity. The paired comparisons within networks, calibration on random placements, and testing across three topologies and three network sizes make the computational results convincing.
  2. The slow-mode coordinate is a principled, population-level quantity, compared with local graph measures like degree and in-strength. The paper shows that degree only matters to the extent that it tracks slow modes, whereas slow-mode alignment works consistently across topologies, tasks and learning. Together with the finding that heterogeneity alone helps little once recurrence is present, this offers a plausible answer to why diverse timescales might exist in cortex: what matters is where they are placed relative to recurrent structure.
Critical weaknesses
  1. Most of the evidence for why aligned placement helps is correlational. Mediation is not tested, since placement is manipulated directly, but the claim that a broader effective timescale spread is what improves performance rests on a partial correlation over seven alignment levels (Fig. 8). The authors acknowledge this doesn’t show mediation. Targeted controls or ablations could make the case much stronger. For example, compare placements matched on mean effective timescale (or slowest-mode decay time) but differing in S_eff, or selectively disrupt the slow modes.
  2. The learning advantage of alignment is transient. It nearly disappears once recurrence is trainable (Fig. 5c).
  3. The neuroscience analysis is a mixed bag, and its central assumption is questionable. MEG timescales aren’t intrinsic timescales, as was acknowledged. MEG timescales are population-level effective timescales, estimated from autocorrelations of regional activity, yet they are mapped onto the model’s intrinsic time constants. Numerous factors can affect the timescales measured based on MEG data; for example, the connectivity itself could produce slower effective timescales in regions that take part in slow connectome modes. If so, finding longer MEG timescales in those regions is close to circular and doesn’t show that intrinsic timescales are placed to match connectivity. The association is borderline significant (ρ = 0.565, spin p = 0.048) to make the 'cortical timescales align with slow mode structure' (Fig. 6b). Authors either need to change their model or use the appropriate data to assess their framework.
  4. The paper treats the spatial ordering of cortical timescales as fixed and maps MEG timescales onto intrinsic unit time constants. However, measured cortical timescales can change with task demands. This supports the concern that MEG timescales might largely reflect network dynamics shaped by connectivity. A discussion of this would help qualify the cortical claims.
Questions for authors
  1. Does the aligned benefit remain when controlling for the mean effective timescale or the slowest-mode decay time? Could you include ablations or matched-placement controls that test whether a broader S_eff causes the performance gain?
  2. How would you distinguish “intrinsic timescales are placed to match slow connectome modes” from “the connectome produces longer effective timescales in those regions”?

Suggestions for future work:

  1. If the intrinsic time constants are trainable, does slow-mode alignment emerge on its own?
  2. How would task-dependent changes in intrinsic timescales interact with the proposed placement principle?
5 comments

We sincerely appreciate the reviewer's acknowledgments on our timescale placement isolation, comparing population-level recurrence vs local structural properties, and the insights it provides for the cortex. We also thank the reviewer's emphasis on distinguishing correlational evidence vs mechanism, intrinsic neuronal properties vs generated activity timescales with recurrence. We conducted two additional mechanics experiments and learnable τ\tau analyses.

 

Weakness 1 & Question 1 — effective timescale spread and computational benefit

We agree with the reviewer that the association between effective timescale spread and task performance in our original framework does not show a mediation. We conducted two additional experiments using 24 independently generated 128-unit networks (eight per topology).

First, we compared timescale placements with matched mean log effective timescales but different spreads. Placements with greater spread showed lower PDM errors, indicating that the computational benefit cannot be attributed to differences in mean log effective timescales alone. Second, we conducted selective perturbations on slow and fast recurrent modes during the memory delay to examine their functional contributions. Slow-mode perturbation reduced the aligned placement benefit more than fast-mode perturbation, see summarisation below.

ExperimentDelayDifference (pp)95% CIp-value
Matched mean log timescale1600 ms+7.53[5.48, 9.44]<.001<.001
Matched mean log timescale2400 ms+4.42[2.89, 6.02]<.001<.001
Slow- vs fast-mode perturbation1600 ms+6.47[4.21, 8.82]<.001<.001
Slow- vs fast-mode perturbation2400 ms+2.57[0.59, 4.41]=.008=.008

Differences are in percentage points of PDM error. The first experiment compares lower vs higher spread placements; the second compares the reduction in aligned placement benefit with slow- vs fast-mode perturbation. These experiments provide further evidence for the functional relevance of effective timescale spread and the contribution of slow recurrent modes to the placement benefit.

Changes in the revised manuscript: Sec. 3.3, lines 374–377; Appendix E, lines 1008–1029.

 

Weakness 2 — recurrent learning reduces the placement advantage

We appreciate the reviewer's point on trainable recurrence reduces in aligned vs anti-aligned performance differences, which is observed in Figure 5c. We further observed that initially anti-aligned networks reorganized their recurrent structure during training, with increased alignment between their fixed intrinsic timescales and slow recurrent modes by 0.137 on average (Figure 5d). However, aligned placement still improved early learning, more so at longer delays (Figure 5a). These findings points to that recurrence training can compensate for unfavourable initial placement, and timescale placement serves as an inductive bias that improves learning efficiency.

Changes in the revised manuscript: Sec. 2.4, lines 219–229; Discussion, lines 514–521.

Weakness 3 & Question 2 — intrinsic timescales versus recurrence-generated activity timescales

We appreciate the reviewer for pointing out MEG based timescales reflect effective neural activity rather than intrinsic neuronal time constants. Recurrent connectivity itself could generate the observed cortical timescale hierarchy. To investigate this alternative, we conducted an additional analysis using the empirical 400 region structural connectome, with identical intrinsic timescales assigned to all regions and recurrent gain fixed at g=0.90g=0.90.

First, we compared the effective timescales generated by this homogeneous recurrent model with empirical MEG timescales. We observed a positive association (Spearman ρ=0.523\rho=0.523, spatial-rotation p=.063p=.063), which suggests that recurrent connectivity can reproduce part of the empirical spatial ordering. Second, we examined whether slow-mode participation provides additional predictive information in addition to the homogeneous model. Including the slow-mode score increased variance explained by ΔR2=0.1069\Delta R^2=0.1069 (spatial rotation p=.020p=.020). This was comparable across reference intrinsic timescales of τ0=4\tau_0=4, 88, and 1616 (ΔR2=0.1067\Delta R^2=0.1067, 0.10690.1069, and 0.10210.1021, respectively). These findings suggest that slow recurrent organization captures information about empirical cortical timescales beyond the homogeneous model. We acknowledge that this does not directly establish intrinsic τ\tau placement in the cortex, and it remains a biological hypothesis; our experiments instead demonstrate the computational consequences of the placement.

Changes in the revised manuscript: Sec. 2.5, lines 249–251; Sec. 3.5, lines 466–474; Appendix H, lines 1150–1159.

 

Weakness 4 — task dependent cortical timescales

Cortical activity timescales can vary with task demands and behavioural states. Our analysis uses a group average MEG timescale map and does not establish whether the observed cortical ordering remains stable in different cognitive tasks. The task dependent variations may reflect changes in local neuronal properties, recurrent dynamics, or their interaction. We have clarified this limitation in our interpretation and discussion, and we do not treat the measured activity timescales as fixed intrinsic neuronal time constants.

Changes in the revised manuscript: Discussion, lines 522–531.

Future Question 3 — does slow mode alignment emerge when intrinsic timescales are learned

We conducted additional experiments enabling intrinsic timescales to be learned in nine independently generated networks (three per topology), using MemoryPro with fixed or trainable recurrent weights. Learning intrinsic timescales improved task performance, but these conditions did not show positive alignment with slow recurrent modes (mean final Spearman ρ\rho = $-$0.017 and $-$0.027). This suggests that task optimization can improve performance without necessarily recovering the aligned timescale arrangement, and that timescale placement serves as an inductive bias for learning rather than a target configuration for convergence.

Changes in the revised manuscript: Sec. 3.4, lines 431–450; Appendix F, lines 1034–1043.

 

Future Question 4 — how would task dependent changes in intrinsic timescales affect the placement principle

We appreciate the reviewer's suggestion on this future direction. Previous studies have shown that measured neural timescales change with task demands (Zeraati et al., 2023) and that intrinsic timescales can adapt to different memory requirements (Khajehabdollahi et al., 2024). If intrinsic timescales change with task demands, their alignment with the recurrent structure may also adapt, potentially altering the spread of effective timescales and computational performance. Our experiments do not directly examine these task dependent changes in intrinsic timescales, which we now include as future work in the Discussion.

Changes in the revised manuscript: Discussion, lines 522–531.

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

The paper asks whether it matters where fast and slow neurons are placed in a recurrent network. The authors keep the network and the set of timescales fixed, and change only which neuron gets which timescale. Placing slow neurons where the network’s own slow activity is strongest gives a wider range of timescales and better performance on memory and decision tasks, especially at longer delays. This placement also speeds up early learning, though training partly makes up for bad placement. In human brain data, slower regions line up with the connectome’s slow activity patterns.

Critical strengths
  1. Changing only which neuron gets which timescale, while keeping the network and the set of timescales fixed, isolates the effect of placement clearly.
  2. Results are tested on many independent networks, three network types, and four tasks, with confidence intervals, effect sizes, and paired tests. The authors are also careful not to overclaim.
  3. Human data show slower regions align with the connectome’s slow patterns, suggesting the idea may apply beyond the model.
Critical weaknesses
  1. The slow mode score marks the neurons that already take part in slow activity. Giving those neurons slow timescales, and the rest fast ones, would naturally make slow parts slower and fast parts faster, which widens the spread of timescales. So Figure 3 may be close to automatic.
  2. The slow mode score uses the full weights and activity, while degree uses only the wiring. Its better performance may simply reflect using more information. A fairer comparison would include other measures that also use the weights, such as eigenvector centrality.
  3. When recurrence is trained, the benefit of mixed timescales almost vanishes, and training partly removes the placement effect. In standard trained RNNs, placement may therefore matter mainly early in learning.
  4. There is no comparison with simply letting each neuron’s timescale be learned. Would learned timescales find the aligned placement on their own
  5. The networks are small, the tasks are simple, and the learning experiment uses only MemoryPro.
Questions for authors
  1. Is the wider timescale spread from slow mode alignment in Figure 3, more than expected from making slow mode neurons slower? What would a simple slower neurons get slower model predict?
  2. How does the slow mode score compare with other measures that use the weights, such as eigenvector centrality?
  3. If each neuron’s timescale is learned along with the recurrent weights, does training end up with an aligned placement?
  4. Do the learning results hold for tasks other than MemoryPro?
3 comments

We sincerely appreciate and are encouraged by the reviewer's recognition of our controlled placement design and our experimental depth. Here, we provide additional analyses addressing the theoretical intuition, weighted centrality baselines, and recurrent learning.

 

Weakness 1 & Question 1 — is the widening of effective timescales expected

We agree that the widening of effective timescales is expected in simple linear dynamical systems. In a simplified linear system with two independent modes, where the recurrent modes have feedback strengths μslow\mu_{\mathrm{slow}} and μfast\mu_{\mathrm{fast}} (μslow>μfast\mu_{\mathrm{slow}} > \mu_{\mathrm{fast}}), the effective decay time approximately follows Ti≈τi/(1−μi)T_i \approx \tau_i/(1-\mu_i). Here, aligning longer intrinsic timescales (higher τi\tau_i) with the slower recurrent mode (μslow\mu_{\mathrm{slow}}) reinforces temporal persistence and increases the separation between slow and fast decay times. We therefore do not claim that this widening is unexpected.

The question in our nonlinear recurrent systems is whether this organization affects computation and which recurrent modes contribute to the computational benefits. We conducted two additional controls to examine these questions. First, we compared placements with matched mean log effective timescales but different timescale spreads. Placements with greater spread resulted in 7.53 [5.48, 9.44] percentage points lower PDM error at 1600 ms (p<.001p<.001). This shows that the computational benefit cannot be explained by differences in mean log effective timescale alone. Second, we selectively perturbed slow and fast recurrent modes during the memory delay, with perturbations calibrated to comparable strength. Slow-mode perturbation reduced the alignment advantage more than fast-mode perturbation (difference = 6.47 [4.21, 8.82] percentage points at 1600 ms, p<.001p<.001). This provides evidence for the functional contribution of slow recurrent modes.

Changes in the revised manuscript: Sec. 3.3, lines 374–377; Appendix E, lines 1008–1029.

 

Weakness 2 & Question 2 — weighted eigenvector centrality baseline

We thank the reviewer for suggesting a weight-based measure. We added weighted eigenvector centrality, computed from the absolute recurrent weights, as an additional placement baseline. We evaluated this control in 24 independently generated 128-unit recurrent networks, eight per topology (Erdős–Rényi, heavy-tail, and modular). Within each network, we kept the recurrent weights, intrinsic timescale set, and PDM evaluation fixed, and compared timescale assignments based on slow-mode participation versus weighted eigenvector centrality.

For each placement rule, we measured the PDM performance benefit (anti-aligned error minus aligned error), following the definition in the manuscript. At 1600 ms, slow-mode placement showed an error reduction of 8.34 percentage points, compared to 3.15 percentage points for weighted eigenvector centrality (paired difference = 5.19 [2.76, 7.47], p<.001p<.001). This advantage remained at 2400 ms (paired difference = 2.59 [0.13, 4.91], p=.003p=.003). Slow-mode placement also produced a greater increase in effective timescale diversity (mean difference in ΔSeff=0.0593\Delta S_{\mathrm{eff}}=0.0593), with positive differences in 23 of 24 networks.

These analyses show that the advantage of slow-mode placement is not simply due to using recurrent connectivity weights as a placement criterion.

Changes in the revised manuscript: Sec. 3.2, lines 334–339; Sec. 3.3, lines 365–370; Appendix E, lines 997–1006.

Weakness 3 — recurrent learning reduces the placement advantage

We agree that recurrent training can compensate for unfavourable timescale placement; in Figure 5c, trainable recurrence reduced the placement gap by 0.090 [0.046, 0.136] on average across 24 networks. We also observed that initially anti-aligned networks reorganised their recurrent dynamics during training, increasing the alignment between their fixed intrinsic timescales and evolving slow recurrent structure by 0.137 [0.108, 0.166] on average (Figure 5d). These results show that recurrent plasticity can partly compensate for unfavourable placement. Nevertheless, aligned placement accelerated early learning, particularly at longer delays (Figure 5a). We therefore interpret timescale placement as an initial inductive bias that improves learning efficiency.

Changes in the revised manuscript: Sec. 2.4, lines 219–229; Discussion, lines 514–521.

 

Weakness 4 & Question 3 — do learned intrinsic timescales naturally align with slow recurrent modes

We appreciate the reviewer's suggestion. We conducted additional experiments where we allowed neuronal intrinsic timescales to be learned, under both fixed and trainable recurrent weights. We evaluated nine independently generated networks (three per topology) using the MemoryPro task. With fixed recurrence, networks with learned τ\tau achieved lower error than with fixed τ\tau (mean reduction = 0.1365, p=.004p=.004). When recurrent weights were also trainable, learning τ\tau provided a much smaller additional improvement (reduction = 0.00109). Neither condition showed positive mean final alignment between learned τ\tau and slow recurrent modes (Spearman ρ=−0.017\rho=-0.017 and −0.027-0.027, respectively).

These results show that task optimization can improve performance without recovering slow-mode alignment. Aligned placement therefore serves as an inductive bias in our setup, rather than a convergence target for the networks.

Changes in the revised manuscript: Sec. 3.4, lines 431–450; Appendix F, lines 1034–1043.

Weakness 5 & Question 5 — task complexity, network size, and generalisation

We appreciate the reviewer's concern about task complexity and generalisation. Our experiments were designed to establish the computational role of intrinsic timescale placement through a progression of increasingly complex settings. We first examined how intrinsic heterogeneity and recurrence shape temporal representations using two diagnostic tasks. We then isolated the effect of timescale placement under fixed recurrent weights and evaluated its computational consequences across four memory and decision tasks and three network topologies. Additional analyses with 64-, 128-, and 256-unit networks tested robustness to network size. Finally, we allowed recurrent weights to adapt and examined how initial placement influences learning.

This progression provides evidence that the placement effect is not restricted to a particular task, topology, or tested network size. Our recurrent learning experiments focused on MemoryPro, and extending this framework to more challenging tasks and larger networks is a natural next step to further evaluate its generalisation.

Changes in the revised manuscript: Discussion, lines 533–537.

Comments on the paper

1 comment

We thank the reviewers for their careful evaluations and constructive feedback. We appreciate their recognition of our controlled timescale placement design and their thoughtful questions about the underlying mechanisms, motivation, and biological interpretation of our findings. In response, we have conducted additional analyses and revised the manuscript to address these points. The main revisions, all highlighted in blue, are summarised below:

Improved clarity: We have revised the Introduction, Results, and Discussion to clarify the research question, experimental progression, and significance and scope of our findings.

New weighted centrality baseline (Secs. 3.2–3.3, Appendix E): We compare slow-mode placement with weighted eigenvector centrality, showing that its dynamical and computational advantages are not simply explained by recurrent connectivity weights.

New functional controls (Sec. 3.3, Appendix E): We implement matched mean log effective timescale comparisons and selective perturbations of slow and fast recurrent modes, providing additional evidence for their functional contributions to temporal computation.

Additional learning experiments (Sec. 3.4, Appendix F): We investigate training intrinsic time constants and further clarify how recurrent plasticity partly compensates for unfavorable initial placement, with alignment providing an advantage during early learning.

Extended cortical analysis (Sec. 3.5, Appendix H): We implement a homogeneous timescale connectome control and assess the additional predictive contribution of slow recurrent modes; we also clarify the distinction between measured MEG activity timescales and intrinsic neuronal time constants.

We believe these revisions have strengthened the manuscript and clarified the contributions and interpretation of our findings, and we hope the reviewers will find our responses and revised manuscript satisfactory.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.