acceptodds
Under review as a conference paper at ICLR 2027

Structure Before Flexibility: Mechanistically Testable Functional Conductances for Continuous-Time Network Dynamics

Abstract

Continuous-time neural models can fit trajectories accurately while assigning the wrong contributions to individual interactions. We introduce SP-KAN-LTC, which fixes the conductance and driving-force factorization on each known edge and learns only a positive bounded state-dependent conductance, represented by a univariate spline network. A matched two-by-two factorization-by-function-class comparison shows that factorization contributes far more to mechanism recovery than the edge function class. In an eight-edge controlled benchmark with known mechanisms, SP-KAN-LTC reduces edge-current and intervention-effect error by a median of 99.8% and 99.5% relative to graph-matched and capacity-matched Current-KAN, with intervention rank correlation rising from -0.70 to +1.00. Within the matched perceptron family, the corresponding reductions are 99.1% and 98.9%, showing that the attribution failure is not specific to the spline parameterization. Against MLP-Conductance, the flexible spline edge contributes a smaller and conditional effect: no consistent advantage is established at four interactions, whereas intervention-effect error improves by 65.9% at eight. On measured electrophysiology, every factorized run recovers the independently recorded synaptic current more accurately than every unfactorized run, though SP-KAN-LTC does not consistently outperform the matched perceptron. Aliasing, driving-force misspecification, hidden interaction memory and held-out optogenetic transfer delimit the interpretation: accurate trajectories do not guarantee identifiable or structurally valid edge mechanisms, and external transfer is not demonstrated on C. elegans. The results support a structure-before-flexibility principle: constraining what an interaction is allowed to mean matters more for mechanistic fidelity than the flexibility used to parameterize it.

Primary area:
learning on time series and dynamical systems
Keywords:
Continuous-time neural networks, Liquid Time-Constant Networks, Kolmogorov-Arnold Networks, structure-preserving learning, functional conductances, mechanistic interpretability, system identification, interaction dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mJUxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SP-KAN-LTC (Structure-Preserving KAN Liquid Time-Constant network).

For any two states, SP-KAN-LTC factorizes the interaction between them as the product of a nonnegative state-dependent interaction strength and an explicit driving force, i.e., Iij=cij(hj)(Eij−hi)I_{ij} = c_{ij}(h_j) ( E_{ij} - h_i ), as mentioned in Eq. (1), and then uses a univariate KAN function to learn the nonnegative and bounded state-dependent interaction strength cijc_{ij}.

In the SP-KAN-LTC, the authors refer to the factorization as the structure and KAN (when compared to the simpler MLP) as the flexibility. By comparing the performance of SP-KAN-LTC (structure + flexibility), MLP-conductance (structure only), current KAN (flexibility only), and current MLP (no structure and no flexibility), the authors find that structure is more important than flexibility.

Critical strengths
  1. The experiments can support the claim "structure before flexibility" in certain scenarios. The experiments include four central methods, i.e., SP-KAN-LTC (structure + flexibility), MLP-conductance (structure only), current KAN (flexibility only), and current MLP (no structure and no flexibility), to compare the contribution of structure and flexibility.

  2. The claim "structure before flexibility" is intuitive and interesting. It may help design space-complexity-friendly neural network architectures in the future.

Critical weaknesses
  1. The experimental verification of the claim "structure before flexibility" is limited to a small number of experiments at a limited scale. It is unclear whether this phenomenon is general or specific.

  2. The claim "structure before flexibility" is conditional. The choice of flexibility is a classical trade-off from the perspective of machine learning theory. On the one hand, higher flexibility may increase the expressiveness of the model and reduce the approximation error. On the other hand, higher flexibility indicates a more complex learning space, which may increase the difficulty of learning and increase the estimation error. Thus, the claim "structure before flexibility" may only be meaningful when higher flexibility can result in a lower excess risk.

Questions for authors

See the weakness part.

3 comments

Thank you for pointing out the approximation and estimation issue. We agree that it is the right way to think about the KAN-versus-MLP comparison, and the Conclusion now says so.

Scale and scope

The controlled benchmark is small on purpose, because we need the true conductance, edge current and intervention effect for every edge. We repeat the comparison across three law families and two interaction counts, and we also evaluate a 75-node contact graph, a Morris–Lecar system with hidden memory, a measured half-center oscillator and C. elegans. We agree that this does not establish a universal result. The paper now limits the mechanism-recovery claim to settings where the factorized interaction form is appropriate (Table 1).

Approximation and estimation versus structure

Within the factorized family, the KAN-versus-MLP result behaves like the usual approximation and estimation trade-off. SP-KAN-LTC helps at eight interactions but not at four, and it is not consistently better on the measured systems (Section 5.2).

The factorization result is different. We increased the capacity of the unfactorized perceptron from 33 to 129 parameters per edge. Its rollout error improves in every law family, for example from 0.0076 to 0.0049 under mixed-shape laws, while its edge-current error increases from 1.19 to 1.87. At the validation-selected width, its rollout error is not distinguishable from MLP-Conductance in two of three families, but its edge-current error is still at least 115 times larger (Section 5.1, Appendix N). The aliasing experiment gives the same separation without changing capacity: as two edge references become indistinguishable, rollout and grouped-current recovery stay accurate while the individual-edge conductance error increases about 25-fold (Section 5.3).

So we agree that "more flexible is better" is subject to the usual learning trade-off. The narrower point of "structure before flexibility" is that low predictive error does not by itself recover the generating edge decomposition (Section 6).

We would be interested to know whether this distinction addresses your concern about the conditional nature of the claim.

Thanks for the response. After reading all the reviews and the corresponding comments, I think that the claim and verification of this paper are conditional, and it is unclear how to check whether the condition of the claim is hold in practice. Thus, I would maintain my recommendation unchanged.

Thank you for reading the responses. We agree that the mechanistic claim has a stated scope, but this scope is not meant to be inferred from the fitted model itself. Different structural assumptions need independent support, for example from known interaction physics, graph structure, reversal references, recorded currents or intervention data. The paper asks whether preserving such independently supported structure improves recovery of the corresponding mechanism.

Within that scope, the revised paper already includes several practical diagnostics and stress tests for these assumptions.

Individual-edge separation. The conditioning of the edge decomposition is computed from the training states, without fitting a model. The conditioning statistic in Section 5.3 and Appendix A.5 falls from 0.087 to numerically zero as two edges become indistinguishable. Over the same sweep, individual-edge conductance error increases about 25-fold while rollout changes by only 8%.

Instantaneous edge law. The paper stress-tests this assumption against a history-capable baseline. In the hidden-memory experiment, SP-KAN-LTC is the best of the four models when the interaction is exactly instantaneous (η=0\eta = 0). At η=0.10\eta = 0.10, CfC is already better on interaction current, rollout and intervention effect (Section 5.3).

Edge references. The paper tests whether these can be learned when the leak physics is known. Learning them from a uniform +0.05Rh+0.05 R_h shift brings SP-KAN-LTC conductance error back to 1.1 to 1.9 times its known-reference level (Appendix P.1). When the leak physics is also free, however, the data admit compensatory decompositions rather than uniquely recovering the generating one.

Conductance form. The oscillator results show that this cannot be established from voltage fit alone: factorized and unfactorized models fit the voltage comparably, so the form has to come from independent knowledge, here the dynamic clamp that imposed it. For C. elegans the interaction law is unknown and there is no edge-level mechanism target, so we do not claim mechanism recovery there.

We therefore view the result as a structure-preserving principle with a clear scope, not as a condition to be discovered from predictive fit. When the interaction form is independently justified, preserving it substantially improves mechanism recovery: even the widened Current-MLP selected on validation still has 115 to 427 times larger median edge-current error across the three controlled law families. When the structural assumptions are not justified, our boundary experiments show why predictive fit alone is insufficient for a mechanistic claim.

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

The paper "Structure Before Flexibility: Mechanistically Testable Functional Conductances for Continuous-Time Network Dynamics" tackles a fundamental problem in scientific machine learning and neural ordinary differential equations (ODEs): accurate trajectory prediction does not mean a model has learned the correct internal mechanisms. A model can successfully fit overall system curves while completely misattributing the contributions of individual internal components. To fix this, the authors propose a "structure-before-flexibility" principle, supported by their experiments, arguing that constraining what an interaction is allowed to mean biophysically is far more important for mechanistic recovery than using unconstrained, highly flexible function approximator

Critical strengths

The paper introduces a specific architecture called Structure-Preserving KAN Liquid Time-Constant Network (SP-KAN-LTC). It uses univariate Kolmogorov-Arnold Networks (KANs) only to model the state-dependent conductance, while strictly preserving a mathematically fixed biological factorization structure. Though KAN based architectures yield very similar results to MLP based (not surprisingly), the paper addresses a blind spot in biological neural modeling by demonstrating that getting the right trajectory fit does not mean the the model has learned the correct internal causal mechanisms or edge currents. Furthermore, it backs this up with a rigorous empirical evaluation across controlled synthetic benchmarks, empirical contact networks, real electrophysiology, and C. elegans optogenetic data. Stengths are:

  1. Systematic evaluation of conductance structure. The paper goes beyond constraining network connectivity and specifically examines the role of conductance factorization in recovering edge-level contributions. The controlled comparisons help distinguish structural effects from those of the edge function class.
  1. Mechanism-oriented evaluation. By evaluating conductance recovery, edge-current errors, and intervention effects and rankings, the authors provide a useful framework for assessing the mechanistic fidelity of neural network models and investigating the mechanisms by which neural circuits implement specific functions. 3.KAN and conductance modeling. Combining KANs with conductance-based dynamics is an interesting and potentially useful approach, allowing flexible interaction laws while preserving a biologically motivated structure. 4.Comprehensive experiments and transparent reporting. The authors examine identifiability, structural misspecification, hidden memory, and intervention transfer. I particularly like their reporting of negative results and of the limitations of the proposed method.
Critical weaknesses

One of the hidden implications of the paper is that connectomics and type of ionic channels are very important for models of neural circuits. It may be useful to emphasize this even more. A weakness is that the SP-KAN-LTC are modestly better than the MLP alternative: should the title reflect this? What is important is the conductance factorization not MLP vs KAN. The paper reports different relationships between predictive performance and mechanistic fidelity across experiments. In the controlled benchmark, factorized models outperform unfactorized models in both rollout accuracy and mechanism recovery. However, in the measured half-center oscillator, rollout errors are comparable while hidden-current errors differ substantially. Mi et al. (2022) also reported a predictive advantage from conductance-based modeling. I think the authors could better explain these differences and clarify under what conditions conductance structure benefits prediction, mechanism recovery, or both.

Matching the number of parameters does not guarantee comparable expressivity or optimization difficulty, as acknowledged in Appendix A.4. If predictive improvement is an important benefit, comparisons across model capacities and training budgets would be useful. If the main claim concerns mechanistic fidelity independently of predictive accuracy, prediction-matched comparisons in the controlled benchmark would provide stronger evidence. The electrophysiology experiment partially addresses the latter, but does not fully resolve the concern.

More broadly, the conclusions about the importance of conductance factorization are primarily supported by comparisons with two relatively small unfactorized architectures, on systems whose generating laws largely follow the assumed structure. It would be valuable to examine a broader range of unfactorized models and clarify how generally the reported structural advantage can be expected to hold.

Since the positive mechanism-recovery results mainly come from systems whose interaction laws are designed to follow the assumed conductance structure, it remains unclear whether these advantages generalize to real biological circuits with unknown interaction laws. Mi et al. (2022) reported positive results for conductance-based modeling in a similar biological setting, although the experimental setups and datasets differ. It would therefore be valuable to investigate under what conditions conductance structure provides meaningful benefits in real neural circuits. Another concern is the flexible nuisance MLP used in the electrophysiology experiment (Appendix L.5). Although its architecture is shared across models, each model trains its own nuisance branch, which may absorb part of the synaptic contribution. An ablation varying its capacity could help determine whether the reported current-recovery advantage is robust.

Questions for authors

1)In the controlled benchmarks and biological applications, how sensitive is SP-KAN-LTC to errors in the driving-force references E_i? 2)What happens if I estimate the E_i in addition to the c_ij? 3) The main controlled neural benchmark contains only four state variables and four or eight interactions. Although the paper includes a larger 75-node SIS benchmark, It would be valuable to evaluate mechanism recovery in larger synthetic neural circuits with more complex interactions, where identifiability and attribution may become more challenging, especially since many biological neural circuits of interest operate at much larger scales. Is this possible, at least in principle? 4) The paper is sometimes difficult to follow because experimental details are scattered across sections and appendices, while the main text frequently relies on abstract descriptions. More explicit descriptions of the ground-truth interaction laws, network configurations, and experimental procedures would improve clarity and reproducibility.

2 comments

Thank you for the constructive review. We ran the two controls you suggested: a widened unfactorized control and a nuisance-capacity sweep on the half-center oscillator.

Framing, title and connectomics

We agree that the main result is about the factorization, not a general KAN-versus-MLP advantage. We kept the title because it already puts structure first, and the abstract and contribution list now say this explicitly. We also agree that the results put weight on knowing the graph and the interaction type, which is where connectomic and biophysical knowledge comes in. Table 1 now states what structural information each benchmark gives the model.

Prediction versus mechanism recovery

The predictive gap depends on capacity. When Current-MLP is widened to five times the per-edge capacity, its rollout error is no longer distinguishable from MLP-Conductance in two of three law families, while the mechanism gap stays large in all three. On the oscillator, the five models already have similar voltage rollout errors, while their hidden-current recovery differs strongly. Mi et al. work in a different setting, so we do not claim to explain their result, but ours show that a predictive advantage alone does not tell whether the decomposition is recovered (Section 6).

Widened Current-MLP

We widened Current-MLP to 33, 65 and 129 parameters per edge and selected the width whose validation rollout error was closest to MLP-Conductance (Section 5.1, Appendix N). At 65 parameters per edge, rollout error is not distinguishable from MLP-Conductance in the mixed-shape and mixed-monotonic families. Its median edge-current error is still 115 and 184 times larger in those two families, and 427 times larger under native-exact laws. From 33 to 129 parameters, rollout improves while edge-current recovery gets worse.

Broader unfactorized models

The width sweep covers larger edge-local current maps with the same information as the factorized models, and the hidden-memory experiment also includes CfC and an augmented neural ODE. We do not claim the result for every unfactorized architecture. For models without an edge-local decomposition, edge current is not directly defined, so intervention effects provide the comparable mechanism-facing target.

Unknown-law biological systems

We agree that this remains a limitation. C. elegans is the only measured setting here with an unknown edge law, and no conductance model is above chance on both transfer measures (Appendix I.1). We use it as a test of external validity, not as evidence of mechanism recovery.

Nuisance-branch capacity

We repeated the oscillator comparison with nuisance widths of 8, 32 and 64 (Appendix H.2). At every width, both factorized models reduce hidden-current error relative to Current-MLP by 51.6% to 65.7%, in 10/10 seeds. The gap narrows mainly because Current-MLP improves, while the factorized errors stay between about 0.45 and 0.53. MLP-Conductance's learned maximum conductance also increases with nuisance width while its current error decreases, which argues against the branch simply absorbing the synaptic contribution. At width 64, MLP-Conductance has lower hidden-current error than SP-KAN-LTC in 10/10 seeds.

Q1. Sensitivity to EE

A +0.05Rh+0.05 R_h shift of every recurrent edge reference increases SP-KAN-LTC conductance error about elevenfold, while rollout error changes much less. MLP-Conductance shows the same pattern at ±0.10Rh\pm 0.10 R_h (Section 5.3). On the oscillator, Esyn=−80E_{syn} = -80 mV is the value used by the clamp itself.

Q2. Can EE be learned?

Yes, when the leak physics is known. From a +0.05Rh+0.05 R_h shift, learning the edge references reduces SP-KAN-LTC conductance error by 83.4% to 94.7%, and a learning-rate-matched control confirms that the gain comes from the learned reference (Appendix P.1). When the leak conductance and equilibrium are also trainable, runs move to compensatory decompositions with lower validation loss but much worse edge-current recovery (Appendices P.2 and P.3).

Q3. Larger circuits

In principle, yes. The edge function is univariate, and the parameter count grows linearly with the number of active edges. We also evaluate on a 75-node measured contact topology with simulated dynamics. This is not the same as mechanism recovery in a large neural circuit, where identifiability is an additional constraint beyond parameter count. Our aliasing experiment shows that individual edges cannot be separated when their driving-force signatures coincide (Appendix A.5), but we did not measure how often this becomes limiting in larger circuits.

Q4. Presentation

Table 1 now summarizes the scale, interaction law, given structure and evaluation target of every benchmark, and Table 4 gives every controlled edge law with its parameters.

Reviewer_vGLrOpenReview ↗

Answer

I thank the authors for their detailed answer and additional experiments. I will keep my positive rating.

Reviewer_3Pt5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SP-KAN-LTC, a continuous-time network in which every known edge keeps a fixed conductance × driving-force form, Iij=cij(hj)(Eij−hi)I_{ij}=c_{ij}(h_j)(E_{ij}-h_i), and only the bounded positive conductance cijc_{ij} is learned with a univariate B-spline KAN. A 2×2 design (factorized vs. unfactorized × KAN vs. MLP edge, 13 parameters per edge) on a 4-state/8-edge synthetic benchmark shows that the factorization, not the function class, drives recovery of edge currents and intervention effects. The KAN edge helps only conditionally (it hurts at 4 edges and helps at 8). Further experiments vary one assumption at a time (reversal aliasing, misspecified EijE_{ij}, hidden synaptic memory in Morris–Lecar, dense vs. sparse graph). Other tests: a semi-synthetic 75-node contact network, a measured half-center oscillator whose synaptic current was withheld, and a C. elegans optogenetic transfer test, which comes out null.

Critical strengths
  1. Evaluation against external mechanism targets rather than trajectory fit. The paper scores edge currents, edge-removal interventions and an independently recorded synaptic current, and treats the decomposition as a hypothesis to be tested. That is the right standard for "interpretable" dynamical models and is rare in this literature.
  2. Honest boundary experiments. The aliasing (ΔE→0\Delta E\to0), driving-force misspecification and hidden-memory (η\eta) studies show clearly that a good fit does not certify an edge mechanism. Negative results (no KAN advantage at 4 edges or on measured data; null C. elegans transfer) are reported plainly, with seed-paired statistics.
Critical weaknesses
  1. The proposed component brings no reliable gain, and the large effects come from a factorization that is not new. The headline reductions (99.8% and 99.5%) compare factorized with unfactorized edges. That factorization is inherited from native LTC (App. A.1: "the factorization itself is preserved rather than introduced"). Its value for predicting held-out activity and withheld synaptic currents was already reported by Mi et al. (2022) and, on this same half-center-oscillator dataset, by Burghi et al. (2025). What is new is the spline conductance, and it is not better than a capacity-matched MLP conductance:

    • At 4 edges the median effect favours the MLP (−63.9%, CI [−92.6, +10.2]).
    • On the measured oscillator it is worse in one transfer direction (+15% error, p=0.008, 8/10 seeds) and tied in the other.
    • On C. elegans it has the lowest top-5 overlap, 0.236, below the 0.267 random-set expectation.
    • It has higher pointwise conductance error than the MLP in most empirical-topology conditions.

    The one positive result (+59.8% at 8 edges) is not explained by law complexity or node degree (Sec. 5.2). So neither the method nor the "structure before flexibility" principle is a new contribution.

  2. The principle is shown only where the assumed edge form is exactly true.

    • Synthetic benchmarks: the generators use the factorized form, and the model is given the true graph, reversal references and conductance cap.
    • "Measured" system: this is a dynamic-clamp dataset (App. L.5). In dynamic clamp, the synaptic current is computed online from a conductance-based synapse model, so the target is factorized by construction. The factorized vs. unfactorized gap is again a well-specified vs. misspecified comparison. Even there, the best hidden-current error is 0.46–0.50 (relative L2), with waveform r = 0.93–0.97: the shape is recovered, the magnitude is not.
    • Fragility: shifting every EijE_{ij} by 5% of the state range raises conductance error 11× (0.0096 → 0.106).
    • Unknown laws: in the only system where the edge law is genuinely unknown (C. elegans, App. I), no model transfers, the true connectome does no better than a shuffled one, and a linear ODE does as well.

    With 4-state benchmarks, the paper never shows that "mechanistically testable" edges can be obtained in a setting where one would need them.

Questions for authors
  1. (Could change my assessment.) Is there any setting where SP-KAN-LTC consistently beats MLP-Conductance, with an explanation of why? If not, what is the method's contribution beyond the (prior) factorization?
  2. (Could change my assessment.) What synapse law and reversal potential did the dynamic clamp implement in the Burghi et al. recordings? If it is g s(Vpre)(E−Vpost)g\,s(V_{pre})(E-V_{post}) with E≈−80E\approx-80 mV, the half-center test is well-specified by construction. Please state this in the paper.
  3. (Clarification.) Can EijE_{ij} be learned jointly, or with a prior, while keeping mechanism recovery? The misspecification study suggests it cannot.
  4. (Clarification.) Why is hidden-current error ≈0.5 even for the best factorized models when the waveform correlation is ≈0.97? Is cmaxc^{max} unidentifiable from voltage alone?
2 comments

Thank you for the careful review. We agree with three of your points, and the paper now states them directly: the KAN edge does not consistently outperform MLP-Conductance, the factorization itself is prior work, and the oscillator current was generated through the same factorized form. The new controls also show where the flexible conductance helps and that the mechanism gap remains when prediction becomes comparable in two of three law families.

Q1. Where does SP-KAN-LTC outperform MLP-Conductance?

At eight controlled interactions. Pooled across law families within each seed, SP-KAN-LTC improves on MLP-Conductance by a median +59.8% in conductance error and +65.9% in intervention-effect error (95% CIs [+22.1, +66.5] and [+29.2, +86.3], Section 5.2). At four interactions both intervals include zero. With learned edge references, SP-KAN-LTC also recovers them more accurately in all three law families (p≤0.008p \le 0.008, Appendix P.1). The cause is open: an occupancy explanation was not supported (Appendix O).

The advantage is not consistent outside the controlled benchmark. On the 75-node contact topology, pointwise conductance error favors MLP-Conductance in most conditions while realized edge-flow error favors SP-KAN-LTC (Section 5.2). At nuisance width 64 on the oscillator, MLP-Conductance has lower hidden-current error than SP-KAN-LTC in 10/10 seeds (Appendix H.2). On C. elegans, SP-KAN-LTC is below chance on top-five overlap but weakly above chance on rank agreement (Appendix I.1). We agree that no experiment shows edge-level recovery in a measured system with an unknown law; the contact dynamics are simulated, and C. elegans has no edge-level target (Section 4, Table 1).

What is new beyond the inherited factorization?

Relative to native LTC, we replace the two-parameter monotone gate with a flexible bounded positive conductance, a spline in SP-KAN-LTC and a perceptron in MLP-Conductance (Appendix A.1). Under the mixed laws, native LTC has 6.5 to 31 times larger mean conductance error than either flexible model (Table 5), though not at matched capacity, and optimization also limits the native gate (Appendix F).

The main contribution does not depend on KAN beating MLP. Mi et al. showed that conductance-based models predict held-out activity better; we ask whether the mechanism gap remains once prediction is comparable. We widened Current-MLP from 13 to 33, 65 and 129 parameters per edge (Section 5.1, Appendix N). At 65 parameters, its test rollout error is not distinguishable from MLP-Conductance in two of three law families, while its median edge-current error remains 115 to 427 times larger.

On the oscillator data, Burghi et al. trained on segments of both trials; our contribution there is the controlled cross-trial comparison, not the direction of the effect.

Q2. What law generated the half-center oscillator current?

The dynamic clamp used Isyn=g(Vpre)(Esyn−Vpost)I_{syn} = g(V_{pre})(E_{syn} - V_{post}) with Esyn=−80E_{syn} = -80 mV and the sigmoid activation of Burghi et al. (Eq. 25). On the deposited voltages, this law reproduces the recorded current with Pearson correlation above 0.999999 on both trials (Section 5.4, Appendix H.1).

So, like the synthetic benchmarks, it tests a narrower question: when the form is correct, which decomposition does training on voltage alone select? All five models fit voltage similarly. Yet among the unfactorized model–width combinations with retained prediction traces, the predicted current has a negative inner product with the recorded current in 116 of 120 edge-runs, and in none of the 280 factorized edge-runs (Appendix H.3). A positive conductance times the driving force cannot make this sign error.

Q3. Sensitivity to EE, and can EE be learned?

The sensitivity is real but not specific to the spline: MLP-Conductance shows the same pattern at shifts of ±0.10Rh\pm 0.10 R_h. With the leak physics known, the edge references can be learned: from a +0.05Rh+0.05 R_h shift, SP-KAN-LTC reduces conductance error by 83.4% to 94.7%, and a learning-rate-matched control confirms the gain (Section 5.3, Appendix P.1).

Learning the leak parameters as well is much harder. Even when started at the true values, all ten runs leave the generating leak physics for lower-loss compensatory decompositions, with a median edge-current error 62 times larger (Appendix P.3). A prior on the leak may be needed, which we did not test.

Q4. Why is the current error near 0.5?

For MLP-Conductance and Sigmoid-Conductance it is mostly amplitude: the best nonnegative rescaling removes 86% and 94% of the squared error. For SP-KAN-LTC it removes 53%, so its error is part amplitude and part shape. We do not claim that the conductance scale is identifiable from voltage alone (Appendix H.3).

Thank you again for the detailed review. We wanted to briefly follow up on Q1 and Q2, which you noted could change your assessment. In short, SP-KAN-LTC shows an advantage over MLP-Conductance at eight controlled interactions, with pooled median improvements of +59.8% in conductance error and +65.9% in intervention-effect error, both with intervals excluding zero, but not at four interactions. The mechanism gap also remains when the widened Current-MLP predicts comparably in two of three law families, with 115 to 427 times larger edge-current error.

For Q2, the clamp law and Esyn=−80E_{syn} = -80 mV are now stated explicitly (Section 5.4, Appendix H.1), and the predicted current points against the recorded current in 116 of 120 unfactorized edge-runs with retained prediction traces, versus 0 of 280 factorized edge-runs.

If anything in our response is still unclear, or if you feel there is a concern we have not fully addressed, we would be happy to clarify it before the discussion closes.

Comments on the paper

1 comment

We would like to thank the reviewers for the detailed reviews. We revised the paper around the main concerns that were raised, and all changes are marked in blue in the revised PDF.

The paper now states its scope more clearly. The conductance factorization is inherited from prior conductance-based models and from native LTC, and we do not claim it as new. The half-center oscillator current was generated by a dynamic clamp that already follows the factorized conductance form. SP-KAN-LTC does not consistently outperform MLP-Conductance, so the main result is not a general KAN-over-MLP advantage.

We added several controls. We widened Current-MLP to 33, 65 and 129 parameters per edge and selected the width whose validation rollout error was closest to MLP-Conductance; its test rollout error was then not distinguishable from MLP-Conductance in two of three law families (Section 5.1, Appendix N). We varied the nuisance-branch capacity on the oscillator (Appendix H.2). We added the exact dynamic-clamp specification and analyzed the amplitude and direction of the recovered current (Appendices H.1 and H.3). We also tested learning the edge references and jointly learning the edge and leak parameters (Section 5.3, Appendix P), scoring on common state support (Appendix O), and a calibrated chance baseline for C. elegans (Appendix I.1).

Table 1 now summarizes the interaction law, the information given to the model and the evaluation target for every benchmark, and Table 4 gives the exact controlled edge laws.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.