- 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., , as mentioned in Eq. (1), and then uses a univariate KAN function to learn the nonnegative and bounded state-dependent interaction strength .
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
-
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.
-
The claim "structure before flexibility" is intuitive and interesting. It may help design space-complexity-friendly neural network architectures in the future.
Critical weaknesses
-
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.
-
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
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 (). At , 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 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.
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.