acceptodds
Under review as a conference paper at ICLR 2027

Activation-Flexible ANN-to-SNN Conversion with Finite-State Markov Neurons

Abstract

Most ANN-to-SNN conversion methods rely on a specific correspondence between the source activation and the spiking neuron dynamics. We propose a finite-state continuous-time Markov chain (CTMC) neuron framework whose stationary spike flux can approximate every continuous nonnegative monotone activation function on a compact interval. For a generalized CTMC family with affine input-dependent transitions, we prove uniform approximation to arbitrary accuracy over this function class and derive an explicit approximation error bound. In practice, two- and three-state CTMCs fit ReLU, sigmoid, softplus, and clipped ReLU on the evaluated input ranges, and we evaluate corresponding MLP conversions for each activation with layerwise rate scaling. Moderate clipping improves the conversion cost-accuracy tradeoff on the MNIST MLP and reduces SynOps by 27% on VGG-11/MNIST at matched ANN-SNN accuracy gap criteria, whereas the trend reverses on VGG-11/CIFAR-10. Mean-field and layerwise diagnostics indicate that finite-window sampling and terminal-layer mismatch are the main residual errors. Overall, our results establish finite-state CTMC neurons as a theoretically grounded framework for activation-flexible ANN-to-SNN conversion beyond fixed activation-neuron correspondences.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Network, ANN-to-SNN Conversion, Continuous-Time Markov chains, Finite-State Markov Neurons, Neuromorphic Computing

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_bV7aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, the authors propose CTMC based ANN-to-SNN conversion method as a generalised framework for different activation functions. Previous methods, which mainly focused on ReLU-based activation functions, in this work they extend it for other activation functions like sigmoid, softplus, and others, particularly for monotonic functions. They propose to use a 3-state CTMC neuron model to match the nonlinearity of the activation functions by matching the CTMC spike emission state with the corresponding ANN activation function. They derive a closed-form expression for the stationary firing rate and optimize the transition-rate parameters by minimizing the approximation error between the CTMC firing rate and the target activation function.

Critical strengths
  1. This paper highlights that the majority of the ANN-to-SNN conversion methods primarily focus on ReLU activation-based ANN models, ignoring other activation function-based ANN models.
  2. The proposed CTMC-neuron model is particularly interesting, and the motivation for using this model is clear.
Critical weaknesses
  1. It's not clear why the CTMC neuron model needed to be a 3-state model, why not another number of states? Particularly, why is this topology preferred?
  2. Why are only a few parameters optimised in the expression as shown in equation 3, why not others (a,b are optimised c, d are not)?
  3. How are the values of c and d set? What kind of optimizer is used to optimise the expression, how many iterations, step size, and other details?
  4. I understand that the majority of ANN-to-SNN conversion methods primarily focus on ReLU activation-based neural networks. However, the experimental evaluation lacks comparisons with recent ANN-to-SNN conversion methods. I suggest the authors compare their method with recent methods such as [1], [2], as well as the relevant baselines reported in these works. [3] is also relevant because it uses probabilistic spiking neurons.
  5. The proposed CTMC neuron model may approximate the spike rate of different non-linear activation functions, however, this method may introduce additional computational overheads, which might reduce the benefits of energy efficiency of SNN. Particularly, during inference, it needs to compute stochastic sampling (for transitions), transition rate calculations. It's not clear how many such operations are needed and what the computational overhead for these operations is.
  6. The presentation of the paper needs improvement, particularly the organization and placement of the figures. The figures take up the majority of the space in the main paper. In Figure 1, subplot A, is it necessary? In Figure B, the subplots could be arranged in a single row instead of 2 rows. Similarly, Figures 3 and 4 could be better organized and aligned to improve readability and reduce the space occupied by the figures.

[1] Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications [2] Differential Coding for Training-Free ANN-to-SNN Conversion [3] Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons

Questions for authors

Please check the weaknesses section.

4 comments

We thank the reviewer for the detailed comments and suggestions.

1. Choice of state count and neuron topology (Q1)

It's not clear why the CTMC neuron model needed to be a 3-state model, why not another number of states? Particularly, why is this topology preferred?

Response: Thank you for raising the question. The three-state topology is not intended to be universally necessary or optimal. Rather, we select compact neuron structures based on the target activation. The two-state model provides a linear positive-input tail suitable for ReLU and softplus, while introducing a third refractory state allows the recovery rate to control the firing-rate ceiling, enabling saturation for sigmoid-like activations.

To further examine the role of state count, we added an ablation with fully connected CTMC neurons ranging from 2 to 20 states (Appendix F, Table 4). Although increasing the number of states substantially improves stationary-rate fitting, it also increases estimated storage and transition counts without improving SNN classification accuracy in our softplus MLP experiment. These results support our choice of compact, activation-dependent topologies as a practical balance between expressivity and complexity.

2. Fitted parameters and optimization details (Q2 & Q3)

Why are only a few parameters optimised in the expression as shown in equation 3, why not others (a,b are optimised c, d are not)?

How are the values of c and d set? What kind of optimizer is used to optimise the expression, how many iterations, step size, and other details?

Response: Thank you for the questions. In our three-state model, all four transition rates are parameterized and optimized, rather than only aa and bb. Specifically, a(H)a(H) and c(H)c(H) are input-dependent, while bb and dd are input-independent but remain trainable. All six parameters (a0,b0,c0,d0,wa,wc)(a_0,b_0,c_0,d_0,w_a,w_c) are jointly optimized for each layer. The prescribed lower bound dmin⁡d_{\min} ensures a positive recovery rate. We have clarified these distinctions in Section 3.2 and Appendix E.1.

We use Adam with an initial learning rate of 10−210^{-2} and cosine annealing, optimizing for 8,000 steps per initialization for MLPs and 16,000 steps for VGG-11. Three-state fitting uses multiple prescribed initializations, retaining the parameters with the lowest fitting loss. Appendix E further specifies the fitting domains, sampling grids, initialization choices, and optimization settings.

3. Comparison with recent ANN-to-SNN conversion methods (Q4)

I understand that the majority of ANN-to-SNN conversion methods primarily focus on ReLU activation-based neural networks. However, the experimental evaluation lacks comparisons with recent ANN-to-SNN conversion methods. I suggest the authors compare their method with recent methods such as [1], [2], as well as the relevant baselines reported in these works. [3] is also relevant because it uses probabilistic spiking neurons.

[1] Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications

[2] Differential Coding for Training-Free ANN-to-SNN Conversion

[3] Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons

Response: We appreciate the reviewer's suggestion. We have expanded Appendix C to include comparisons with conventional IF, Inference-Scale (Bu et al., 2025), and DCGS (Huang et al., 2025), including both plain and delayed evaluation for Inference-Scale. We also discuss the related two-phase probabilistic spiking approach (Bojković et al., 2025) in Section 2.

The results confirm that recent methods achieve substantially lower SynOps on ReLU networks. For VGG-11/CIFAR-10, DCGS achieves 0.209G SynOps, while Inference-Scale with delayed evaluation requires 0.486G under the reported accuracy-gap criterion (Table 2).

Our primary contribution lies in activation flexibility at the spiking-neuron level. While Inference-Scale primarily targets ReLU conversion, DCGS extends to non-ReLU operations through differential graded units that explicitly evaluate the original nonlinear functions. In contrast, CTMC neurons approximate diverse nonnegative activations directly through their stationary spike flux, without requiring separate activation-specific computation modules.

Our uniform approximation theorem establishes the theoretical foundation for this approach, while the successful conversion of sigmoid, softplus, and ReLU networks demonstrates its practical feasibility using compact two- and three-state neurons.

4. Computational overhead during inference (Q5)

The proposed CTMC neuron model may approximate the spike rate of different non-linear activation functions, however, this method may introduce additional computational overheads, which might reduce the benefits of energy efficiency of SNN. Particularly, during inference, it needs to compute stochastic sampling (for transitions), transition rate calculations. It's not clear how many such operations are needed and what the computational overhead for these operations is.

Response: Thank you for raising this question. To clarify the computational requirements, we have examined the per-step operations of our CTMC implementation (Appendix D). For each neuron, the two-state model updates its filtered input, evaluates two input-dependent transition rates, computes two exponential transition probabilities and a branching probability, and performs three uniform random draws in the current implementation.

These operations are implemented through parallel tensor computations on GPU. The sampling procedure is implementation-dependent and can be further simplified, for example by using a single state-dependent random draw to sample from the corresponding transition outcomes. The reported operation counts therefore characterize our current simulator rather than the minimum computational requirements of the neuron model.

A comprehensive assessment of hardware energy efficiency would require implementation-specific measurements beyond the scope of the present study.

5. Figure organization and presentation (Q6)

The presentation of the paper needs improvement, particularly the organization and placement of the figures. The figures take up the majority of the space in the main paper. In Figure 1, subplot A, is it necessary? In Figure B, the subplots could be arranged in a single row instead of 2 rows. Similarly, Figures 3 and 4 could be better organized and aligned to improve readability and reduce the space occupied by the figures.

Response: We thank the reviewer for the suggestions on figure organization and readability. We have revised the layouts of Figures 2 and 3, arranging their panels more compactly and consistently to improve alignment, readability, and space efficiency. We have also revised Figure 4A to present absolute SNN accuracy, facilitating a more direct interpretation of the accuracy–cost trade-off.

Regarding Figure 1, we have chosen to retain its original structure, including Panel A. As our work connects ANN-to-SNN conversion with finite-state stochastic neuron dynamics, we believe this introductory illustration provides useful context, particularly for readers less familiar with the field. We retain the arrangement in Panel B to maintain a clear association between the IF and LIF membrane trajectories and their corresponding firing-rate curves. Together, these panels establish the conceptual connection between conventional activation functions and spiking-neuron responses before introducing our CTMC construction.

I would like to thank the authors for their response. I liked the CTMC method as a generalised framework for ANN-SNN conversion, but I believe the manuscript still has some limitations. After reading their discussion with other reviewers and considering the current version of the paper, I will stand by my score. I request that the authors improve the presentation of the manuscript and include the modifications suggested by the reviewers to make it complete.

I believe this work will also benefit from showing how feasible it is to deploy on actual hardware; for example, it needs sampling for transitions between states and others.

Thank you for your follow-up and for recognizing the value of our generalized CTMC framework for ANN-to-SNN conversion. We would like to emphasize that we have responded to each of the concerns raised in your original review and respectfully request that you reconsider your assessment.

1. Scope and primary contribution

The objective of this work is not to outperform existing conversion methods in computational efficiency, but to develop a generalized ANN-to-SNN conversion framework that accommodates different activation functions. There is no universally optimal activation function. Our framework enables the conversion of ANNs while preserving their original activation behavior, supported by a theoretical approximation guarantee and experimental validation. We also recognize the importance of hardware deployment, although hardware-level efficiency has never been a claimed contribution of this work.

From the outset, we have clearly defined the scope of this work and acknowledged its efficiency limitations. Superiority in computational efficiency is neither the objective nor a claimed contribution of our framework. We therefore respectfully ask that these acknowledged limitations not serve as the primary basis for maintaining the original rating.

2. Additional experiments and manuscript revisions

In direct response to your original concerns, we have made the following revisions:

  • Q1: Additional state-count ablation and justification of neuron topology (Appendix F).
  • Q2 & Q3: Detailed parameterization, optimization, and calibration procedures (Section 3.2, Appendices E and G).
  • Q4: Additional experimental comparisons with Inference-Scale and DCGS, and discussion of the recommended probabilistic spiking method (Appendix C, Section 2).
  • Q5: Quantification of computational overhead, stochastic sampling, storage, and state-transition costs (Appendices D and F).
  • Q6: Revised figure layouts and improved manuscript presentation.

In response to your latest comment, we have further revised the manuscript's layout and figure organization to improve readability. We remain open to further improvements. However, given our point-by-point responses, additional experiments, and substantial revisions, we would appreciate clarification on which substantive concerns remain insufficiently addressed.

We respectfully ask that the scientific contributions of our framework and the revisions addressing your original concerns be fully taken into account in your overall assessment.

Thank you again for your time and consideration.

Reviewer_iBm6OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes converting a trained artificial neural network into a spiking neural network by replacing each neuron with a small Markov chain. The chain moves randomly between a few states, some transitions produce spikes, and the long-run spike rate is fitted to match the original activation function (e.g. sigmoid or ReLU). The authors prove that, with enough states, such chains can approximate a broad class of activation functions to arbitrary accuracy. Experiments on MNIST and CIFAR-10 compare activation choices and measure the trade-off between accuracy loss and the number of spikes.

Critical strengths
  1. The main idea, using the long-run firing rate of a Markov chain as an activation function, is intuitive and well illustrated (Figure 1), even for a reader outside the area.
  2. The approximation theorem (Appendix B) is, as far as I can check, correct and cleanly argued using Bernstein polynomials.
Critical weaknesses
  1. The practical benefit is unclear to me. I am not familiar with the SNN conversion literature, but the standard baseline reported in the paper's own Appendix C appears to need fewer spikes at similar or better accuracy than the proposed method (e.g. on VGG-11/MNIST, 0.38 G SynOps for the baseline versus 2.64 G in Table 1). The text describes this as "comparable" (L. 430–431), which I found hard to reconcile with the numbers. I would appreciate a clearer explanation of when and why the proposed neurons are preferable.

  2. The theory and the experiments seem loosely connected. The theorem relies on large chains, while the experiments use only 2–3 states, so it is not clear how much the theoretical result supports the practical method.

Questions for authors

Clarification requests:

  1. When is the proposed method preferable to standard conversion? The integrate-and-fire baseline in Appendix C appears to use fewer spikes at similar or better accuracy than the CTMC networks in Table 1. If the authors can show a setting where the proposed neurons are cheaper or more accurate, or explain why the comparison is not like-for-like, I would view the practical contribution more favorably.

  2. Absolute accuracies for the clipping experiments. The paper notes that the gap measures conversion fidelity rather than absolute accuracy. Could you report the absolute accuracies, so readers can see whether clipping saves spikes at equal final accuracy?

  3. Experimental details. I could not find the definition of the parameter rr in (T,τ,r)(T,\tau,r), nor how the layer scales αℓ\alpha_\ell are chosen. Clarifying these would help me judge whether the results are reproducible.

  4. Assumptions of Theorem 1. The theorem assumes the activation is nondecreasing, but I could not see where this is used in the proof. Is the assumption needed? If not, the result is more general than stated.

6 comments

We thank the reviewer for the constructive questions and suggestions.

1. Practical benefits of CTMC conversion (W1 & Q1)

The practical benefit is unclear to me. I am not familiar with the SNN conversion literature, but the standard baseline reported in the paper's own Appendix C appears to need fewer spikes at similar or better accuracy than the proposed method (e.g. on VGG-11/MNIST, 0.38 G SynOps for the baseline versus 2.64 G in Table 1). The text describes this as "comparable" (L. 430–431), which I found hard to reconcile with the numbers. I would appreciate a clearer explanation of when and why the proposed neurons are preferable.

When is the proposed method preferable to standard conversion? The integrate-and-fire baseline in Appendix C appears to use fewer spikes at similar or better accuracy than the CTMC networks in Table 1. If the authors can show a setting where the proposed neurons are cheaper or more accurate, or explain why the comparison is not like-for-like, I would view the practical contribution more favorably.

Response: Thank you for raising the question. We agree that conventional IF conversion achieves the reported accuracy-gap criterion with fewer SynOps on VGG-11/ReLU, and we have revised the description of this comparison.

The primary advantage of CTMC neurons lies in their flexibility in converting ANNs with different activation functions within a unified framework. Our theoretical construction establishes that finite-state CTMC neurons can approximate a broad class of activation functions through their stationary spike flux. Moreover, this flexibility is practically realizable with compact two- and three-state neuron models. Using the same fitting and layerwise scaling principles, we successfully convert MNIST MLPs with sigmoid, softplus, and ReLU activations to within one percentage point of their respective source ANNs (Section 5.1, Figure 3).

This provides a common conversion framework that preserves the activation choices of the source ANN without requiring their replacement by ReLU-like functions, while retaining spike-based interlayer communication. CTMC conversion is therefore particularly relevant when the objective is to preserve non-ReLU source activations through spiking-neuron dynamics.

2. Absolute accuracy under clipping (Q2)

Absolute accuracy for the clipping experiments. The paper notes that the gap measures conversion fidelity rather than absolute accuracy. Could you report the absolute accuracy, so readers can see whether clipping saves spikes at equal final accuracy?

Response: Thank you for the suggestion. We have revised Figure 4A to report absolute SNN accuracy rather than the conversion gap, allowing a direct comparison of spike efficiency at comparable final accuracy.

ClipReLU6 reaches approximately 97.5% accuracy at 9.8k SynOps per sample. Unclipped ReLU requires roughly 13.2k SynOps at a similar accuracy and reaches higher accuracy at larger budgets. The updated results demonstrate that the SynOps savings from moderate clipping can persist when evaluated at comparable absolute SNN accuracy.

3. Experimental details and rate scaling (Q3)

Experimental details. I could not find the definition of the parameter r in () , nor how the layer scales are chosen. Clarifying these would help me judge whether the results are reproducible.

Response: Thank you for pointing out these missing details. We have clarified that rr denotes the base target firing rate used to control layerwise scaling (Section 3.3). We now explicitly describe how ANN activations are mapped to target firing rates and how the corresponding weights are rescaled.

Appendix E specifies the scaling choices for MLPs, while Appendix G provides the VGG-11/CIFAR-10 calibration procedure, including the layerwise rate multipliers, preactivation-based scales, and offline weight adjustments.

4. Monotonicity assumption in Theorem 1 (Q4)

Assumptions of Theorem 1. The theorem assumes the activation is nondecreasing, but I could not see where this is used in the proof. Is the assumption needed? If not, the result is more general than stated.

Response: We thank the reviewer for this careful observation. The monotonicity assumption is indeed not required by our proof. We initially stated the theorem for nondecreasing activations to maintain consistency with the activation functions examined in our experiments, rather than because of a theoretical restriction.

We appreciate the reviewer for highlighting that this presentation understated the generality of our result. We have accordingly revised Theorem 1 and the related descriptions to explicitly state that the uniform approximation guarantee applies to any continuous, nonnegative activation function on a compact interval.

5. Connection between the theoretical and experimental models (W2)

The theory and the experiments seem loosely connected. The theorem relies on large chains, while the experiments use only 2–3 states, so it is not clear how much the theoretical result supports the practical method.

Response: The theoretical result establishes the expressive capacity of the general finite-state CTMC family through its stationary spike flux. Our practical two- and three-state neurons are compact realizations of this design principle, with transition structures selected according to the shape of the target activation. The theorem establishes expressivity for the general family rather than guaranteeing arbitrary approximation accuracy for a fixed low-state topology.

To further investigate the relationship between expressivity and practical conversion, we added a state-count ablation using fully connected CTMC neurons with 2, 3, 6, 10, and 20 states on the same softplus MLP, while keeping the layerwise rate scales fixed (Appendix F, Table 4).

Increasing the state count substantially improves stationary-rate fitting but increases estimated storage and internal transition counts, without corresponding improvements in classification accuracy. These results illustrate that large state spaces are not necessarily required for effective conversion and support the use of compact neuron models for the activation functions considered in our experiments.

I want to thank the authors for answering my questions, and clarifying the concerns I raised in my review. In particular, I am satisfied with the replies to my points 2, 3 and 4. It is also more transparent to me now that one of the main contributions of this work is about the flexibility regarding activation functions rather than efficiency. I consider this as an important problem to be addressed, but I still see limited practical benefit: in their experiments ReLU is still the cheapest option. Due to this reason I maintain my score.

Reviewer_L4TZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies activation-flexible ANN-to-SNN conversion based on finite-state continuous-time Markov neurons. It fits stationary firing curves and applies layerwise scaling to bridge ANN activations and SNN spike rates. The authors provide a general construction showing that, on compact intervals, the model can uniformly approximate continuous, nonnegative, nondecreasing activation functions. They then evaluate low-state models through activation fitting and conversion experiments on MNIST and CIFAR-10. In some MNIST settings, moderate clipping reduces event cost while keeping the conversion gap within a target range, but CIFAR-10 shows the opposite trend, suggesting that the benefit of clipping is dataset- and setting-dependent.

Critical strengths
  1. The stationary-flux formulation and the constructive approximation theorem give the paper a solid theoretical basis for activation flexibility. The authors also make a clear distinction between the general expressivity result and the more restricted low-state models used in experiments.
  2. The clipping experiments provide helpful evidence about the limits of bounded activations. The different outcomes on MNIST and CIFAR-10 support conditional conclusions about conversion fidelity and event cost.
Critical weaknesses
  1. The connection between the low-state experimental models and the claimed affine transition rates is unclear. The relevant section does not fully specify the transition-rate functions used in the experiments, so it is hard to tell whether the fitted curves come from affine transitions, piecewise transformations, or other nonlinear mechanisms.
  2. The interpretation of error sources is not fully supported. The mean-field expression is presented as an operational definition, but it does not clearly explain why the results depend on simulation length in the reported figures. The claim that sampling error dominates also appears broader than what the current ReLU configurations support.
Questions for authors
  1. What are the exact transition-rate functions used in the activation fitting? Are they affine across the entire fitting domain, and do they include clipping or piecewise transformations?
  2. Is the mean-field model best understood as a static stationary-rate approximation, or as a time-dependent deterministic model? Please clarify why its predictions depend on simulation length, and indicate which configurations support the claim that sampling error dominates.
5 comments

We thank the reviewer for pointing out where the transition-rate parameterization and the time dependence of mean-field inference needed clarification.

1. Transition-rate functions and their relation to the theorem (W1 & Q1)

The connection between the low-state experimental models and the claimed affine transition rates is unclear. The relevant section does not fully specify the transition-rate functions used in the experiments, so it is hard to tell whether the fitted curves come from affine transitions, piecewise transformations, or other nonlinear mechanisms.

What are the exact transition-rate functions used in the activation fitting? Are they affine across the entire fitting domain, and do they include clipping or piecewise transformations?

Response: Section 3.2 now explicitly specifies the transition-rate functions used in our experiments. The input-dependent rates are parameterized as

a(H)=softplus⁡(a0+waH),c(H)=softplus⁡(c0+wcH).a(H)=\operatorname{softplus}(a_0+w_aH), \qquad c(H)=\operatorname{softplus}(c_0+w_cH).

The input-independent rates are

b=softplus⁡(b0),d=softplus⁡(d0)+dmin⁡.b=\operatorname{softplus}(b_0), \qquad d=\operatorname{softplus}(d_0)+d_{\min}.

where dd is specific to the three-state model.

While the theoretical construction uses affine input-dependent rates, the practical models employ nonlinear softplus links to ensure positive transition rates and facilitate gradient-based optimization. For clipped ReLU, the input is clamped before evaluating the rate functions, producing the desired saturation without fitting separate parameter sets across different input regions (Appendix E). Both constructions follow the same stationary-flux approximation principle.

2. Dynamic mean-field inference and interpretation of error sources (W2 & Q2)

The interpretation of error sources is not fully supported. The mean-field expression is presented as an operational definition, but it does not clearly explain why the results depend on simulation length in the reported figures. The claim that sampling error dominates also appears broader than what the current ReLU configurations support.

Is the mean-field model best understood as a static stationary-rate approximation, or as a time-dependent deterministic model? Please clarify why its predictions depend on simulation length, and indicate which configurations support the claim that sampling error dominates.

Response: Thank you for raising the question. Our mean-field evaluation for VGG-11 employs a time-dependent deterministic approximation of the stochastic CTMC dynamics, rather than a static stationary-rate substitution. Section 3.4 now explicitly describes the two-state MF recurrence.

Specifically, the mean-field model evolves the state occupancy probabilities from the same initial conditions as the stochastic SNN. At each time step, the expected spike output is computed and propagated through the network using the same temporal filtering and finite-window readout as the stochastic SNN. Consequently, mean-field predictions depend on the inference duration.

We have also refined the terminology and narrowed the original claim regarding sampling error. The mean-field analysis distinguishes discrepancies already present in deterministic finite-time conversion from those arising when stochastic spike generation is introduced.

On VGG-11/MNIST, most of the accuracy loss at T=30T=30 occurs between MF and SNN for both ReLU and ClipReLU6. At the same τ=0.02\tau=0.02 and r=3r=3, extending the inference window to T=50T=50 reduces this loss for both activations (Figure 5B). These results demonstrate the substantial contribution of stochastic spike realization in these configurations. We interpret the two gaps as an operational accuracy decomposition rather than a formal statistical variance decomposition.

Thank you for the detailed response and revisions. Your clarifications in Section 3.2 regarding the softplus link functions and input clamping, as well as the explicit time-dependent mean-field recurrence and narrowed error claims in Section 3.4, address my previous technical questions well. However, I notice that the revised main text now extends onto page 10 (including Sections 6 and 7), exceeding the official 9-page limit for the main paper. Given this page-limit violation alongside the remaining reservations regarding the limited empirical scope and dataset-dependent gains, I will maintain my rating of 2.

Thank you for reviewing our revised manuscript and for acknowledging that our revisions address your previous technical questions. We would like to clarify a possible misunderstanding regarding the page limit. According to the official ICLR 2027 Author Guidelines (https://iclr.cc/Conferences/2027/AuthorGuidelines), the main-text page limit is extended from 9 to 10 pages during the discussion/rebuttal phase and for the camera-ready version. Our revised manuscript therefore complies with the official page limit. While we respect your remaining concerns regarding the empirical scope and dataset-dependent gains, we kindly ask that you reconsider your assessment in light of this clarification. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.