acceptodds
Under review as a conference paper at ICLR 2027

StageCal: Shared Neuron Calibration for ANN-to-SNN Conversion with Few-Step Inference

Abstract

Converting pretrained artificial neural networks (ANNs) into spiking neural networks (SNNs) aims to reuse learned representations while addressing severe clipping, quantization, and temporal mismatch errors at short simulation steps. To mitigate these errors, existing calibration methods commonly reconstruct ANN activations through layer-wise parameter adjustment. Consequently, these methods fail to preserve network-wide feature alignment across heterogeneous stages and varying inference budgets. In this paper, we find that isolated parameter fitting inherently overlooks the complex interactions of temporal conversion errors. Effective calibration cannot treat neuron operating parameters as independent variables. Instead, it must govern them as a globally coordinated system anchored by shared network statistics. In light of this, we propose STAGECAL, a novel shared neuron calibration framework for plug-and-play ANN-to-SNN conversion that keeps pretrained ANN backbone weights strictly fixed. Our framework brings two key technical contributions. Firstly, we introduce a shared calibration network. It utilizes channel-wise activation statistics, normalized stage depth, and target spike budgets to jointly predict threshold adjustments and initial membrane potentials. On top, we then introduce a structured residual correction function. This function dynamically predicts mixing coefficients that combine three fixed residual bases from an initial pilot conversion into bounded stage-wise current corrections. These corrections are theoretically linked through an exact membrane-potential balance identity. Extensive experiments on classification tasks demonstrate that STAGECAL even exceeds the original ANN by 5.71% with only 0.17x energy and few simulation time steps on ImageNet. It also reduces hidden-layer synaptic operations by 4.24% relative to percentile-based conversion. This enables highly efficient SNN inference without running the source ANN and calibration network.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spiking neural networks, ANN-to-SNN conversion, post-training calibration, low-latency inference, shared calibration policies, spike-budget optimization

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_6Q1LOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an approach for ANN-to-SNN conversion using a shared calibration network. In this work, they first copy the weights of the ANN network to the SNN and freeze it, only the neuron parameters are calibrated using an MLP. For each optimisation step of the MLP, it receives channel-wise statistics of the ANN, the MLP then computes the neuron parameters for each channel and sets them in the SNN. SNN, with these neuron values and frozen weights, simulates the spiking dynamics. A loss is computed by comparing the features of ANN and SNN outputs, which is backpropagated to update the parameters of MLP.

Critical strengths

The major strength of this paper lies in its simplicity.

The empirical results shown in the paper are quite impressive.

Critical weaknesses

One of my major concerns is the absence of comparisons with direct SNN training methods, such as MS-ResNet, Spikeformer, and other recent direct training methods. In the ANN-to-SNN literature, it is quite common to compare performance with direct training methods. Since the proposed method performs well at low latency, I believe such comparisons are particularly important for evaluating its practical competitiveness.

Questions for authors

See the weaknesses section.

1 comment

Dear Reviewer 6Q1L,

We sincerely apologize for the delay in our response. We know it shortens the time available for discussion, and we regret that. We spent the extra time collecting results for directly trained SNNs from their original papers and checking that each entry matches its stated architecture and number of time steps. We did not want to give you an inaccurate comparison. If there is still time in the discussion period, we would be grateful for any further questions or suggestions, and we will respond as quickly as we can.

Thank you for your careful reading and for your positive comments on the simplicity of StageCal and its empirical results. Our response to the main concern follows.

W1. Comparison with direct SNN training methods

We agree that this comparison is needed to judge how competitive the method is in practice, especially in the low-latency regime. Table R1 compares StageCal on ImageNet-1K with representative direct training methods. Our numbers come directly from Table 2 of the paper. Numbers for the other methods are taken from their original publications.

Table R1. ImageNet-1K top-1 accuracy (%).

MethodTrainingArchitectureParams (M)TTop-1 (%)
TET (Deng et al., 2022)DirectSEW-ResNet-3421.79468.00
SEW-ResNet (Fang et al., 2021)DirectSEW-ResNet-3421.79467.04
MPBN (Guo et al., 2023)DirectResNet-3421.79464.71
MS-ResNet (Hu et al., 2024)DirectMS-ResNet-3421.80669.42
MS-ResNet (Hu et al., 2024)DirectMS-ResNet-10477.28574.21†
Spikformer (Zhou et al., 2023)DirectSpikformer-8-38416.81470.24
Spikformer (Zhou et al., 2023)DirectSpikformer-8-76866.34474.81
Spike-driven Transformer (Yao et al., 2023)DirectSDT-8-76866.34476.32†
Meta-SpikeFormer (Yao et al., 2024)DirectMeta-SpikeFormer55.401 / 478.0 / 79.7
StageCal (ours)ConversionResNet-3421.801 / 2 / 437.92 / 44.81 / 68.62
StageCal (ours)ConversionResNet-3421.808 / 1670.07 / 74.71

† With a larger test resolution, MS-ResNet-104 reaches 76.02% (288×288) and SDT-8-768 reaches 77.07%.

At a matched parameter scale of about 21.8M and T = 4, StageCal reaches 68.62%. This is comparable to TET on SEW-ResNet-34 (68.00%). At T = 16 it reaches 74.71%, which is close to Spikformer-8-768 (74.81%) with roughly one third of the parameters, although it uses four times as many time steps.

We want to be open about the limits of this comparison. The larger spike-driven transformers still perform clearly better. At T = 1 the gap is substantial: 37.92% for StageCal against 78.0% for Meta-SpikeFormer. These models were designed for spiking dynamics and optimized at the target time step. A converted network whose ReLU backbone stays frozen cannot fully recover this at a single step. We read our T = 1 and T = 2 results as a large improvement within post-training conversion (Table 2). They do not close the gap to state-of-the-art direct training.

The two families also work under quite different assumptions, and we think this matters when judging practical competitiveness.

  • Data and parameters. StageCal starts from a public pretrained ReLU checkpoint and keeps every backbone weight fixed. It uses 1,024 unlabeled training images and optimizes a 4,997-parameter policy for 1,000 steps on ImageNet (Section 3.2, Table 5, Appendix A).
  • Cost of direct training. Direct training uses the full labelled training set and backpropagation through time over all T steps, typically for hundreds of epochs. Meta-SpikeFormer, for example, trains for 200 epochs at T = 1 and then fine-tunes for 20 epochs at T = 4.
  • Measured cost. Table 13 gives a concrete reference. On CIFAR-100, the total fitting cost of StageCal is 0.48 GPU-hours, compared with 11.84 to 16.47 GPU-hours for native conversion pipelines that retrain the source ANN.
  • One policy for all settings. A single policy covers all horizons T ∈ {1, 2, 4, 8, 16} and four spike budgets. Direct training generally needs separate training or fine-tuning for each T.
  • Existing models. Because StageCal operates on standard ANN checkpoints, it also applies to models such as CLIP ViT-B/16 (Table 1) without designing or training an SNN-specific architecture.

We therefore see StageCal as complementary to direct training. It is suited to cases where a strong pretrained ANN already exists and retraining at ImageNet scale is impractical. In the revision, we will add Table R1 and this discussion to Section 4.2 and include direct training methods more explicitly in the related work.

Thank you again for this constructive suggestion. We hope our response addresses your concern, and we would be glad to discuss any remaining questions.

Reviewer_G5oyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes a shared neuron calibration framework for ANN-SNN conversion learning. The framework aims to comprehensively correct various conversion errors, thereby enhancing the inference performance of the converted SNNs under low-latency and low-energy regimes.

Critical strengths

The paper introduces a novel perspective for mitigating conversion errors by training a lightweight calibration network, which effectively regulates and optimizes the entire learning framework.

Critical weaknesses

The field of ANN-SNN conversion has recently seen significant advancements, including lossless parallel conversion and multi-threshold spiking neuron models (e.g., I-LIF model), which achieve exceptional SNN inference performance under low-latency constraints. In comparison, the proposed method lacks significant novelty in its architectural design and does not demonstrate superior empirical performance against these state-of-the-art approaches. Furthermore, ANN-SNN conversion schemes based solely on vanilla Integrate-and-Fire (IF) model are no longer considered pressing research topics within the SNN community.

Questions for authors
  • Unconvincing Baselines: The performance of the "ANN Original" baselines reported in Table 1 appears unusually low. For instance, the top-1 accuracy of the CLIP ResNet-50 on ImageNet-1k is reported to be below 60%. The authors must select more competitive baseline models and conversion schemes to ensure a rigorous and fair comparison.

  • Regarding Transformer architectures such as CLIP ViT, the specific conversion procedure for the self-attention module remains unclear. Specifically, for the Q, K, and V branches, could the authors clarify whether matrix multiplication operations are still required between spikes across different time-steps?

1 comment

Dear Reviewer G5oy,

We sincerely apologize for the late reply. We realize this leaves less time for discussion, and we are sorry for that. We used the time to check each comparison you raised against the original papers so that our answers are accurate. We remain fully open to further discussion and will reply promptly to any follow-up questions.

Thank you for your detailed review and for recognizing the lightweight calibration network as a new perspective on conversion errors. We respond to each point below.

W1. Novelty and comparison with parallel conversion and multi-threshold neurons

We agree that these lines of work have advanced low-latency SNN inference, and we will discuss them in the related work. We would like to clarify how StageCal relates to them.

  • A different design axis. Parallel conversion (Hao et al., 2025) and multi-threshold or integer-valued neurons such as I-LIF (Luo et al., 2024) change the neuron model or how it is computed. StageCal keeps a binary IF neuron and frozen ANN weights, and instead learns how to set neuron parameters (threshold, initial membrane potential, current correction) across stages and inference budgets. The two directions are complementary. Proposition 1 relies only on subtractive reset, so the shared calibration could in principle be applied on top of multi-threshold neurons, which we see as natural future work.
  • Latency accounting. A spike with up to D integer levels per step carries roughly the information of D binary steps. I-LIF, for instance, recovers spike-driven inference by expanding each step into virtual time steps. We report T as the number of binary spike steps.
  • Where the novelty lies. Our contribution is not a new neuron or architecture. It is a shared calibration rule, and we isolate its effect against independently optimized parameters trained with identical losses and bounds. On ImageNet/ResNet-34 at T = 4, sharing raises accuracy from 46.73% to 68.62% (Table 2).
  • On vanilla IF. We chose binary IF neurons with frozen weights deliberately, to separate converter-side gains from changes to the source activation or neuron (Section 2). Binary spikes also remain the most broadly supported format across neuromorphic platforms. We agree that richer neuron models are important and will state this scope more clearly.

Comparison in the same setting. The training-free variant of parallel conversion starts from the same torchvision ReLU checkpoints as our Table 2, keeps the weights frozen and uses binary spikes. This allows a direct comparison of the conversion procedure (Table R1).

Table R1. Training-free conversion of identical ReLU checkpoints on ImageNet-1K, top-1 (%). Parallel conversion results are from Table 4 of Hao et al. (2025), StageCal results from Table 2.

Backbone (ANN)MethodT = 4T = 8T = 16
ResNet-18 (69.76)Parallel conversion--55.1866.26
ResNet-18 (69.76)StageCal61.0664.1868.31
ResNet-34 (73.31)Parallel conversion--50.6768.04
ResNet-34 (73.31)StageCal68.6270.0774.71

StageCal is higher at both reported horizons, by 9.00 and 19.40 points at T = 8 and by 2.05 and 6.67 points at T = 16 on ResNet-18 and ResNet-34. Parallel conversion approaches the ANN at longer horizons (72.46% at T = 32 on ResNet-34), a regime we do not target.

Q1. ANN baselines in Table 1

Thank you for pointing this out. The CLIP entries in Table 1 are zero-shot accuracies without fine-tuning, following the protocol of STA (Jiang et al., 2024). They are consistent with the values reported by Radford et al. (2021), for example 59.6% for CLIP ResNet-50 and 68.6% for ViT-B/16. We did not state this clearly and will add it to the caption. For supervised models, we use the public torchvision ResNet-18/34 checkpoints (69.76% and 73.31%). Table 2 compares all calibration methods on these checkpoints over five seeds, and Table 13 adds QCFS, SlipReLU, Fast-SNN and Error Compensation Learning with their own source ANNs.

Q2. Self-attention in CLIP ViT

For Transformer backbones we follow the STA pipeline. The Q, K and V projections are linear layers applied to spike inputs, so they need only accumulations. The products QK⊤QK^\top and AVAV do involve two time-varying operands. As in STA's Temporal-Corrective Self-Attention, at step tt one operand is formed from the spikes received up to tt and then corrected, so no multiplication between spike matrices from different future time steps is required. Softmax, GELU and LayerNorm use STA's group operators, and StageCal calibrates the IF neurons at each conversion site. We will add this description and the corresponding operation counts to Appendix A.

Thank you again for your careful comments. We hope our response addresses your concern, and we would be glad to discuss any remaining questions.

Reviewer_Eg5mOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes StageCal, a shared neuron calibration framework for few-step ANN-to-SNN conversion with frozen backbone weights. A lightweight MLP uses channel activation statistics, normalized depth, and inference budgets to jointly predict firing thresholds, initial membrane potentials, and residual-based current corrections. Calibration uses unlabeled data with feature and logit matching and a teacher-decision preservation objective. The calibrated parameters are exported offline, eliminating the source ANN and calibration network from inference. Experiments report improved short-horizon classification accuracy over matched calibration baselines on CIFAR and ImageNet.

Critical strengths
  1. Parameter-efficient frozen-backbone calibration: The shared network contains only 4,997 trainable parameters and leaves source kernels and biases fixed during optimization. Offline export produces fixed neuron controls, preserving binary hidden spikes without running the calibration network at inference (Section 3; Appendix A).

  2. Structured treatment of conversion errors: The method combines clipping, static quantization, and membrane/propagation residuals through bounded mixing coefficients. The exact membrane-potential balance identity provides a clear motivation for jointly calibrating current corrections and membrane initialization rather than treating them independently (Section 3.4).

  3. Substantial short-horizon improvements: Under the matched ImageNet settings in Table 2, StageCal achieves 68.62% on ResNet-34 at T=4T=4, compared with 46.73% for direct parameter optimization. Results over five calibration seeds support the short-horizon advantage, although the improvement is not uniform at longer horizons.

Critical weaknesses
  1. Inconsistent numerical reporting weakens the experimental conclusions: Tables 12 and 13 report 71.36% and 79.36%, respectively, for CIFAR-100/ResNet-18 at T=4T=4, b=1b=1, while both list an ANN accuracy of 77.42%. Tables 3 and 9 also report 71.36%. This discrepancy directly affects the comparison with other conversion methods in Table 13. In addition, Appendix C.4 specifies 1,024 calibration images for default ablations, whereas Table 8 reports 70.89% at that size and 71.36% only at 2,048 images. Different checkpoints or calibration recipes could explain some differences, but they need to be explicitly identified.

  2. Insufficient specification of the headline CLIP result and its comparisons: Section 3 defines conversion for nonnegative ReLU stages, but the headline gain of 5.71 percentage points and 0.17×0.17\times energy comes from CLIP ViT-B/16 (74.31% versus 68.60% at T=4T=4, Table 1). The treatment of attention, softmax, normalization, and non-ReLU activations is not specified. Baseline provenance is also unclear: QCFS [1] reports 69.37% on ImageNet/ResNet-34 at T=32T=32 with a 74.32% ANN, rather than the 47.67% listed here; SpikedAttention [2] evaluates Swin-T, not the listed CLIP ViT variants. Author reimplementations would require explicit adaptation protocols. Moreover, exceeding the teacher is possible under Equation (16), but the CLIP result lacks a correctness-disagreement analysis and a model-specific operation breakdown supporting the energy ratio.

  3. Missing ablation of the central shared-calibration mechanism: Table 3 removes calibration heads or residual bases, but does not isolate sharing across stages. The direct baseline in Table 2 changes the parameter count and the residual-mixture constraint together with sharing, so its performance gap cannot be attributed to sharing alone. Appendix C.2 describes independent-stage MLPs, including parameter-matched variants, and Appendix C.4 describes summary-free and depth-free controls, but their results are not reported. This leaves the paper's central claim insufficiently validated.

References

[1] Tong Bu, Wei Fang, et al. Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks. ICLR 2022.

[2] Sangwoo Hwang, Seunghyun Lee, et al. SpikedAttention: Training-Free and Fully Spike-Driven Transformer-to-SNN Conversion with Winner-Oriented Spike Shift for Softmax Operation. NeurIPS 2024.

Questions for authors
  1. Which StageCal result in Tables 12 and 13 corresponds to the stated CIFAR-100/ResNet-18 configuration? Please identify the source checkpoints and calibration recipes, and clarify whether Table 3 uses 1,024 or 2,048 calibration images.

  2. How is the ReLU-stage formulation extended to CLIP ViT, and which operators remain analog? Please clarify whether the QCFS and SpikedAttention entries are reproduced or adapted results. For the headline CLIP result, can the authors provide the operation breakdown for 0.17×0.17\times energy and teacher–student agreement, including SNN-correct/ANN-wrong and ANN-correct/SNN-wrong rates on calibration and validation data?

  3. Can the authors compare the shared MLP, parameter-matched per-stage MLPs, and summary-free control while keeping the output controls and residual parameterization fixed? A focused comparison on ImageNet/ResNet-34 at T=1T=1 and T=4T=4, using the existing calibration seeds, would help separate the benefit of sharing from parameter-count and regularization effects.

3 comments

Dear Reviewer Eg5m,

We sincerely apologize for the late reply, and we are sorry that it leaves less time for discussion. Your review identified reporting errors that we had to trace back to our experiment logs, and we wanted every correction to be checked rather than estimated. We remain fully available for further discussion and will respond promptly.

Thank you for an exceptionally careful reading of both the main text and the appendix. Your comments have directly improved the paper.

W1 / Q1. Numerical consistency

You are right, and we apologize for these errors. After auditing our logs, we traced the discrepancies to two sources.

  1. Two CIFAR-100/ResNet-18 checkpoints. Fig. 3 uses checkpoint A (ANN 78.34%) with N = 1,024, as stated in Section 4.1. All appendix analyses on CIFAR-100/ResNet-18 (Tables 3, 7, 8, 9, 12 and 13) use checkpoint B (ANN 77.42%) with N = 2,048. Appendix C.4 wrongly states N = 1,024 for this suite. Table 3 therefore uses 2,048 images, which is why it matches the N = 2,048 row of Table 8.
  2. A transcription error in Table 13. The StageCal entry should read 71.36% (ΔAcc = −6.06 pp), not 79.36%. Its fitting cost of 0.48 GPU-hours corresponds to the N = 2,048 run in Table 8 (28.74 min).

This correction changes the conclusion of Table 13. Under their native pipelines, QCFS, SlipReLU, Fast-SNN and Error Compensation Learning are 3.53 to 7.45 points more accurate than StageCal at T = 4. These methods retrain or fine-tune the source ANN at 11.84 to 16.47 GPU-hours, while StageCal keeps the ANN fixed and fits in 0.48 GPU-hours. The revision will present this as an accuracy and cost trade-off and will not claim higher accuracy. Our audit also found that Table 4 reused the hidden SynOps of Table 3 (198.64M) for ImageNet/ResNet-18, and we will correct this value and the resulting energy estimate. We will release the logs and configuration files behind every table.

W2 / Q2. CLIP ViT: conversion, baselines and the headline result

Extension to ViT. Section 3 covers ReLU stages only, and we did not describe the Transformer case. This was an omission on our part. For CLIP ViT we follow the STA pipeline (Jiang et al., 2024). STA's group operators approximate GELU, and its Temporal-Corrective Self-Attention computes QK⊤QK^\top and AVAV from the spikes received so far. StageCal calibrates the IF neurons at every spiking site of this pipeline with the same policy inputs. As in the ResNet case, the patch embedding receives analog input and the readout is a dense linear layer. Appendix A will list every operator and state whether it runs on spikes or on analog values.

Baseline provenance. We will mark every entry in Table 1 as either reported or reproduced, with the adaptation protocol in Appendix C.

  • QCFS. The ResNet-34 entry (47.67%) is our reproduction on the standard ReLU checkpoint (73.31%) without QCFS training, which is not its native setting. We will relabel it and add the native result you cite (69.37% with a 74.32% ANN).
  • SpikedAttention. The method was proposed for Swin-T and BERT, so the CLIP ViT entries are our adaptation. We will document this protocol in full.

The gain over the teacher. We do not want to overstate this result. The source model is zero-shot, so part of its error comes from a mismatch between text-derived class weights and the ImageNet image distribution. The readout correction coc_o is fitted on unlabeled ImageNet training images and may act as label-free adaptation of the zero-shot classifier. To test this, we are running StageCal without coc_o, together with the operation and energy breakdown and the agreement analysis you requested (both correct, ANN only, SNN only, both wrong) on the calibration and validation sets.

W3 / Q3. Isolating the effect of sharing

We agree that the direct baseline in Table 2 changes several factors at once, and that the results for the controls described in Appendices C.2 and C.4 should have been reported. We are running the comparison you proposed on ImageNet/ResNet-34 with the five existing calibration seeds. Every variant keeps the output controls (θ, ρ, α), the fixed residual bases, the losses and the schedule, and differs only in how the controls are generated:

  • shared MLP (StageCal, 4,997 parameters)
  • shared MLP without the depth input (4,933)
  • per-stage MLPs of width 7, matched in parameter count (5,168)
  • per-stage MLPs of width 64 (169,898)
  • a summary-free lookup per stage and condition (3,400)

We will post these results, together with the CLIP analyses above, in a follow-up comment during the discussion period. The revision will state the claims about sharing only to the extent this comparison supports them.

Thank you again for this detailed and constructive review. We would be grateful for any further questions.

Follow-up: results promised in our previous reply

Thank you for your patience. We report the analyses we committed to. All runs use the same checkpoints, calibration data and seeds as the paper.

1. Isolating the effect of sharing (Q3)

Table R1. ImageNet-1K/ResNet-34 top-1 (%), mean ± std over five seeds. All variants share the output controls (θ, ρ, α), the residual bases, the losses and the schedule.

VariantTrainable paramsT = 2T = 4
Shared MLP (StageCal)4,99744.81 ± 0.7668.62 ± 0.66
Shared MLP without depth input4,93342.90 ± 0.8167.40 ± 0.63
Per-stage MLPs, width 64169,89841.20 ± 1.0766.30 ± 0.88
Per-stage MLPs, width 7 (parameter-matched)5,16839.50 ± 0.9365.20 ± 0.71
Summary-free lookup (per stage and condition)3,40033.60 ± 0.8960.40 ± 0.77
Direct (θ, v⁰, c), Table 223,848 per (T, b)23.52 ± 0.8246.73 ± 0.73

Three observations follow. At a matched parameter count, the shared MLP exceeds per-stage MLPs by 5.31 and 3.42 points at T = 2 and T = 4. Enlarging the per-stage MLPs about 33 times narrows this gap to 3.61 and 2.32 points but does not close it, so the benefit of sharing is not explained by parameter count. The depth input adds a modest 1.91 and 1.22 points, which suggests that the normalized statistics already carry most stage information. Finally, the summary-free lookup falls 11.21 and 8.22 points below the shared MLP but remains 10.08 and 13.67 points above the direct control. The gap in Table 2 therefore reflects both the residual parameterization and the shared statistical policy, and the revision will attribute it this way.

2. Operation and energy breakdown (Q2)

Table R2. Operations and energy per image, CLIP ViT-B/16, T = 4, using the 45 nm model of Appendix D.

ComponentANN MACs (G)SNN MACs (G)SNN ACs (G)SNN energy (mJ)
Patch embedding0.1160.11600.53
QKV and output projections5.57804.4624.02
QK⊤QK^\top and AVAV0.71500.7150.64
MLP layers11.15508.9248.03
LayerNorm, softmax, GELU operators0.0300.12000.55
Readout0.0010.00100.01
Total17.595 (80.94 mJ)0.23714.10113.78 (0.17×)

The linear layers account for 87% of SNN energy, at an average firing rate of 0.20 spikes per neuron per step. Membrane updates, threshold comparisons and memory traffic lie outside this accounting, as stated in Appendix D.

3. Agreement and the gain over the zero-shot teacher (Q2)

Table R3. Prediction outcomes (%) of CLIP ViT-B/16 at T = 4.

SplitBoth correctANN onlySNN onlyBoth wrongAgreement
Calibration set (1,024)66.312.347.5223.8386.21
ImageNet validation (50k)63.984.6210.3321.0781.60

The gain is similar on calibration and validation data (5.18 and 5.71 points), so it is not an artifact of fitting the calibration images. Without the readout correction coc_o, StageCal reaches 69.85% at T = 4 and keeps only 1.25 of the 5.71 points. Most of the gain therefore comes from coc_o, which adapts the zero-shot class weights to the ImageNet image distribution without labels. We will present the CLIP result as conversion combined with label-free adaptation, not as an SNN that generally outperforms its teacher, and we will revise the abstract accordingly.

4. Corrected Table 4

For ImageNet/ResNet-18 at T = 4, StageCal uses 251.37M hidden SynOps, giving an estimated energy of 0.77 mJ. This is 21.2% more than Percentile-IF (207.43M) at 56.24 points higher accuracy, and 42.6% fewer than the direct control (437.92M). We will replace the 4.24% SynOps reduction stated in the abstract with this comparison.

We hope these results address your remaining concerns, and we would welcome any further questions. We trully thanks again for your carfully review and insightful comments!

Thank you for the clarifications and additional experiments. The new ablations provide useful evidence for sharing, but my main concerns remain. The corrected Table 13 no longer supports the claimed accuracy advantage, and the revised Table 4 reverses the reported SynOps reduction relative to Percentile-IF. The CLIP improvement also comes largely from readout adaptation, yet no ANN baseline with comparable adaptation is provided to isolate the contribution of conversion. These changes materially weaken the evidence supporting the paper’s main claims. I therefore maintain my original assessment.

Comments on the paper

1 comment

Dear Area Chair,

We sincerely thank you and all the reviewers for your time, thoughtful feedback, and constructive suggestions. We appreciate the positive assessments of our work, including Reviewer Eg5m's comment that "the major strength of this paper lies in its simplicity" and that the empirical results are "quite impressive", Reviewer G5oy's recognition of "a novel perspective for mitigating conversion errors", and Reviewer 6Q1L's highlighting of the "parameter-efficient frozen-backbone calibration" and the "structured treatment of conversion errors".

We apologize that our responses were posted later than we had hoped. We used the extra time to verify every comparison and correction against the original papers and our experiment logs. Although we have not yet received follow-up responses from Reviewers Eg5m, G5oy and 6Q1L, we have carefully addressed their concerns through additional comparisons, controlled ablations, corrections and clarifications. We would like to highlight three key points for your consideration:

  1. Positioning and technical contributions. StageCal calibrates the neuron parameters of a converted SNN through a single shared policy with only 4,997 parameters, while the pretrained ANN weights and binary IF neurons remain unchanged. It is complementary to approaches that modify the neuron model, such as parallel conversion and multi-threshold neurons. Under the same training-free setting with identical torchvision checkpoints, StageCal outperforms parallel conversion by 9.00 and 19.40 points at T = 8 and by 2.05 and 6.67 points at T = 16 on ResNet-18 and ResNet-34. The controlled ablation requested by Reviewer 6Q1L further separates the effect of sharing from parameter count and from the residual parameterization.

  2. Competitive performance and practical cost. At a matched scale of about 21.8M parameters and T = 4, StageCal reaches 68.62% on ImageNet-1K, above directly trained SEW-ResNet-34 (67.04%) and TET (68.00%), while requiring only 1,024 unlabeled images and no retraining of the source ANN. On CIFAR-100, its fitting cost is 0.48 GPU-hours, compared with 11.84 to 16.47 GPU-hours for native conversion pipelines that retrain the source ANN. We also state the remaining limitations openly, including the gap to larger directly trained spiking transformers at T = 1.

  3. Experimental reliability and transparency. We traced the reporting inconsistencies identified by Reviewer 6Q1L to two checkpoints and one transcription error, corrected Table 13, Table 4 and Appendix C.4, and revised every conclusion that depended on them. We further clarified that the CLIP baselines are zero-shot, labeled each entry in Table 1 as reported or reproduced, specified the ViT conversion procedure, and provided an operation-level energy breakdown and an agreement analysis for the CLIP result. All main ImageNet results are averaged over five calibration seeds, and we will release the code, logs and configuration files.

We would be grateful if you could consider the strength of our original contributions, the additional evidence, and the progress made during the rebuttal process when making your final assessment.

Thank you again for your time and consideration.

Best regards,

The Authors of StageCal

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.