acceptodds
Under review as a conference paper at ICLR 2027

FastSNN: Accurate Post-Training Compression for Accelerating Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) have shown promising performance on a wide range of perception tasks, making them attractive for edge deployment. However, these powerful SNNs require massive memory and computational resources, often exceeding the tight budgets of edge devices. Model compression is therefore essential, and post-training compression (PTC), which efficiently compresses a pretrained model without retraining, is particularly appealing. Existing PTC methods for SNNs minimize the local reconstruction error of each layer's signals, but this objective overlooks the discrete and temporal nature of spike transmission. In SNNs, a weight perturbation introduced by compression alters the transmitted signal only through spike flips, i.e., a spike is either inserted or deleted. We find that model performance after PTC depends less on how many flips perturbations cause than on whether flips are balanced. Building on this flip-balance perspective, we propose FastSNN, a PTC framework that jointly accounts for signal reconstruction and flip balance during SNN compression. By modeling the relation between weight perturbations and spike flips in a unified form, FastSNN applies to a broad range of compression types, including pruning and quantization. Across multiple datasets and SNN models, FastSNN outperforms the state-of-the-art method by 9.09% on average under 87.5% unstructured pruning and delivers up to a 4.2 inference speedup on a prototype neuromorphic chip.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Post-Training Compression

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Y8TUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a post-training compression framework for spiking neural networks (SNNs), named FastSNN. FastSNN collects two types of information for compression. First, it uses the local reconstruction Hessian matrix to describe the impact of weight changes on the reconstruction of the membrane potential. Then, it uses the Fisher matrix to estimate the impact of weight changes on the final task loss by gradients. Experimental results show that FastSNN outperforms state-of-the-art methods under high sparsity.

Critical strengths
  • This paper provides on-device evaluation, including CPUs and neuromorphic hardware. Experimental results demonstrate its potential for acceleration in deployments.
  • FastSNN outperforms state-of-the-art methods under high sparsity, including 2:8 structured sparsity and 87.5% unstructured sparsity.
Critical weaknesses
  • Although this paper claims to be motivated by flip balance, the two metrics actually used in the proposed method do not directly measure flip balance but are indirectly reflected through the Fisher term in the final loss. The proposed method more closely resembles the Hessian term from existing SBC [1] and a Fisher term that uses the gradient of the final loss to guide layer-by-layer compression, similar to GuidedQuant [2].
  • This paper uses temporal flip balance BlB_l to measure the flip balance. It is defined as the ratio of the balanced spikes to the total flipped spikes. However, this ratio does not directly reflect the total number of unbalanced spikes. I suggest using either the total number of unbalanced spikes or the unbalanced spike firing rate to measure flip balance.
  • The proposed FastSNN does not always outperform state-of-the-art methods on low-bit quantization of spiking transformers.

[1] Shi, Lianfeng, et al. "Spiking brain compression: exploring one-shot post-training pruning and quantization for spiking neural networks." Optimization for Machine Learning. 2025.

[2] Kim, Jinuk, et al. "GuidedQuant: Large language model quantization via exploiting end loss guidance." International Conference on Machine Learning. 2025.

Questions for authors
  1. Please compare FastSNN with SBC on the total number of unbalanced spikes or the unbalanced spike firing rate.
  2. Please explain in detail how FastSNN improves flip balance.
2 comments

To Reviewer Y8TU

We sincerely thank the reviewer for the detailed and expert comments, which helped us state our contribution more precisely.

W1: novelty over SBC and GuidedQuant

FastSNN builds on both: H\mathbf{H} is the SMP Hessian of SBC, and guiding layerwise compression by the end loss follows GuidedQuant. It adds two things.

The first is a Fisher term that measures flip balance. GuidedQuant squares the loss-weighted error of each row and then sums, ∑i,j(∂ℓ/∂Zij⋅ΔZij)2\sum_{i,j}(\partial\ell/\partial Z_{ij}\cdot\Delta Z_{ij})^2. In a spiking layer the rows of one sample are its timesteps, so opposite flips cannot cancel. We sum over timesteps before squaring (Eq. 12). At the last spiking layer the term then equals the squared net spike-count change weighted by the loss sensitivity (Appendix A.2). It is thus a first-order measure of count imbalance itself, which answers the concern that our metrics reflect flip balance only indirectly.

The second is the adaptive mixture, because neither part is sufficient alone. For MS-ResNet-56 on CIFAR-100 under 2:8 pruning (Fig. 2(b)), SBC reaches 33.5%, our Fisher term alone 48.4%, and the mixture 51.7%. The split-half weight that sets the mixture without a search is in neither prior method.

W2 & Q4: BlB_l versus the number of unbalanced spikes

We agree that BlB_l, as a ratio, does not give the number of unbalanced spikes, so we measured both suggested quantities. At the final spiking layer FastSNN has fewer unbalanced spikes than SBC in every setting (6%, 21% and 17% fewer), although it produces more flips.

SEW-ResNet on CIFAR10-DVS, all 1,000 test samples, final spiking layer, SBC / FastSNN:

SparsityAcc. (%)BLB_L (%) ↑ULU_L ↓rU,Lr_{U,L} (%) ↓
2:868.3 / 71.247.9 / 56.2394.7 / 369.94.82 / 4.52
1:856.9 / 66.026.2 / 50.2585.0 / 461.47.14 / 5.63
87.5%63.3 / 70.738.2 / 53.8481.6 / 399.55.88 / 4.88

Ul=∑i∣si+−si−∣U_l=\sum_i|s_i^+-s_i^-| counts, per sample, the flips that cannot be paired with an opposite flip in the same neuron (Eq. 10), and rU,l=Ul/(NlT)r_{U,l}=U_l/(N_lT) is the unbalanced spike firing rate.

Both follow from statistics we already report. Layer ll has RlKlR_lK_l flips, where KlK_l is the number of dense spikes, and a fraction BlB_l of them is balanced, so Ul=(1−Bl)RlKlU_l=(1-B_l)R_lK_l. The reviewer's metrics and ours therefore carry the same information and rank the methods identically. We keep BlB_l in the paper because it isolates the unbalanced fraction from the flip count and has a count-matched chance level (dashed lines in Fig. 1(b) and 2(a)).

W3: low-bit quantization of spiking transformers

On Transformer-series models FastSNN trails the best baseline by about 1% on average at 2 bits (Appendix A.6). However, the best baseline changes across settings. Over the 48 transformer quantization settings, SBC leads at 3 and 4 bits but falls up to 18.4% behind at 2 bits, and GPTQ leads at 2 bits but falls up to 5.6% behind at 3 bits. FastSNN has the smallest mean gap to the best method (0.75%, against 0.98% for GPTQ and 1.79% for SBC) and the smallest worst case (2.4%). Its gain grows with the damage caused by compression, and at 2 bits the best baseline loses 5.2% on transformers against 8.5% on ResNets.

Q5: how FastSNN improves flip balance

For a weight perturbation δr\boldsymbol{\delta}_r, the two metrics penalize different things: δrHδr⊤∝∑n,t∣δVn,r[t]∣2,δrFrδr⊤=∑n(∑tan,r[t])2,\boldsymbol{\delta}_r\mathbf{H}\boldsymbol{\delta}_r^{\top}\propto\sum_{n,t}\\|\delta V_{n,r}[t]\\|^2,\qquad \boldsymbol{\delta}_r\mathbf{F}_r\boldsymbol{\delta}_r^{\top}=\sum_n\Big(\sum_t a_{n,r}[t]\Big)^2, where δVn,r[t]\delta V_{n,r}[t] is the deviation of the surrogate membrane potential and an,r[t]a_{n,r}[t] is its signed first-order effect on the loss at timestep tt (Appendix A.2). The Hessian of SBC is a sum of squares over timesteps, so deviations of opposite sign both add to the cost. The Fisher term is a square of a sum, so contributions of opposite sign cancel before squaring.

ExactOBS (Eq. 8) uses the metric to choose which weight to remove and to compensate the rest. With H(α)\mathbf{H}(\alpha), a weight whose removal causes a one-signed count drift tends to be kept, and the compensation may add opposite deviations that restore the net count (Fig. 1(c)).

Thank you again for the careful and constructive review.

Sorry for my late reply. The rebuttal has fully addressed my concerns. I believe this paper is worthy of acceptance.

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

This paper presents FastSNN, a post-training compression framework for spiking neural networks (SNNs). To address spike insertions and deletions induced by compression, the authors investigate the relationship between the balance of these changes over time and model performance. Building on a temporally averaged readout structure, the framework combines a local membrane-potential reconstruction matrix with the empirical Fisher information matrix of the final task loss, estimating their mixing weights by splitting the calibration set. The framework uses labeled calibration data, requires no retraining, and supports unstructured pruning, N pruning, and weight quantization. The paper evaluates the framework on image classification, event-based visual recognition, object detection, and wireless sensing tasks, and reports inference speedups on CPUs and neuromorphic chips.

Critical strengths
  1. The paper considers the discrete and temporally dependent nature of spike transmission, observing that the number of spike flips alone is insufficient to explain performance changes following compression. This observation is supported by a joint analysis of spike-flip balance, firing rates, and accuracy. These findings offer a valuable perspective on compression errors in SNNs and provide an intuitive rationale for using sensitivity to the final task loss to guide local compression.

  2. A unified compression metric can be combined with existing solvers to support multiple pruning patterns and weight quantization, while the adaptive mixing mechanism reduces the need to select mixing coefficients manually. The experiments cover a range of network architectures and tasks, and further demonstrate practical inference speedups on CPUs and neuromorphic chips. These results highlight the framework’s practical value for deployment under resource constraints.

Critical weaknesses
  1. The hardware evaluation reports both reductions in BOPs and measured inference speedups. It would be helpful to clarify how BOPs are calculated and to provide the basic settings used for inference time measurements, such as the batch size and the scope of the timed execution. It would also be useful to distinguish clearly between computational cost metrics and actual energy consumption, so that readers do not interpret reductions in BOPs as directly implying proportional reductions in energy consumption. These additions would improve the interpretability and reproducibility of the results.

  2. The paper is clearly organized overall. However, the presentation becomes relatively dense when transitioning from intuitive observations about spike-flip balance to the mathematical modeling of sensitivity to the final loss. A more explicit explanation of this transition, together with greater consistency in the relevant terminology, would help readers understand the connections among the various design choices.

  3. Should provide the Memory and other parm. of Orin, probably.

Questions for authors
  1. The paper takes the balance between spike insertions and deletions as its primary motivation, while the proposed method uses the empirical Fisher information matrix of the final task loss to guide compression. For the temporally averaged readout used in the paper, could this design be interpreted as preferentially suppressing spike-count imbalances to which the final task is sensitive, while permitting certain count changes that have relatively little effect on predictions? In other words, how might the cancellation encouraged by Fisher-guided compression be related more precisely to the spike-flip balance defined for individual neurons in the motivation section? I believe that a brief discussion of this connection would further strengthen the intuitive explanation of the method.
3 comments

To Reviewer eFYZ

We sincerely thank the reviewer for the encouraging assessment and the constructive suggestions. We respond to each point below.

W1. How BOPs are computed, and the settings of the latency measurements.

BOPs. We use the standard bit-operation count and, as in OBC (Frantar & Alistarh, 2022)[1], compute it over the compressible layers. For layer ll, BOPsl=MACsl⋅bw,l⋅ba,l⋅(1−sl)\mathrm{BOPs}_l = \mathrm{MACs}_l \cdot b_{w,l} \cdot b_{a,l} \cdot (1-s_l), where bw,lb_{w,l} and ba,lb_{a,l} are the weight and activation bit-widths and sls_l is the sparsity (sl=0.5s_l = 0.5 for 2:4). We also referred BOPs on A.4 Detailed Statistics for interpretability.

We list detailed settings of inference experiments below:

Timing settings.

  • Jetson AGX Orin (32 GB). JetPack 6.1 and TensorRT 10.3. MAXN power mode.
  • Execution. Batch size 1, 640×640 input. FP16 for the dense model; INT8 with 2:4 sparse tensor-core kernels for the compressed models.
  • CPU and Darwin3 (Fig. 2(d)). For CPU and Darwin3 inference, we also use Batch Size = 1 for SpikingVGG16-UT-HAR workload, applying block sparsity detailed in A.5 for deployment. For CPU, the entire model is scoped as timed execution. In Darwin3, due to its hardware constraints, the first encoding convolution and classification header are excluded for timed execution.

W2. BOPs are a computational-cost metric, not energy.

We measure the real energy consumption fo better illustration. The energy is evaluated in Jetson AGX Orin following the guidance under: https://github.com/omnia-unist/Miro - https://embeddeddl.wordpress.com/2018/04/25/convenient-power-measurements-on-the-jetson-tx2-tegra-x2-board/ - https://docs.nvidia.com/jetson/archives/r34.1/DeveloperGuide/text/SD/PlatformPowerAndPerformance/JetsonOrinNxSeriesAndJetsonAgxOrinSeries.html

Results are presented below (three runs, each 60 s of continuous inference on COCO val2017 images):

Model (SpikeYOLO-s)BOP red.mAP50:95Latency (ms)SpeedupEnergy/img (mJ)Energy red.
Dense FP161×44.787.301.00×434.9 ± 2.31.00×
FastSNN 5×5.06×41.933.791.93×203.0 ± 0.82.22×
FastSNN 8×8.18×41.653.781.93×194.2 ± 0.12.35×
FastSNN 10×10.28×41.613.701.97×192.2 ± 0.32.37×

A 5–10× BOP reduction thus gives a 1.9–2.0× speedup and a 2.2–2.3× reduction in energy, which adheres the 2:4 sparse tensor core's features. However, larger BOP reduction do not lead to more energy saving. This is because 2:4 sparse tensor core can only reduce computational cost by at most 2x. We therefore present BOPs explicitly as a hardware-agnostic metric and keep the statements in Sec 4.2 still, and report the measured latency and energy in Appendix A.4 (Table 11) of the revised paper.

W3: the transition from flip balance to the end-loss sensitivity, and terminology

We agree that this step was too compressed, and Appendix A.2 now spells it out (in blue). Section 3.2 shows that flip balance matters, but the balance BlB_l of a layer cannot be optimized directly: flips are discrete, and balance in one layer does not carry over to the next. At the last spiking layer, however, the temporal-mean readout (Eq. 4) sees only spike counts, so the end loss depends on the flips only through the terminal imbalance b\mathbf{b} (Eq. 11). Backpropagating that loss therefore scores a perturbation in any layer by its predicted effect on b\mathbf{b}, which is the Fisher term of Eq. 12.

Q1: how Fisher-guided cancellation relates to per-neuron flip balance

Yes, your reading is correct. At the last spiking layer the temporal-mean readout gives a spike the same weight at every timestep, and the Fisher term reduces to

δrFrδr⊤=∑n(gˉn,r,c^n,r,⊤)2,\boldsymbol{\delta}_r\mathbf{F}_r\boldsymbol{\delta}_r^{\top} = \sum_n \big(\bar{g}_{n,r}\\,\widehat{c}_{n,r}^{\\,\top}\big)^2,

the squared net spike-count change c^n,r\widehat{c}_{n,r} weighted by the loss sensitivity gˉn,r\bar{g}_{n,r}. A paired insertion and deletion in one neuron, which is what BlB_l counts, changes no count and costs nothing. A count change costs more the more the loss depends on that neuron, so changes with little effect on the prediction are tolerated. Fisher guidance is thus a task-weighted form of per-neuron flip balance. In earlier layers the sensitivity varies over time, and a pair cancels only as far as the two sensitivities agree. Appendix A.2 now includes this discussion (in blue).

Thank you again for your thoughtful review. If anything remains unclear, we are happy to elaborate.

[1] Frantar, Elias, and Dan Alistarh. "Optimal brain compression: A framework for accurate post-training quantization and pruning." Advances in Neural Information Processing Systems 35 (2022): 4475-4488.

Good job, I keep my scores.

Thank you again for your compliment on our work. I would be pleased to answer any further questions. Thank you.

Reviewer_3SrwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents FastSNN, a post-training compression framework for spiking neural networks with temporally averaged readouts. Its central motivation is that compression-induced spike insertions and deletions should be considered not only through their frequency, but also through whether their effects balance over time, particularly at the final spiking layer where preserving spike counts preserves the readout. Experiments cover multiple SNN architectures, datasets, and deployment settings. The paper also demonstrates joint compression for deployment and reports inference speedups of up to 2.3× on a CPU and 4.2× on a neuromorphic chip, showing its effectiveness.

Critical strengths
  1. The motivation for the problem is concrete.
  2. Some of the gains at high sparsity levels are substantial.
  3. Experiments on CPU and Darwin3 hardware demonstrate the method's practical effectiveness.
Critical weaknesses
  1. Figure 2(c) shows the performance of two detection models under different BOP budgets, compared with the dense models. However, it does not include sequential pruning–quantization baselines, such as SBC→GPTQ and OSSCAR→GPTQ, under the same BOP budgets. Similarly, the on-chip speedups in Figure 2(d) primarily demonstrate the benefits of compression over dense deployment, without establishing an accuracy–latency advantage over other compression methods.
  2. The claimed unbiasedness of the split-half variance estimator in Eq. (15) is not justified by the derivation as written. In Eq. (13), κ=tr(H)/tr(F)\kappa = tr(H)/tr(F) is computed from the calibration data, but Appendix A.3 treats it as a deterministic constant when deriving Σ=2κFˉ\Sigma = 2\kappa\bar{F} and the split-half variance estimator.
Questions for authors

The main experimental setup states that all methods use the same BN recalibration protocol, whereas the timing discussion for Table 9 states that SBC uses folded BN without recalibration, while the other methods retain and recalibrate BN. Is this difference confined to the timing experiment, or does it also apply to the accuracy comparisons?

1 comment

To Reviewer 3Srw

We sincerely thank the reviewer for these suggestions and questions, and we appreciate the close reading of both the experiments and the appendix. We respond to each point below.

W1: Accuracy comparison with sequential pruning–quantization baselines under the same BOP budgets Following the reviewer's suggestion, we added both baselines to in BOP experiments, and the manuscript has been revised (Fig. 3).

Results. EMS-YOLO on COCO val2017, mAP@50:95(dense: 31.26).

BOP reductionSBC→GPTQOSSCAR→GPTQFastSNN
5×30.0130.0630.51
6×29.9229.9530.34
7×29.8229.9830.29
8×29.4929.3630.06
9×29.4529.5330.21
10×29.5329.2930.05

FastSNN is the most accurate at every budget, and its lead over the stronger baseline in mAP@50:95 is 0.31 to 0.52. Its mAP@50:95 at 10× BOP reduction (30.05) matches what the two baselines reach at 5× (30.01 and 30.06), so FastSNN reaches the same accuracy on this metric at twice the BOP reduction.

W2: Accuracy-Latency advantages over other compression methods We added a same-hardware comparison on the AMD CPU against SBC and manuscript has been revised, with results as follows:

(Averaged over three individual runs)

Block sparsityLatency (ms)Speedup vs. denseSBC acc. (%)FastSNN acc. (%)Δ
65%18.31.99×97.5 ± 0.198.1 ± 0.1+0.5
75%17.42.09×97.3 ± 0.497.7 ± 0.1+0.4
80%16.82.16×97.2 ± 0.497.3 ± 0.1+0.1
87.5%16.22.24×95.7 ± 0.297.1 ± 0.3+1.4
90%16.02.27×95.1 ± 0.496.3 ± 0.5+1.2
92.5%15.92.29×93.9 ± 0.494.9 ± 0.5+1.0

At High-sparsity regime, where the CPU speedup is largest, FastSNN is consistently more accurate there:

  • At 87.5% sparsity (16.2 ms, 2.24× faster than dense), FastSNN reaches 97.1% against SBC's 95.7%.
  • At 90% and 92.5% the gaps are 1.2 and 1.0 points.
  • To stay above 97% accuracy, SBC must stop at 80% sparsity (16.8 ms), while FastSNN can go to 87.5% (16.2 ms).

W3 and W4: unbiasedness of the split-half estimator, and κ\kappa treated as a constant

Thank you for the careful reading of Appendix A.3. It caught an error in our derivation.

κ=tr⁡(H)/tr⁡(F)\kappa = \operatorname{tr}(\mathbf{H})/\operatorname{tr}(\mathbf{F}) depends on the same calibration samples as F(1)\mathbf{F}^{(1)} and F(2)\mathbf{F}^{(2)}, but our derivation moved it outside the expectation, so E[σ^2]=σ2\mathbb{E}[\widehat{\sigma}^2] = \sigma^2 held only for a fixed κ\kappa. The derivation of α⋆\alpha^\star is unaffected: it uses only E[G−Σ]=0\mathbb{E}[\mathbf{G} - \boldsymbol{\Sigma}] = \mathbf{0}, which holds by the definition Σ=E[G]\boldsymbol{\Sigma} = \mathbb{E}[\mathbf{G}] whether or not κ\kappa depends on the data.

We now normalize each half by its own trace:

G(k)=tr⁡(H)tr⁡(F(k)) F(k),Gˉ=12(G(1)+G(2)),σ^2=14∥G(1)−G(2)∥F2.\mathbf{G}^{(k)} = \frac{\operatorname{tr}(\mathbf{H})}{\operatorname{tr}(\mathbf{F}^{(k)})}\,\mathbf{F}^{(k)}, \qquad \bar{\mathbf{G}} = \tfrac{1}{2}\big(\mathbf{G}^{(1)} + \mathbf{G}^{(2)}\big), \qquad \widehat{\sigma}^2 = \tfrac{1}{4}\big\|\mathbf{G}^{(1)} - \mathbf{G}^{(2)}\big\|_F^2 .

Because G(1)\mathbf{G}^{(1)} and G(2)\mathbf{G}^{(2)} are i.i.d., σ^2\widehat{\sigma}^2 is exactly unbiased for the variance of Gˉ\bar{\mathbf{G}}, and each scale κk\kappa_k stays inside the expectation (Proposition 1).

Eq. 15, which our experiments use, is this estimator when the two halves have equal trace. Otherwise it differs by one term: with η=(tr⁡F(1)−tr⁡F(2))/tr⁡F\eta = (\operatorname{tr}\mathbf{F}^{(1)} - \operatorname{tr}\mathbf{F}^{(2)})/\operatorname{tr}\mathbf{F}, the exact form is κ2∥F(1)−F(2)−η F∥F2/(1−η2)2\kappa^2\|\mathbf{F}^{(1)} - \mathbf{F}^{(2)} - \eta\,\mathbf{F}\|_F^2/(1-\eta^2)^2. Section 3.3 now says this instead of calling the estimates unbiased without qualification. The revised appendix also states its assumptions (i.i.d. samples, equal halves chosen independently of the data, H\mathbf{H} treated as fixed) and claims unbiasedness only for σ^2\widehat{\sigma}^2 and Q=∥H−Gˉ∥F2Q = \|\mathbf{H} - \bar{\mathbf{G}}\|_F^2. β^\widehat{\beta} and α^\widehat{\alpha} are plug-in estimates.

We have revisded Appendix A.3 to correct our derivations.

Q1: Whether SBC used the samed BN-fused protocol for accuracy experiments Thank you for spotting this inconsistency. The difference is confined to the timing experiment. In every accuracy comparison, SBC retains BN and its running statistics are recalibrated with the same protocol as all other methods, as stated in Section 4. We have re-timed SBC under the common protocol on the same hardware. It takes 103.4 s, compared with 76.7 s before, and Table 9 and the timing paragraph in Appendix A.4 now use this number. It had no effect on any accuracy result.

Thank you again for the time and care you put into this review. We would be glad to clarify any remaining point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.