acceptodds
Under review as a conference paper at ICLR 2027

MTA-SNN: Revitalizing Convolutional Spiking Neural Networks via Multi-Threshold Parallel Neurons and Global-Local Aggregation

Abstract

Spiking Neural Networks (SNNs) offer a promising avenue for energy-efficient computing via event-driven paradigms, but a significant trade-off exists between the hardware-friendly nature of convolutional architectures and the superior representational capacity of Transformer-based models, whose reliance on complex attention mechanisms can complicate asynchronous deployment. To bridge this gap, we propose MTA-SNN, which revitalizes convolutional SNNs by integrating attention-like capabilities using fully deployable primitives. Our approach introduces three core innovations: (1) Multi-Threshold Parallel Neurons (MTPN), which employ heterogeneous firing thresholds to expand information capacity in the channel dimension; (2) Global Local Aggregation (GLA) that captures long-range dependencies; and (3) Multi-Scale Position Encoding (MSPE) via dilated convolutions to explicitly enrich spatial context. Extensive experiments demonstrate that MTA-SNN establishes new state-of-the-art results among low-latency SNNs, achieving 83.07% on CIFAR-100 and 78.37% on ImageNet with a single timestep. This performance significantly outperforms existing convolutional SNNs and matches transformer-based counterparts requiring multiple timesteps, proving that high-performance vision SNNs can be realized within a strictly neuromorphic-friendly convolutional framework.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spiking Neural Networks

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

This paper proposes MTA-SNN, which aims to bridge the gap between the hardware-friendliness of convolutional SNNs and the high representational power of transformer-based SNNs. Three core modules are introduced: (1) Multi-Threshold Parallel Neurons (MTPN) that employ multiple LIF neurons with heterogeneous firing thresholds sharing the same weights in parallel, expanding information capacity in the channel dimension without increasing timesteps; (2)Global Local Aggregation (GLA) that combines depthwise convolution (local path) with global average pooling + 1×1 conv (global path) to approximate self-attention; and (3) Multi-Scale Position Encoding (MSPE) via dilated depthwise convolutions at rates {1, 2, 4} to inject multi-scale spatial context. With a single timestep (T=1) in a 3-stage hierarchical design, MTA-SNN achieves 83.07% on CIFAR-100 and 78.37% on ImageNet, significantly surpassing existing convolutional SNNs and matching transformer-based counterparts requiring multiple timesteps.

Critical strengths
  1. The paper makes good latency-performance trade-off. It achieves high accuracy at T=1. This is highly suitable for asynchronous neuromorphic hardware.

  2. The decomposition of “attention-like” behavior into depthwise convolutions and global pooling is neuromorphic-friendly and avoids multi-path spike matrix multiplications.

  3. The comparison between heterogeneous and identical thresholds supports the importance of threshold diversity. The ablations on spatial encoding and aggregation further help explain how the components contribute to the final performance.

Critical weaknesses
  1. The classification results are encouraging, but they do not establish whether threshold diversity also benefits spatially detailed tasks. Evidence on such tasks is needed to strengthen the contribution.

  2. Classic CNN attention modules also combine local features and global context, yet their behavior under the same spiking constraints remains unclear. A controlled comparison is needed to substantiate which aspects of GLA are particularly suitable for single-timestep SNNs.

  3. Existing SNN transformers employ various position encoding strategies: Spikformer uses a single-scale 3×3 conv-based conditional position encoding (CPE), QKformer relies on implicit positional information from hierarchical conv stems, and recent work proposes Gray-PE and Log-PE based on binary encoding for relative position representation. MSPE's multi-scale dilated depthwise convolution design is distinct from these approaches, as it explicitly injects spatial context at multiple receptive field scales. The authors should provide a concise comparison and clearly explain the relevant implementation differences.

  4. Parallel thresholds reduce the need for temporal integration but introduce additional spike activity and state or communication costs. Parameter count and timestep count alone are insufficient to support claims of practical efficiency. The authors should provide a clearer breakdown of these costs and identify the conditions under which MTPN offers meaningful efficiency gains.

  5. The effectiveness of heterogeneous thresholds leaves an important methodological question unresolved: how should their number and locations be selected for a new task? The authors should provide clearer guidance and relate these choices to activation statistics or firing behavior.

Questions for authors
  1. Does MTPN also improve dense prediction tasks such as detection and segmentation?

  2. For GLA, did you compare it with classic CNN attention modules under the same spiking constraints?

  3. Can the authors provide direct comparisons of MSPE against existing SNN PE methods, such as Spikformer's single-scale CPE, QKformer's implicit PE via hierarchical convolutional stems, and Gray-PE/Log-PE?

  4. Can the authors provide a clearer cost breakdown for MTPN, including SOPs, event count, memory access, routing overhead, and latency?

  5. How are the thresholds of MTPN chosen, and can the authors provide guidelines for threshold selection across different datasets and tasks?

4 comments

Comment1: Does MTPN also improve dense prediction tasks such as detection and segmentation?

Responses to Comment1: Thank you for asking us to highlight this part. We already include dense-transfer results in Appendix H. We transfer the ImageNet-pretrained MTA-SNN-10-384 backbone, retaining MTPN, GLA, and MSPE, to COCO object detection with Mask R-CNN and ADE20K semantic segmentation with Semantic FPN.

TaskMethodTMetricParamsEnergy
COCO detectionMeta-SpikeFormer144.0 [email protected]34.9M49.5 mJ
COCO detectionMeta-SpikeFormer151.2 [email protected]75.0M140.8 mJ
COCO detectionMTA-SNN-10-384150.0 [email protected]17.31M29.7 mJ
ADE20K segmentationMeta-SpikeFormer435.3 mIoU58.9M183.6 mJ
ADE20K segmentationMTA-SNN-10-384141.70 mIoU17.31M20.9 mJ

These results show that the MTPN-containing backbone transfers beyond classification with competitive dense-prediction quality and lower estimated backbone energy. We will state the scope clearly: these are full-backbone transfer results rather than an isolated MTPN-only dense-task ablation.

Comment2: For GLA, did you compare it with classic CNN attention modules under the same spiking constraints?

Response to Comment2: Thank you for this question. Appendix K (Table 13) contains a controlled plug-and-replace comparison under matched T=1 spiking conditions. All variants use the same backbone, MTPN/MSPE modules, insertion locations, optimizer, augmentation, and training schedule; only the attention module is replaced.

Attention moduleTop-1 (%)Difference from GLA
GLA (Ours)83.16--
SE-Block80.41-2.75
CBAM66.97-16.19

These results show that conventional CNN attention modules do not transfer robustly to the single-timestep sparse-spike regime. SE and CBAM rely on multiplicative gating estimated from sparse binary activations, which can suppress informative spikes. GLA avoids such multiplicative gates and instead performs local/global aggregation through convolution, pooling, concatenation, and linear fusion. We will highlight this comparison more prominently.

Comment3: Can the authors provide direct comparisons of MSPE against existing SNN PE methods, such as Spikformer's single-scale CPE, QKformer's implicit PE via hierarchical convolutional stems, and Gray-PE/Log-PE?

Responses to Comment3: Thank you for this suggestion. Controlled comparisons are already provided in Appendix G, Table 8, which we reproduce below for convenience. All configurations use the same backbone and training protocol on CIFAR-100, with single-timestep inference (T=1).

ConfigurationTime stepsAccuracy (%)
Base + CPE (Spikformer)179.84
Base + Implicit PE (QKformer)180.58
Base + Gray-PE179.72
Base + Log-PE180.16
Base + MSPE (Ours)181.16

MSPE achieves the highest accuracy among these configurations, improving over CPE and implicit PE by 1.32 and 0.58 percentage points, respectively, and over Gray-PE and Log-PE by 1.44 and 1.00 points. These controlled results support the benefit of explicitly incorporating spatial context at multiple receptive-field scales. Specifically, MSPE uses dilated depthwise convolutions with rates {1,2,4}, extending spatial aggregation beyond single-scale convolutional encoding and the positional information implicitly supplied by hierarchical stems.

For implementation clarity, as specified in Appendix G, Gray-PE and Log-PE are labels for our controlled implementations of Gray-code-inspired positional coding and logarithmically spaced sinusoidal/Fourier coding, respectively; they should not be interpreted as exact reproductions of particular published SNN implementations. We will strengthen the main-text reference to Table 8 and make these implementation distinctions more explicit.

Comment4: Can the authors provide a clearer cost breakdown for MTPN, including SOPs, event count, memory access, routing overhead, and latency?

Responses to Comment4: Thank you for requesting clarification. Appendix E analyzes MTPN’s computation and memory costs, and Appendix A.1 addresses event-dependent latency. The ImageNet comparison uses the same MTA-SNN-10-384 backbone:

Cost componentSingle-threshold baselineMTPN
Threshold branches KK14
Timesteps TT41
Expected spikes per neuron group per inference1.401.13
Normalized SOP estimate1.0000.807
Estimated energy (mJ/image)7.164.92
Selected memory traffic¹32 B: membrane reads/writes4 B: byte-aligned branch-map reads
Learned weight sets per layer11, shared

¹ Selected traffic terms modeled in Appendix E.3, not measured total memory traffic.

SOPs and events. All branches are counted:

SOPl=FlT∑i=1Krl,i,\mathrm{SOP}_l = F_l T \sum_{i=1}^{K} r_{l,i},

where FlF_l is the dense synaptic-operation count and rl,ir_{l,i} the branch firing rate. The measured rates sum to 0.49+0.36+0.17+0.11=1.130.49+0.36+0.17+0.11=1.13, versus 4×0.35=1.404\times0.35=1.40, giving 19.3% fewer normalized SOPs. Weight sharing saves learned parameters, not branch events or accumulations.

Memory and routing. Single-timestep execution removes repeated temporal-state accesses but adds parallel threshold evaluations and branch outputs. Physical weight replication, buffering, and routing depend on hardware mapping. Emitted spikes are distinct from routed traffic, which also depends on fan-out, multicast, and placement; routing overhead has not been separately measured.

Latency. Appendix A.1 uses L=αT+βErouteL=\alpha T+\beta E_{\mathrm{route}}. MTPN is faster when

< \alpha(T_{\mathrm{baseline}}-T_{\mathrm{MTPN}}),$$ meaning temporal savings outweigh additional event-handling costs. Separately, Appendix I reports **training-time** GPU runtime decreasing from **3.90 to 1.55 ms/sample** and peak memory from **183.75 to 68.22 MB/sample**, comparing $(K,T)=(1,3)$ with $(3,1)$. These are not chip-level inference-latency measurements. We will consolidate these results and assumptions in the revision. **Comment5:** How are the thresholds of MTPN chosen, and can the authors provide guidelines for threshold selection across different datasets and tasks? **Responses to Comment5:** Thank you for this insightful question. We agree that the threshold configuration should be properly motivated. In our method, we intentionally employ a fixed threshold configuration to prioritize hardware robustness and reproducibility over dataset-specific tuning. The chosen set {0.5, 1.0, 2.0} is centered around the standard STN baseline threshold (1.0). It extends this baseline with a lower threshold (0.5) to capture weak or low-amplitude responses, and a higher threshold (2.0) to preserve strong, high-amplitude features. This design naturally aligns with the signal-intensity distribution, enabling a coarse-to-fine response decomposition within a single timestep. Regarding sensitivity, we have added an ablation study in Appendix Table 7. The results demonstrate that performance remains stable across a reasonable range of threshold triplets. However, overly compressed or excessively shifted threshold sets are less effective. The selected {0.5, 1.0, 2.0} configuration provides the most consistent trade-off between accuracy and computational efficiency. We will update the manuscript to clarify this rationale.
Reviewer_yRVdOpenReview ↗

Review

Thank you for the detailed response and additional experiments.

The COCO and ADE20K results evaluate the full MTA-SNN backbone, so they establish transferability but leave MTPN’s individual contribution to dense prediction unresolved. A matched single-threshold baseline at (T=1) would help isolate that contribution. For hardware efficiency, the response describes the ImageNet firing rates as measured, while Table 5 identifies two branch rates as adjusted or extrapolated. Please clarify which rates were measured directly, since they inform the reported SOP reduction. The Speck2e experiment demonstrates that a compact multi-threshold variant can run on physical hardware, but it does not measure routing overhead or inference latency for the full model.

The claimed latency advantage therefore remains a theoretical prediction. I maintain my original rating.

We appreciate your constructive follow-up questions. Our experiments provide complementary evidence at the neuron, architecture, and hardware levels. With the branch count fixed, heterogeneous thresholds improve CIFAR-100 accuracy from 80.15% to 83.16% over identical thresholds. Neuron replacement alone also improves ResNet-18 accuracy by 4.31 percentage points, supporting a benefit from threshold diversity beyond increasing neuron count. The COCO and ADE20K results further demonstrate that the complete architecture transfers to spatially detailed tasks. The suggested single-threshold baseline at T=1T=1 would isolate MTPN’s task-specific contribution more directly; we will distinguish this question from the component effectiveness and architectural transferability established by the existing experiments.

Regarding firing rates, the distinction between measurements and estimates requires clarification. The CIFAR-100 rates of 0.450.45, 0.350.35, and 0.180.18 in Table 5 were directly measured. Its ImageNet entries include adjustment and extrapolation: specifically, 0.480.48 and 0.120.12 should not be described as direct measurements. The separate rate profile in Appendix E.2 also needs to be reconciled with the underlying experimental records. We will verify the provenance of these values and make the reporting and associated calculations consistent. The reported 19.3% reduction, calculated as 1−1.13/1.401-1.13/1.40, is a normalized SOP estimate conditional on the stated firing rates, not a measured chip-level energy reduction. Our accounting explicitly includes events and synaptic accumulations from every threshold branch; it does not infer efficiency solely from T=1T=1.

The Speck2e experiment directly establishes physical execution of the tied-weight multi-threshold mechanism on asynchronous neuromorphic hardware. The compact model runs on four DynapCNN cores, achieving 93.63% accuracy and measured net dynamic energy of 293.65 μJ/sample. These results provide concrete deployment evidence. Full-model routing overhead and inference latency constitute a further level of system-performance characterization and are not established by this compact deployment. We will clearly separate physical measurements, SOP-based estimates, and latency modeling so that the complementary contributions of these results are explicit.

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

This paper proposes MTA-SNN, a convolutional spiking neural network that aims to match spiking transformers in accuracy while using only neuromorphic-friendly operations instead of spike-based self-attention. Its main component is Multi-Threshold Parallel Neurons (MTPN): parallel LIF neurons with different firing thresholds that share convolution weights, which the authors argue increases information capacity at a single timestep instead of adding timesteps. Two further modules replace transformer components. Global Local Aggregation (GLA) combines a depthwise convolution with global average pooling as an attention substitute, and Multi-Scale Position Encoding (MSPE) uses dilated depthwise convolutions to add spatial context. At T=1, the model reports 97.03%/83.07% on CIFAR-10/100 and 78.37% on ImageNet, plus 84.8% on CIFAR10-DVS at T=16. These results are supported by CIFAR-100 ablations, simple theoretical results on quantization levels and gradient flow, estimated energy savings, transfer to detection and segmentation, and a small-scale run on a Speck2e neuromorphic chip.

Critical strengths

1.   Model gain 83.07% on CIFAR-100 and 78.37% on ImageNet using single T=1 which easily compete spiking transformer.

2.   A simple hardware is used (Speck2e chip) for deployment using a basic network settings. 

3.   Promising ablation for each module.

4.   Admits the limitations and trade-off between accuracy and deployment.

Critical weaknesses

1.   False claims of single threshold. Comparison to MT-SNN and MLF which propose multi-threshold summing.

2.   Comparison with SOTA is confusing with number 97.03% and 97.04%. broken base line for CBAM because it drops to 66.97%.

3.   It hard to know how GLA works asynchronously.

4.   Number are in consistent in table 1, 4, 11 and abstract.

5.   Table 2 need Meta-SpikeFormer 80.0% top-1 score at 55M parameters).

6.   QKformer numbers need to revised to its original gain. Third-party evaluation is unfair compare to original. 

7.   Novelty concerns in GLA. A 7x7 Con. and GAP of 1x1 similarity to SE or PoolFormer.

Questions for authors

Please address the points from critical weaknesses for possible improve rating.

  1. How does MTPN differ from MLF (IJCAI 2022), MT-SNN (parallel multi-threshold mode), and I-LIF/SpikeYOLO? Please cite and compare against them.
  2. Which CIFAR-100 result is final: 83.07% (Abstract, Table 1) or 83.16% (Contributions, Sec. 5.1, Tables 4, 7, 11)?
  3. How is GLA’s global average pooling executed asynchronously, given that it needs the whole feature map before broadcasting, which is the same issue Sec. 4.4 raises against Q×K attention?
  4. What makes GLA novel compared with SE- or PoolFormer-style mixing? Was CBAM tuned at all, given that it collapses to 66.97% in App. K?
  5. Why is Meta-SpikeFormer’s ImageNet result (80.0%, 55M) omitted from Table 2? Why is only the re-evaluated QKformer number reported rather than its original result? If a strict-spiking standard is applied, does it also cover MTPN’s summed, non-binary outputs?
  6. How is the CIFAR-10 state-of-the-art claim justified when MaxFormer reports 97.04% vs. your 97.03% in Table 1?
3 comments

Comment1: How does MTPN differ from MLF (IJCAI 2022), MT-SNN (parallel multi-threshold mode), and I-LIF/SpikeYOLO? Please cite and compare against them.

Responses to Comment1: Thank you for this constructive comment. MTPN does not claim to be the first multi-threshold SNN method, and we will correct the overly general single-threshold statement. Its novelty lies in combining heterogeneous thresholds, independent binary-spike branches, shared convolutional weights, and parallel single-timestep execution in a fully convolutional architecture.

MLF [1] also uses parallel heterogeneous-threshold LIF neurons with summed outputs; thus, parallel multi-threshold firing alone is not our claimed novelty. Parallel-MT [2] compares thresholds against a shared membrane, Cascade-MT [2] sequentially updates residual potentials, and I-LIF/SpikeYOLO [3] expands integer activations through virtual timesteps. MTPN instead uses independent standard binary LIF branches with tied convolutional weights, followed by spike-domain aggregation, distributing multi-level representation across parallel binary channels compatible with asynchronous, event-driven execution.

For a P-parameter convolution with K branches, weight sharing retains approximately P, rather than KP, stored weights, while additional neuron-state and spike-processing costs remain. Our ablation isolates the benefit of threshold diversity: replacing {0.5,1.0,2.0} with three identical thresholds reduces CIFAR-100 accuracy from 83.16% to 80.15%. Neuron replacement alone also improves ResNet-18 accuracy by 4.31 percentage points.

Speck2e deployment validates physical execution, but does not establish resource-matched superiority over prior methods. We will clarify these structural distinctions and the space-for-time trade-off, separating ablation and deployment evidence from hardware-matched comparisons.

[1] Feng et al., Multi-Level Firing with Spiking DS-ResNet, IJCAI, 2022.

[2] Wang and Zhang, MT-SNN: Enhance Spiking Neural Network with Multiple Thresholds, arXiv:2303.11127.

[3] Luo et al., Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection, ECCV, 2024.

Comment2: Which CIFAR-100 result is final: 83.07% (Abstract, Table 1) or 83.16% (Contributions, Sec. 5.1, Tables 4, 7, 11)?

Responses to Comment2: Thank you for pointing out the unclear reporting. Both values correspond to the same model configuration and experimental protocol. 83.16% is the result obtained with the random seed used for our ablation study, whereas 83.07% ± 0.11% is the mean ± standard deviation across three independent random seeds, reported in Table 1 and the Abstract. The ablation-study run also achieved the highest accuracy among these three runs. Thus, the difference reflects single-seed versus multi-seed reporting, rather than changes in architecture, training configuration, or model version.

We will consistently report 83.07% ± 0.11% as the main CIFAR-100 result and explicitly identify 83.16% as the single-seed result used in the ablation study. We will clarify this distinction in the relevant text and table captions and provide the three seed identifiers and corresponding accuracies for reproducibility.

Both values use the same model and protocol. 83.16% is the ablation-study seed’s result, also the best of three independent seeds; 83.07%±0.11% is their mean±SD (Table 1/Abstract). The difference is single- vs multi-seed reporting, not architecture, training or model-version changes. We will use 83.07%±0.11% as the main CIFAR-100 result, label 83.16% as the single-seed ablation result in text/captions, and provide all three seed IDs and accuracies for reproducibility.

Comment3: How is GLA’s global average pooling executed asynchronously, given that it needs the whole feature map before broadcasting, which is the same issue Sec. 4.4 raises against Q×K attention?

Responses to Comment3: Thank you for this constructive comment. GLA’s global pooling requires a finite observation window to define the completed average, but does not require the whole feature map before accumulation begins. Appendix L.1 demonstrates this event-domain formulation in the asynchronous simulation of MTA-SNNtiny.

Specifically, the local pathway applies depthwise convolution to arriving events, while the global pathway incrementally accumulates spatial contributions using fixed pooling weights, with the normalization factor 1/(HW) absorbed into the synaptic mapping. Both pathways route their outputs to downstream spiking neurons for current accumulation and fusion, without requiring lockstep execution. The exact finite-window average is available only after the window closes; asynchronous accumulation does not remove this dependency or the associated communication costs.

We also clarify the implementation scope: the current Speck2e low-level operators do not support the global-pooling and path-fusion mappings required by GLA. Thus, GLA’s asynchronous execution is supported by simulation, while the physical-chip experiment validates only the compact MTSNN’s tied-weight multi-threshold mechanism.

We will revise Sec. 4.4 to avoid implying that global dependencies alone make QK attention incompatible with asynchronous execution. The relevant distinction is GLA’s fixed-weight spatial accumulation versus input-dependent query–key multiplication, rather than the absence of a global dependency.

Comment4: What makes GLA novel compared with SE- or PoolFormer-style mixing? Was CBAM tuned at all, given that it collapses to 66.97% in App. K?

Responses to Comment4: Thank you for this constructive comment. GLA’s contribution lies in combining learned local filtering and global-context aggregation for single-timestep SNNs, rather than introducing either operation individually. Its local pathway uses depthwise convolution, while a separate global pathway performs spatial aggregation. The pathways are fused through concatenation and learned linear projection, without multiplicative gating. This linear fusion can be expressed as branch-specific projections followed by downstream current accumulation in the event-domain formulation. GLA thus differs from SE-style channel recalibration and PoolFormer-style pooling-based mixing. Appendix J and Figure 6 support the complementary roles of its two pathways.

CBAM underwent convergence checks and limited learning-rate and weight-decay tuning. We replaced its ReLU with spiking neurons and substituted it for all GLA modules, retaining MTPN, MSPE, and the backbone. It achieved 66.97%, compared with GLA’s 83.16%. We will correct Appendix K’s “no module-specific tuning” statement and document the tuning scope and convergence checks.

These results support GLA’s effectiveness in the evaluated single-timestep configuration, rather than a general superiority over CBAM. They suggest that direct activation replacement with limited tuning is insufficient in this setting. Although multiplicative gating may be sensitive to sparse spikes, our experiment does not establish this as the cause; CBAM may require further SNN-specific adaptation.

Comment5: Why is Meta-SpikeFormer’s ImageNet result (80.0%, 55M) omitted from Table 2? Why is only the re-evaluated QKformer number reported rather than its original result? If a strict-spiking standard is applied, does it also cover MTPN’s summed, non-binary outputs?

Response to Comment5: Thank you for raising this question. We will clarify our comparison criteria.

Our baseline is MaxFormer’s QKFormer, a membrane-shortcut variant preserving binary outputs, achieving 77.78% at T=1T=1. MaxFormer modifies the original pre-spike shortcut because its ternary outputs {0,1,2}\{0,1,2\} require multi-spike execution and additional costs. The original QKFormer’s 85.65% uses T=4T=4 and 384×384384 \times 384 inputs, versus T=1T=1 and 224×224224 \times 224 here. We will label this modified baseline explicitly.

We will also include Meta-SpikeFormer’s 80.0%, annotating its 55.4M parameters, four timesteps, and knowledge distillation, which differ from our single-timestep, non-distilled setting.

The same criterion applies to MTPN:

\qquad S_i \in \{0,1\}.$$ Branches transmit binary events; their non-binary sum represents postsynaptic current, not an analog activation transmitted as a spike. **All branch events enter our SOP accounting.** We will distinguish event representation from internal accumulation and apply consistent representation and cost-accounting criteria to competing methods, including multi-spike approaches.

Comment6: How is the CIFAR-10 state-of-the-art claim justified when MaxFormer reports 97.04% vs. your 97.03% in Table 1?

Responses to comment6: We agree that our claim requires qualification. MTA-SNN achieves 97.03% at T=1, nearly matching MaxFormer’s 97.04% at T=4, rather than exceeding it. It also outperforms Spikformer, Spikingformer, and QKformer (T=4) by 1.52, 1.08, and 0.85 percentage points, respectively. These results support effective single-timestep representation, but timestep count alone does not establish hardware speedup. We will replace the unqualified state-of-the-art claim with “the highest CIFAR-10 accuracy among the single-timestep methods compared in Table 1, competitive with leading multi-timestep SNNs.”

Reviewer_a38YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MTA-SNN, a convolutional spiking neural network for low-latency visual recognition. It introduces Multi-Threshold Parallel Neurons (MTPN), which use parallel neurons with heterogeneous firing thresholds and shared convolutional weights to improve information encoding within a single timestep. Global-Local Aggregation (GLA) combines depthwise convolution and global average pooling to capture spatial context, while Multi-Scale Position Encoding (MSPE) uses dilated convolutions to enhance spatial representations. The authors also analyze the information capacity and approximation properties of MTPN. Experiments on CIFAR-10/100, ImageNet, CIFAR10-DVS, and preliminary neuromorphic hardware demonstrate competitive accuracy, including approximately 83% on CIFAR-100 and 78.37% on ImageNet at a single timestep.

Critical strengths

Strength 1: The paper presents a clear design for low-latency convolutional SNNs by replacing temporal integration with parallel multi-threshold processing. MTPN shares convolutional weights across branches, while GLA and MSPE improve spatial feature extraction. This design has potential advantages for neuromorphic deployment.

Strength 2: The paper achieves competitive single-timestep accuracy on CIFAR-10/100 and ImageNet, supported by component-wise ablations. On CIFAR-100, heterogeneous MTPN improves accuracy by 3.01 percentage points over identical-threshold parallel neurons. Preliminary simulation and physical-chip experiments also provide some evidence of hardware feasibility.

Critical weaknesses

Weakness 1: The methodological novelty relative to existing multi-threshold SNNs is unclear. SpikeAtConv [1], Deep Multi-Threshold Spiking-UNet [2], Stair-LIF [3], and One-Hot Multi-Level LIF [4] have already explored multi-threshold or multi-level spike encoding. MTPN mainly differs in its shared-weight parallel convolutional design, but these differences are not sufficiently discussed or experimentally compared. GLA and MSPE also largely build on established convolution, pooling, and positional encoding techniques.

Weakness 2: Some theoretical claims require correction or stronger justification. Proposition 2 in Appendix A.4 claims a single-threshold approximation error lower bound of (R/2), but choosing (v=R/3) and (w=2R/3) gives a maximum error of (R/3). Theorem 1 establishes a comparison of discrete encoding levels rather than efficiency under matched hardware budgets. The gradient analysis also overlooks possible cancellation between branch gradients.

Weakness 3: The independent contribution of GLA and MSPE is limited. GLA combines depthwise convolution and global pooling, while MSPE uses conventional multi-scale dilated convolutions. Table 4 shows that replacing GLA with SSA reduces CIFAR-100 accuracy by only 0.06 percentage points. The paper also lacks direct evidence that GLA provides meaningful asynchronous execution benefits over existing attention mechanisms.

Weakness 4: The claimed energy efficiency is not consistently supported by matched comparisons. On ImageNet, MTA-SNN-10-768 improves accuracy over QKformer by only 0.59 percentage points, while estimated energy increases from 6.79 mJ to 14.11 mJ. Although Appendix E reports savings over a multi-timestep baseline, these results rely mainly on analytical energy estimates and do not establish a superior accuracy–energy trade-off against strong SNN alternatives.

Weakness 5: The hardware experiments do not fully validate the asynchronous deployment claims. Appendix L deploys only a compact two-threshold MTSNN on Speck2e rather than the complete MTA-SNN architecture. Physical-chip latency is not reported, and the main energy results remain estimates. The implementation costs of global pooling and feature broadcasting under asynchronous execution are also insufficiently explained.

Weakness 6: Several experimental details need clarification. Table 1 reports 83.07% CIFAR-100 accuracy, whereas Table 4 reports 83.16%. The paper also lacks systematic comparisons under matched parameter counts, synaptic operations, and neuron budgets. Moreover, the energy analysis combines measured and estimated firing rates, limiting the reproducibility of the efficiency claims.

[1] Liao, W., Chen, F., Liu, C., Wang, W., and Liu, H. SpikeAtConv: An Integrated Spiking-Convolutional Attention Architecture for Energy-Efficient Neuromorphic Vision Processing. Frontiers in Neuroscience, 19:1536771, 2025.

[2] Li, H., Zhang, Y., Xiong, Z., and Sun, X. Deep Multi-Threshold Spiking-UNet for Image Processing. Neurocomputing, 586:127653, 2024.

[3] Luo, J., Chen, Y., Liu, Y., Wang, J., Yu, Z., and Xiao, S. Stair-LIF: Boosting the Representation of Spiking Neural Networks with Learnable Incremental Multi-Threshold Neurons. In IEEE International Conference on Multimedia and Expo (ICME), 2025.

[4] Abillama, P., Lee, C., Bejarano-Carbo, A., Zhang, Q., Sylvester, D., Blaauw, D., and Kim, H.-S. One-Hot Multi-Level Leaky Integrate-and-Fire Spiking Neural Networks for Enhanced Accuracy-Latency Tradeoff. IEEE Access, 13:38163–38180, 2025.

Questions for authors

Q1: Given the existing multi-threshold SNN methods [1–4], could the authors clarify the independent novelty of MTPN, particularly the benefits of its shared-weight design? Could direct comparisons with these approaches be provided?

Q2: Proposition 2 claims an approximation error lower bound of (R/2), but (v=R/3) and (w=2R/3) yield an error of (R/3). Could the authors correct this result and clarify how Theorem 1 supports efficiency under matched hardware budgets?

Q3: Replacing GLA with SSA changes CIFAR-100 accuracy by only 0.06 percentage points. Could the authors clarify GLA's independent advantages and compare its actual latency, synaptic operations, and energy consumption with existing attention modules?

Q4: MTA-SNN-10-768 reports roughly twice the estimated energy of QKformer for a 0.59-point accuracy improvement. Could the authors provide matched accuracy–latency–energy comparisons and quantify the computational overhead of multi-threshold parallelism?

Q5: The physical-chip experiments evaluate only a compact MTSNN. Could the authors explain how GLA, global pooling, and multi-threshold branches would operate asynchronously, and provide more complete chip-level latency, energy, and memory measurements?

Q6: Could the authors clarify the discrepancy between 83.07% and 83.16% CIFAR-100 accuracy and report random-seed statistics? Could they also compare different threshold counts and multi-timestep baselines under matched computational and synaptic-operation budgets?

5 comments

Comment1: Given the existing multi-threshold SNN methods [1–4], could the authors clarify the independent novelty of MTPN, particularly the benefits of its shared-weight design? Could direct comparisons with these approaches be provided?

Responses to Comment1: Thank you for this constructive comment. We clarify that MTPN does not claim to be the first multi-threshold SNN method. Its novelty lies in combining heterogeneous thresholds with independent binary-spike branches, shared convolutional weights, and parallel single-timestep execution in a fully convolutional architecture.

Specifically, Spiking-UNet [2] sequentially fires and resets a shared membrane, Stair-LIF [3] generates multi-level responses through learnable incremental thresholds and multi-spike firing, and One-Hot Multi-Level LIF [4] uses mutually exclusive interval-based coding. Although SpikeAtConv [1] also considers parallel heterogeneous thresholds, MTPN differs in using independent standard binary LIF branches with tied convolutional weights, followed by spike-domain aggregation. Thus, MTPN distributes multi-level representation across parallel binary channels while preserving asynchronous, event-driven execution.

The shared-weight design is important because a convolution with P parameters requires approximately KP stored parameters with untied weights, whereas MTPN retains approximately P parameters for K threshold branches. This substantially reduces synaptic weight storage and avoids replicating learned filters, while the additional spike-processing cost is explicitly acknowledged. Our ablation further shows that the gain comes from threshold diversity: replacing {0.5,1.0,2.0} with three identical thresholds reduces CIFAR-100 accuracy from 83.16% to 80.15%.

Since [1–4] are not designed for asynchronous execution, direct numerical comparisons would not constitute a hardware-matched evaluation. We will clarify these methodological distinctions and the associated space-for-time trade-off in the revision.

[1] Liao, W., Chen, F., Liu, C., Wang, W., and Liu, H. SpikeAtConv: An Integrated Spiking-Convolutional Attention Architecture for Energy-Efficient Neuromorphic Vision Processing. Frontiers in Neuroscience, 19:1536771, 2025.

[2] Li, H., Zhang, Y., Xiong, Z., and Sun, X. Deep Multi-Threshold Spiking-UNet for Image Processing. Neurocomputing, 586:127653, 2024.

[3] Luo, J., Chen, Y., Liu, Y., Wang, J., Yu, Z., and Xiao, S. Stair-LIF: Boosting the Representation of Spiking Neural Networks with Learnable Incremental Multi-Threshold Neurons. In IEEE International Conference on Multimedia and Expo (ICME), 2025.

[4] Abillama, P., Lee, C., Bejarano-Carbo, A., Zhang, Q., Sylvester, D., Blaauw, D., and Kim, H.-S. One-Hot Multi-Level Leaky Integrate-and-Fire Spiking Neural Networks for Enhanced Accuracy-Latency Tradeoff. IEEE Access, 13:38163–38180, 2025.

Comment2: Proposition 2 claims an approximation error lower bound of (R/2), but (v=R/3) and (w=2R/3) yield an error of (R/3).

Responses to Comment2: Thank you for identifying this error. We agree that the R/2R/2 lower bound is incorrect. Under f(x)=xf(x)=x on [0,R][0,R] and f^(x)=wΘ(x−v)\hat f(x)=w\Theta(x-v),

∥f^−f∥∞=max⁡�0�.\|\hat f-f\|_\infty = \max�0�.

Its minimum is R/3R/3, attained at v=R/3v=R/3 and w=2R/3w=2R/3. We will correct Proposition 2 accordingly.

For our KK-threshold construction, vi=iR/Kv_i=iR/K and w=R/Kw=R/K provide an upper bound ∥f^−f∥∞≤R/K\|\hat f-f\|_\infty \le R/K. We will revise Theorem 1 to state that KK ordered thresholds provide at most K+1K+1 instantaneous levels, while TT binary spike counts provide at most T+1T+1, requiring T≥KT \ge K to match the level count.

Since

= \sum_i a_i h'(u-v_i)$$ can vanish through cancellation or zero upstream gradients, our general claim is limited to expanded local surrogate-gradient support. Strictly reduced zero-derivative probability requires a common potential distribution, retention of the baseline threshold, and positive mass in added support. Shared linear aggregation prevents branchwise cancellation with nonzero equal upstream gradients and nonnegative surrogate derivatives, but cannot guarantee non-vanishing network gradients.

Comment3: Could the authors clarify GLA's independent advantages and compare its actual latency, synaptic operations, and energy consumption with existing attention modules?

Response to Comment 3: Thank you for this constructive comment. GLA’s main contribution is hardware-compatible local–global aggregation, combining depthwise convolution and global pooling rather than relying solely on a small accuracy gain. We added the following comparisons at T=1, aggregated over four attention modules:

ModuleMeasured GPU latency¹ (ms)Synaptic operations² (M/image)Estimated synaptic energy² (mJ/image)
GLA0.46644.6250.040
SSA0.44280.1170.072
¹ RTX 4090, FP32, batch size 1, CUDA Graph replay; summed independently measured module latencies.
² Measured spike activity across the full CIFAR-100 test set; energy estimated at 0.9 pJ per accumulation.
GLA reduces SOPs and estimated synaptic energy by 44.3%, but shows no GPU latency advantage.
The architectural distinction concerns native operator support: implementing binary products through accumulations does not remove SSA’s input-dependent matrix operations, which our target platform does not natively support. GLA instead uses convolution and pooling primitives. We will add these comparisons and distinguish measured GPU latency, estimated synaptic energy, simulator execution, and physical-chip evidence.

Comment4: Why does MTA-SNN-10-768 use ~2× QKformer’s estimated energy for a 0.59-point accuracy gain?

Responses to Comment4: We agree that efficiency should be assessed at comparable operating points. The 768-channel model targets peak accuracy; smaller variants offer better accuracy–energy trade-offs:

  • Matched accuracy, T=1T=1: MTA-SNN-Maxpool-10-384 achieves 77.84% at 5.09 mJ versus QKformer-10-768’s 77.78% at 6.79 mJ—25.0% lower estimated energy and 73.4% fewer parameters, subject to the discussed max-pooling deployment caveats.
  • Default average pooling: MTA-SNN-10-384 achieves 76.78% at 4.92 mJ versus QKformer-10-384’s 76.48% at 5.52 mJ, with comparable parameters and T=1T=1 versus T=4T=4.

Our energy estimates include all threshold branches:

SOPsMTPN=∑lFlT∑i=1Klrl,i,SOPs_{MTPN} = \sum_l F_l T \sum_{i=1}^{K_l} r_{l,i},

where FlF_l denotes dense operations and rl,ir_{l,i} denotes branch firing rates. Appendix E’s rates sum to 0.98, implying 2.8× the computation of a single-step branch firing at 0.35. Weight sharing preserves parameter count, but parallel execution adds threshold evaluations and neuron states; physical storage and communication depend on mapping.

Thus, MTPN exchanges spatial resources for fewer temporal steps. We will emphasize benefits at selected operating points, explicit branch costs, and the distinction between timestep count and measured latency.

Comment5: The physical-chip experiments evaluate only a compact MTSNN. Could the authors explain how GLA, global pooling, and multi-threshold branches would operate asynchronously, and provide more complete chip-level latency, energy, and memory measurements?

Responses to Comment5: Thank you for this direct and useful concern. Appendix L provides complementary evidence: asynchronous simulation of MTA-SNNtiny containing both GLA and MTPN, and physical execution of a compact MTSNN validating the tied-weight multi-threshold mechanism.

MTPN. Input events are routed to shared-weight branches with heterogeneous thresholds, which emit spikes independently. Their spike streams are merged by downstream synaptic accumulation, so the branches do not need to fire simultaneously or wait for one another. The Speck2e experiment implements two tied branches with thresholds {0.5,1.0}{0.5,1.0}{0.5,1.0}.

GLA. The local pathway applies depthwise convolution to arriving events. The global pathway performs event-driven spatial accumulation using fixed pooling weights, with the pooling normalization absorbed into the synaptic mapping. The outputs of the two paths are then routed to downstream spiking neurons, where their currents are accumulated and fused. A finite observation window is used to define the global pooling result, but the two paths do not require lockstep execution or a frame-based matrix operation. Appendix L.1 implements this event-domain formulation in the asynchronous simulator (The simulator emulates SNN execution in a fully asynchronous environment and has been aligned with the physical chip, substantially reducing the complexity of deployment. We plan to open-source it in a future release. ).

Physical measurements and scope. The current Speck2e low-level operators do not support the global-pooling and path-fusion mappings required by GLA. Therefore, the physical-chip experiment evaluates only the compact MTSNN with the tied-weight multi-threshold mechanism, rather than GLA or the full MTA-SNN. It achieves 93.63% accuracy on the evaluated N-MNIST subset, with a net dynamic energy of 293.65 μ\muμJ/sample and 0.0698 μ\muμJ/input event, excluding idle consumption. Accordingly, chip-level latency, energy, and memory measurements for GLA are unavailable in the current hardware version; its asynchronous evidence is limited to the simulator. We will clarify this distinction and avoid extrapolating the compact-chip measurements to the full MTA-SNN.

Comment6: Could the authors clarify the discrepancy between 83.07% and 83.16% CIFAR-100 accuracy and report random-seed statistics? Could they also compare different threshold counts and multi-timestep baselines under matched computational and synaptic-operation budgets?

Responses to Comment6: Thank you for highlighting the reporting inconsistency and need for budget-controlled comparisons.

Accuracy reporting. Both values come from the same three independent runs: 83.07% ± 0.11% is the mean ± standard deviation; 83.16% is the best run. We apologize for the ambiguity. We will use mean ± standard deviation for headline results, label best-run values explicitly, and report seeds and individual accuracies.

Threshold counts and computational budgets. Appendix F, Table 7, fixes the remaining architecture and training protocol at T=1T=1, reporting 81.18% (K=1K=1), 81.90–82.28% (K=2K=2), 83.16% (K=3K=3), and 83.36% (K=4K=4), suggesting diminishing returns beyond three thresholds. Appendix I compares (K,T)=(3,1)(K,T)=(3,1) with (1,3)(1,3) under matched training settings, obtaining 83.16% versus 81.18%.

However, identical architectures or equal KTKT do not ensure equal computational costs. Spike-driven operations depend on measured branch firing rates:

SOPs=∑lFlT∑i=1Krl,i,SOPs = \sum_l F_l T \sum_{i=1}^{K} r_{l,i},

where FlF_l is the layer operation count and rl,ir_{l,i} is its branch firing rate. These comparisons are therefore architecture-controlled, not strictly SOP-matched. Equal-budget evaluation requires layerwise SOP measurements, separate accounting of neuron-update and other non-synaptic costs, and repeated-seed comparisons at a common budget. We will state this limitation and avoid claiming superiority under strictly matched computational costs.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: MTPN’s Contributions and GLA’s Asynchronous Execution

Dear Area Chair,

We thank you and the reviewers for the careful evaluation. We would like to clarify two central issues: MTPN’s relationship to existing multi-threshold methods and GLA’s asynchronous execution.

MTPN combines heterogeneous thresholds, independent binary-spike branches, shared convolutional weights, and parallel single-timestep execution within a fully convolutional architecture. The seven works raised by the reviewers can be grouped by their encoding mechanisms: Spiking-UNet, Stair-LIF, and Cascade-MT construct multi-level responses through multi-level firing or sequential updates; SpikeAtConv’s MBPL, MLF, and Parallel-MT employ parallel multi-threshold encoding; One-Hot Multi-Level LIF uses mutually exclusive interval coding; and I-LIF/SpikeYOLO expands integer activations through virtual timesteps during inference.

MTPN implements spatially parallel encoding through independent standard binary LIF branches and shared-weight aggregation, without sequential inter-branch residual updates, exclusive lane selection, or virtual-timestep expansion. We acknowledge its shared foundations with parallel methods such as MLF and will revise the related-work discussion accordingly. Our contribution further encompasses the organization and validation of this mechanism within a single-timestep architecture: at a fixed branch count, heterogeneous thresholds improve CIFAR-100 accuracy from 80.15% to 83.16% over identical thresholds; neuron replacement alone also improves ResNet-18 accuracy by 4.31 percentage points. Weight sharing retains one convolutional parameter set of size PP, compared with KPKP for independently parameterized branches, while additional state and event-processing costs remain.

GLA’s asynchronous execution is supported by simulation. The local pathway applies depthwise convolution to arriving events, while the global pathway incrementally accumulates spatial contributions through fixed pooling weights. Both pathways feed downstream spiking neurons, where their currents accumulate and fuse without simultaneous firing or lockstep completion. A finite observation window defines the completed global statistic but does not prevent event processing within that window. Relative to the compared QK attention, GLA replaces input-dependent matrix products with fixed-weight convolutional and pooling aggregation. Its rationale rests on this computational organization and demonstrated event-driven execution, rather than eliminating all synchronization.

Appendix L provides complementary validation: MTA-SNNtiny, containing both GLA and MTPN, executes in the asynchronous simulator, while the tied-weight multi-threshold mechanism executes physically on Speck2e, achieving 93.63% accuracy on the evaluated N-MNIST subset and 293.65 μJ/sample net dynamic energy. The current chip’s low-level operators do not support the global-pooling and path-fusion mappings required by GLA; accordingly, GLA’s execution evidence comes from simulation, and the physical measurements concern the compact MTSNN.

We respectfully ask that the contribution be assessed through these complementary results: isolated component gains, single-timestep architectural performance, asynchronous simulation, and physical implementation of the multi-threshold mechanism. Together, they support the contribution of MTA-SNN to spike representation and event-driven architecture design.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.