- 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
-
The paper makes good latency-performance trade-off. It achieves high accuracy at T=1. This is highly suitable for asynchronous neuromorphic hardware.
-
The decomposition of “attention-like” behavior into depthwise convolutions and global pooling is neuromorphic-friendly and avoids multi-path spike matrix multiplications.
-
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
-
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.
-
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.
-
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.
-
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.
-
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
-
Does MTPN also improve dense prediction tasks such as detection and segmentation?
-
For GLA, did you compare it with classic CNN attention modules under the same spiking constraints?
-
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?
-
Can the authors provide a clearer cost breakdown for MTPN, including SOPs, event count, memory access, routing overhead, and latency?
-
How are the thresholds of MTPN chosen, and can the authors provide guidelines for threshold selection across different datasets and tasks?
4 comments
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 component | Single-threshold baseline | MTPN |
|---|---|---|
| Threshold branches | 1 | 4 |
| Timesteps | 4 | 1 |
| Expected spikes per neuron group per inference | 1.40 | 1.13 |
| Normalized SOP estimate | 1.000 | 0.807 |
| Estimated energy (mJ/image) | 7.16 | 4.92 |
| Selected memory traffic¹ | 32 B: membrane reads/writes | 4 B: byte-aligned branch-map reads |
| Learned weight sets per layer | 1 | 1, shared |
¹ Selected traffic terms modeled in Appendix E.3, not measured total memory traffic.
SOPs and events. All branches are counted:
where is the dense synaptic-operation count and the branch firing rate. The measured rates sum to , versus , 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 . 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.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 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 , , and in Table 5 were directly measured. Its ImageNet entries include adjustment and extrapolation: specifically, and 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 , 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 .
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.
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.
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.
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).
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.