Research map / Architectures and efficiency / Spiking networks
Unveiling the Spatial-temporal Effective Receptive Fields of Spiking Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces a spatio-temporal effective receptive field (ST-ERF) analysis for Transformer-based SNNs, finds weak global ERF in existing models, and proposes MLP-based channel mixers (MLPixer, SRB) that improve Meta-SDT on COCO detection and ADE20K segmentation.
Problem
Spiking neural networks lag behind ANNs on visual long-sequence modeling tasks (dense prediction), and existing Transformer-based SNNs fail to form a global effective receptive field across timesteps.
Approach
Extends the gradient-based ERF to the temporal dimension (ST-ERF) via loss-derived gradients, and replaces convolutional channel mixers in early stages with pixel-wise MLP-based mixers (MLPixer) and a hybrid conv-MLP mixer (SRB) to reduce locality bias and enlarge the global ERF.
Key results
- SRB (eps=4) on Meta-SDT-Base achieves 48.9% AP^b_50 on COCO val2017 Mask R-CNN (vs 46.9% for SDTv3-B), with 37M params vs 39M.
- SRB on SDTv3-B reaches 43.7% mIoU on ADE20K (+2.6 over SDTv3-B's 41.1%) while reducing parameters by 1.2M.
- On SDTv3-T, SRB improves ADE20K mIoU from 34.9% to 38.2% (+3.3) with 0.3M fewer parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the channel mixer with MLPixer ε4 (ablation) | Meta-SDT-Tiny · COCO 2017 val2017 | APb50: 35.5 -> 37.0 | Helped |
| replace the channel mixer with MLPixer ε6 (ablation) | Meta-SDT-Tiny · COCO 2017 val2017 | APb50: 35.5 -> 38.5 | Helped |
| replace the channel mixer with SRB ε4 (ablation) | Meta-SDT-Tiny · COCO 2017 val2017 | APb50: 35.5 -> 39.2 | Helped |
| replace the channel mixer with MLPixer ε4 (ablation) | Meta-SDT-Base · COCO 2017 val2017 | APb50: 46.9 -> 47.6 | Helped |
| replace the channel mixer with MLPixer ε6 (ablation) | Meta-SDT-Base · COCO 2017 val2017 | APb50: 46.9 -> 48.8 | Helped |
| replace the channel mixer with SRB ε4 (ablation) | Meta-SDT-Base · COCO 2017 val2017 | APb50: 46.9 -> 48.9 | Helped |
| replace the channel mixer with MLPixer ε4 (ablation) | Meta-SDT-Tiny · ADE20K | mIoU: 34.9 -> 34.9 | No clear effect |
| replace the channel mixer with MLPixer ε6 (ablation) | Meta-SDT-Tiny · ADE20K | mIoU: 34.9 -> 35.9 | Helped |
| replace the channel mixer with SRB ε4 (ablation) | Meta-SDT-Tiny · ADE20K | mIoU: 34.9 -> 38.2; 3.3% | Helped |
| replace the channel mixer with MLPixer ε4 (ablation) | Meta-SDT-Base · ADE20K | mIoU: 41.1 -> 42.0 | Helped |
| replace the channel mixer with MLPixer ε6 (ablation) | Meta-SDT-Base · ADE20K | mIoU: 41.1 -> 43.4 | Helped |
| replace the channel mixer with SRB ε4 (ablation) | Meta-SDT-Base · ADE20K | mIoU: 41.1 -> 43.7; 2.6% | Helped |
Surprising findings
- Spikformer exhibits diffuse receptive fields across all stages, and SDT-V1 loses its spatial ERF in the final stage.
- Replacing convolutional channel mixers with MLPs in early stages yields both accuracy and parameter gains rather than a trade-off.
Limitations
- Does not study how neuronal dynamics parameters (e.g., decay, threshold) affect ST-ERF.
- Event-based tracking gains are mixed; several metrics do not improve over SDTrack (e.g., VisEvent AUC drops for MLPixer and SRB).
- ERF analysis relies on random inputs and gradient stimuli averaged over 60 iterations, not on trained-model statistics.
Experimental setup
Datasets: ImageNet-1K COCO 2017 ADE20K FE108 VisEvent
Benchmarks: COCO 2017 ADE20K FE108 VisEvent
Metrics: AP^b AP^b_50 AP^m mIoU AUC PR
Baselines: SDTv3-T SDTv3-B SDT-V1 Spikformer QKFormer CAFormer-s18 SDTrack
Builds on: Effective Receptive Field Meta-SDT CAFormer Spikformer MLP-Mixer
Closest papers by meaning
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.89)
- FSTA-SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks (AAAI 2025 · similarity 0.89)
- SpikF: Spiking Fourier Network for Efficient Long-term Prediction (ICML 2025 · similarity 0.88)
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.87)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.87)
- Toward Relative Positional Encoding in Spiking Transformers (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.