Research map / Architectures and efficiency / Spiking networks
Efficient Transformer Attention for SNNs via Hadamard Simplification
Tingting Jiang, Jiangrong Shen, Long Chen, Yaxin Li, Qi Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace patch splitting with 3x3 convolution (ablation) | MM-Attn · CIFAR-100 | accuracy: 78.15% to 77.37% | Hurt |
| replace MM-Attn with full Hadamard attention (ablation) | MM-Attn (3x3 conv) · CIFAR-100 | accuracy: 77.37 -> 75.91 | Hurt |
| remove scale factor (ablation) | FH-Attn · CIFAR-100 | accuracy: 75.91 -> 76.63 | Improved |
| remove multi-head attention (ablation) | FH-Attn · DVS-Gesture | accuracy: 94.09 -> 95.48 | Improved |
| remove multi-head attention (ablation) | FH-Attn · CIFAR-100 | accuracy: 76.63 -> 76.21 | No clear effect |
| use single mask on KV branch instead of double mask (ablation) | SSA late-spiking · CIFAR-100 | accuracy: 76.21 -> 78.33 | Improved |
| use early-spiking instead of late-spiking gating (ablation) | SSA single KV mask · CIFAR-100 | accuracy: 78.33 -> 79.45 | Improved |
| use early-spiking instead of late-spiking gating (ablation) | SSA single KV mask · DVS-Gesture | accuracy: 96.52 -> 97.56 | Improved |
| compress K branch into channel-wise gate (USSA, 3x3 conv) (ablation) | SSA · CIFAR-100 | accuracy: 79.45 -> 78.30 | Hurt |
| use 1x1 instead of 3x3 conv for K branch (ablation) | USSA channel-wise · CIFAR-100 | accuracy: 78.30 -> 78.22 | Hurt |
| use channel-wise instead of spatial-wise reduction (ablation) | USSA 1x1 conv · CIFAR-100 | accuracy: 77.84 -> 78.22 | Improved |
| replace MDSSA with SSA | SpikingResformer · CIFAR-100 | accuracy: 78.15 -> 79.45 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Spikingformer — Spikingformer: A Key Foundation Model for Spiking Neural Networks (AAAI 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer (CVPR 2025 · similarity 0.92)
- Towards Efficient Spiking Transformer: a Token Sparsification Framework for Training and Inference Acceleration (ICML 2024 · similarity 0.89)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.89)
- Spiking Vision Transformer with Saccadic Attention (ICLR 2025 · similarity 0.89)
- Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips (ICLR 2024 · similarity 0.88)
- STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking (NeurIPS 2025 · similarity 0.88)
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.
