Research map / Architectures and efficiency / Spiking networks
TEFormer: Structured Bidirectional Temporal Enhancement Modeling in Spiking Transformers
Sicheng Shen, Mingyang Lv, Bing Han, Dongcheng Zhao, Guobin Shen, Feifei Zhao, Yi Zeng
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 |
|---|---|---|---|
| add T-MLP and TEA together (bidirectional temporal fusion) (ablation) | H-Spikformer baseline · CIFAR10 (4-384) | Acc@1 (%): 95.91 -> 96.24 | Improved |
| add T-MLP alone (ablation) | H-Spikformer baseline · CIFAR10 (4-384) | Acc@1 (%): 95.91 -> 95.98 | No clear effect |
| add TEA alone (ablation) | H-Spikformer baseline · CIFAR10 (4-384) | Acc@1 (%): 95.91 -> 95.85 | Hurt |
| assign forward fusion to TEA and backward to T-MLP instead of both backward (ablation) | TEFormer · CIFAR10 (4-384) | Acc@1 (%): 96.09 -> 96.24 | Improved |
| use single-gate T-MLP instead of GRU-style gate (ablation) | TEFormer · CIFAR10 | Acc@1 (%): 95.90 -> 96.24 | Improved |
| use single-gate T-MLP instead of EMA-style update (ablation) | TEFormer · CIFAR100 | Acc@1 (%): 79.55 -> 79.84 | Improved |
| use learnable alpha instead of fixed 0.75 (ablation) | TEA · CIFAR10 | Acc@1 (%): 96.04 -> 96.24 | Improved |
| use learnable alpha instead of fixed 0.75 (ablation) | TEA · CIFAR10-DVS | Acc@1 (%): 81.2 -> 81.9 | Improved |
| increase temporally enhanced layers from 2 to 4 (ablation) | TEFormer · CIFAR10 | Acc@1 (%): 96.04 -> 95.66 | Hurt |
| enhance 2 layers instead of 1 (ablation) | TEFormer · CIFAR10 | Acc@1 (%): 95.24 -> 96.04 | Improved |
| reduce T-MLP MLP ratio from 4 to 2 (ablation) | TEFormer · CIFAR10 | Acc@1 (%) / energy: 96.24 -> 95.98; 5.18 mJ -> 2.39 mJ | Mixed |
| add TEFormer under phase encoding vs QKFormer | TEFormer · CIFAR10 (phase encoding) | Acc@1 (%): 87.76 -> 89.92 | 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 2 of 2 shared results
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
- STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking (NeurIPS 2025 · similarity 0.90)
- Towards Efficient Spiking Transformer: a Token Sparsification Framework for Training and Inference Acceleration (ICML 2024 · similarity 0.90)
- Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer (CVPR 2025 · similarity 0.89)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.88)
- LMUFormer: Low Complexity Yet Powerful Spiking Model With Legendre Memory Units (ICLR 2024 · similarity 0.88)
- Spiking Transformer with Spatial-Temporal Attention (CVPR 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.
