Research map / Architectures and efficiency / Spiking networks

Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips

ICLR 2024 · Architectures and efficiency › Spiking networks · approach: Spiking transformers · 19 citations (OpenAlex)

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.

Proposes Meta-SpikeFormer, a meta Transformer-based spiking neural network with Conv-based early stages, spike-driven self-attention (SDSA) variants, and membrane shortcuts, reaching 80.0% top-1 on ImageNet-1K and SOTA SNN results on detection and segmentation.

Problem

Transformer-based SNNs are not truly spike-driven and have not outperformed CNN-based SNNs, and no SNN backbone had handled classification, detection and segmentation in a unified, directly trained way.

Approach

Extends Spike-driven Transformer into a meta architecture with Conv-based SNN blocks in early stages, RepConv-generated spiking Q/K/V, new linear-complexity SDSA-2/3/4 operators using only sparse addition, and membrane shortcuts for spike-driven identity mapping.

Key results

• Meta-SpikeFormer (55.4M) reaches 80.0% top-1 on ImageNet-1K with T=4 using distillation, versus 76.3% for Spike-driven Transformer (66.3M) and 75.3% for MS-Res-SNN.
• On COCO with Mask R-CNN, Meta-SpikeFormer (75.0M, T=1) reaches 51.2 mAP@0.5 versus 50.1 for EMS-Res-SNN (14.6M).
• On ADE20K with Semantic FPN, the 16.5M Meta-SpikeFormer (T=1) reaches 32.3% mIoU at 22.1 mJ, versus ResNet-18 at 32.9% and 147.1 mJ.

Measured effects reported in the paper

Surprising findings

• Removing the SepConv token mixer in Conv-based blocks cut power by 29.5% while losing only 0.3% accuracy.
• Dropping the K_S branch (SDSA-2) costs only 0.4% accuracy versus SDSA-1, and SDSA-4 matches SDSA-3.

Limitations

• Only vision tasks evaluated; language tasks left for future work.
• Multi-timestep training increases training cost, and detection/segmentation rely on ImageNet pretraining and fine-tuning.
• Energy numbers are theoretical estimates rather than measured on neuromorphic hardware.

Experimental setup

Datasets: ImageNet-1K HAR-DVS COCO ADE20K VOC2012

Benchmarks: ImageNet-1K HAR-DVS COCO val2017 ADE20K VOC2012

Metrics: top-1 accuracy mAP@0.5 mIoU parameters (M) estimated energy (mJ)

Baselines: Spike-driven Transformer Spikformer MS-Res-SNN Att-MS-Res-SNN SEW-Res-SNN EMS-Res-SNN Spiking-Yolo Spike Calibration ViT RSB-CNN-152 ResNet-18 PVT-Tiny

Builds on: Spike-driven Transformer CAFormer MetaFormer MS-ResNet Spikformer

Closest papers by meaning

• Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer (CVPR 2025 · similarity 0.91)
• Sorbet: A Neuromorphic Hardware-Compatible Transformer-Based Spiking Language Model (ICML 2025 · similarity 0.90)
• SpikedAttention: Training-Free and Fully Spike-Driven Transformer-to-SNN Conversion with Winner-Oriented Spike Shift for Softmax Operation (NeurIPS 2024 · similarity 0.90)
• Spike2Former: Efficient Spiking Transformer for High-performance Image Segmentation (AAAI 2025 · similarity 0.89)
• Multiplication-Free Parallelizable Spiking Neurons with Efficient Spatio-Temporal Dynamics (NeurIPS 2025 · similarity 0.89)
• Quantized Spike-driven Transformer (ICLR 2025 · similarity 0.89)