Research map / Architectures and efficiency / Spiking networks
TTFSFormer: A TTFS-based Lossless Conversion of Spiking Transformer
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 TTFSFormer, a generalized TTFS neuron with input/output transform kernels and zero reference time that enables lossless ANN-to-SNN conversion of Vision Transformers (ViT, EVA) with TTFS coding on ImageNet-1K.
Problem
ANN-to-SNN conversion of Transformers has relied on rate coding, which needs many spikes; prior TTFS conversion methods were limited to ReLU MLPs/CNNs because TTFS spike times represent only [0,1] and cannot handle GELU/SiLU, softmax, LayerNorm, or multiplication.
Approach
A generalized TTFS neuron with configurable input transform kernel, output transform kernel, and zero reference time expands the representable range and lets neurons compute nonlinear functions (GELU/SiLU, log-sum-exp for softmax, LayerNorm, multiplication via log-domain tricks) exactly, enabling lossless conversion of all Transformer layers.
Key results
- Converted ViT-L/16 reaches 85.78% top-1 vs 85.83% ANN, and EVA02-L reaches 90.03% vs 90.05%, with accuracy loss below 0.1%.
- Estimated energy is about 20% of the original ANN (e.g., 4.9 mJ for ViT-S/16 at 22.3%).
- Without fine-tuning, accuracy matches the ANN at time precision 1024 or higher; accuracy declines at lower precision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Convert the pretrained Transformer to an SNN using TTFS coding and generalized neurons | TTFSFormer · ImageNet-1K | top-1 accuracy | No clear effect |
| Use TTFS coding for Transformer conversion instead of the compared rate-coding conversions | Transformer SNN conversion · ImageNet-1K | top-1 accuracy | Helped |
| Use TTFS coding for Transformer conversion | ViT-S/16 · ImageNet-1K | energy consumption ratio: 22.3% | Helped |
| Use TTFS coding for Transformer conversion | ViT-B/16 · ImageNet-1K | energy consumption ratio: 21.0% | Helped |
| Use TTFS coding for Transformer conversion | ViT-L/16 · ImageNet-1K | energy consumption ratio: 20.7% | Helped |
Limitations
- Only ANN-to-SNN conversion; no TTFS training framework for spiking transformers is proposed.
- Requires high time precision (about 1024 or more) to match ANN accuracy; accuracy degrades at lower precision.
- Energy is an analytical estimate from operation counts rather than measured on neuromorphic hardware.
Experimental setup
Datasets: ImageNet-1K
Benchmarks: ImageNet-1K
Metrics: top-1 accuracy energy consumption (mJ) accuracy vs time precision
Baselines: STA ECMT SpikeZIP-TF MST AdaFire Lossless TTFS SRP Spikingformer Spike-Driven E-Spikeformer
Builds on: Stanojevic et al. lossless TTFS conversion SpikeZIP-TF ECMT STA
Closest papers by meaning
- SpikeZIP-TF: Conversion is All You Need for Transformer-based SNN (ICML 2024 · similarity 0.90)
- Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer (CVPR 2025 · similarity 0.89)
- Differential Coding for Training-Free ANN-to-SNN Conversion (ICML 2025 · similarity 0.89)
- SpikedAttention: Training-Free and Fully Spike-Driven Transformer-to-SNN Conversion with Winner-Oriented Spike Shift for Softmax Operation (NeurIPS 2024 · similarity 0.88)
- Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips (ICLR 2024 · similarity 0.88)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · 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.