TTFSFormer: A TTFS-based Lossless Conversion of Spiking Transformer
Authors: Lusen Zhao, Zihan Huang, Jianhao Ding, Zhaofei Yu
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on different models demonstrate that our proposed method can achieve high accuracy with significantly lower energy consumption. We evaluate our method on various pre-trained Transformer models, including Vi T and EVA, using the Image Net-1K dataset. Experimental results demonstrate that our approach achieves performance comparable to ANN counterparts. |
| Researcher Affiliation | Academia | 1 Peking University, China. Correspondence to: Zhaofei Yu <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Converting ANN into TTFS-based SNN |
| Open Source Code | Yes | The source code of the proposed method is available at https://github.com/ Forest On The Land/TTFSFormer.git. |
| Open Datasets | Yes | In this section, we evaluate our TTFS-based converted SNN methods on the Image Net-1k dataset (Deng et al., 2009) |
| Dataset Splits | No | The paper mentions using the ImageNet-1k dataset but does not explicitly describe any specific training, validation, or test splits, nor does it refer to standard splits with specific percentages or counts. |
| Hardware Specification | No | The paper estimates energy consumption and discusses hardware implementation limitations regarding time precision but does not specify any concrete hardware components (e.g., GPU models, CPU types) used for the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the implementation. |
| Experiment Setup | Yes | The whole conversion process is shown in Algorithm 1. We ll discuss some details in this part. A.1. Setting the Constants Since we re using adjustable parameters τ and Tref, we can set the [a, b] such that nearly all outputs lie within the range. More specifically, if the output range is [a, b], we can set bτ = Tref Temit, aτ = Tref Tend, (29) which indicates that τ = δ b a and Tref = Temit + bτ. |