Spiking Transformer Evaluation Platform (STEP)
收藏资源简介:
STEP是一个统一的基准框架,用于评估Spiking Transformers,支持广泛的任务,包括分类、分割和检测,适用于静态、基于事件和序列数据集。它提供了模块化支持,包括脉冲神经元、输入编码、代理梯度和多个后端。通过STEP,我们重现和评估了几个代表性的模型,并对注意力设计、神经元类型、编码方案和时序建模能力进行了系统性的消融研究。我们提出了一个统一的能量估计分析模型,考虑了脉冲稀疏性、位宽和内存访问,并表明量化的ANNs可能提供可比或更好的能量效率。我们的结果表明,当前的Spiking Transformers严重依赖于卷积前端,缺乏强大的时序建模,这突出了对脉冲本机架构创新的需求。
STEP is a unified benchmark framework for evaluating Spiking Transformers, which supports a wide range of tasks including classification, segmentation, and detection, and is applicable to static, event-based, and sequential datasets. It offers modular support covering spiking neurons, input encoding, surrogate gradients, and multiple backends. Using STEP, we reproduced and evaluated several representative models, and conducted systematic ablation studies on attention design, neuron types, encoding schemes, and temporal modeling capabilities. We propose a unified energy estimation analytical model that accounts for spiking sparsity, bit width, and memory access, and demonstrates that quantized ANNs can achieve comparable or better energy efficiency. Our results show that current Spiking Transformers heavily rely on convolutional frontends and lack robust temporal modeling, which highlights the need for spiking-native architectural innovations.

- 1STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking中国科学院自动化研究所脑认知实验室 · 2025年



