Efficient and Effective Time-Series Forecasting with Spiking Neural Networks

Authors: Changze Lv, Yansen Wang, Dongqi Han, Xiaoqing Zheng, Xuanjing Huang, Dongsheng Li

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct experiments to investigate the following research questions:
Researcher Affiliation Collaboration The work was conducted during the internship of Changze Lv (czlv22@m.fudan.edu.cn) at Microsoft Research Asia. 1School of Computer Science, Fudan University, Shanghai, China 2Microsoft Research Asia, Shanghai, China.
Pseudocode No The paper describes the methodology in prose and mathematical equations but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https: //github.com/microsoft/Seq SNN.
Open Datasets Yes Metr-la (Li et al., 2017b); Pems-bay (Li et al., 2017b); Electricity (Lai et al., 2018); Solar (Lai et al., 2018)
Dataset Splits Yes We partitioned the forecasting datasets into train, validation, and test sets following a chronological order. The statistical characteristics and specific split details can be found in Table 4.
Hardware Specification Yes We run our experiments on 4 NVIDIA RTX A6000 GPUs.
Software Dependencies No The paper mentions ‘Snn Torch’ and ‘Spiking Jelly’ as Pytorch-based frameworks, and ‘Adam’ optimizer, but does not provide specific version numbers for any of these software dependencies.
Experiment Setup Yes To construct our proposed SNNs, we use two Pytorch-based frameworks: Snn Torch (Eshraghian et al., 2021) and Spiking Jelly (Fang et al., 2020b). For all SNNs, we set the time step Ts = 4. For all LIF neurons in SNNs, we set threshold Uthr = 1.0, decay rate β = 0.99, α = 2 in surrogate gradient function. ... we set the batch size as 128 and adopt Adam (Kingma & Ba, 2014) optimizer with a learning rate of 1 10 4. We adopt an early stopping strategy with 30 epochs tolerance.