Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Multiplication-Free Parallelizable Spiking Neurons with Efficient Spatio-Temporal Dynamics

Authors: Peng Xue, Wei Fang, Zhengyu Ma, Zihan Huang, Zhaokun Zhou, Yonghong Tian, Timothée Masquelier, Huihui Zhou

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our methods are validated in neuromorphic Spiking Heidelberg Digits voices, sequential CIFAR images, and neuromorphic DVS-Lip vision datasets, achieving superior performance over SOTA spiking neurons. Training speed results demonstrate the effectiveness of our acceleration methods, providing a practical reference for future research.
Researcher Affiliation Academia Peng Xue1,2,6 Wei Fang3 Zhengyu Ma1 Zihan Huang4 Zhaokun Zhou1,3 Yonghong Tian1,3,4 Timothée Masquelier5 Huihui Zhou1 1Peng Cheng Laboratory 2Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 3School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University 4School of Computer Science, Peking University 5Centre de Recherche Cerveau et Cognition (CERCO), UMR5549 CNRS Université Toulouse 3 6University of Chinese Academy of Sciences
Pseudocode Yes Algorithm 1 Autoselect acceleration algorithm
Open Source Code Yes Our code is available at Github.
Open Datasets Yes Our methods are validated in neuromorphic Spiking Heidelberg Digits voices, sequential CIFAR images, and neuromorphic DVS-Lip vision datasets, achieving superior performance over SOTA spiking neurons. ... We evaluate the long-term dependencies learning ability of mul-free channel-wise PSN in three widely used classification tasks, including the Spiking Heidelberg Digits (SHD) spoken digit dataset [37], the sequential CIFAR dataset, and the high spatial-temporal resolution automatic lip-reading DVS-Lip dataset [38].
Dataset Splits Yes Half of the words in the dataset are visually similar pairs in the LRW dataset [41] (e.g., "America" and "American"). The training and testing sets are derived from different speakers, posing a challenge for the model to exhibit robust generalization capabilities with respect to speaker characteristics.
Hardware Specification Yes The experiments are carried out on a Debian GNU/Linux 11(bullseye) server with an Intel(R) Xeon(R) Platinum 8336C CPU, an Nvidia A100-SXM4-80GB GPU and 32GB RAM.
Software Dependencies No The paper mentions Py Torch, NVIDIA CUDA Deep Neural Network (cu DNN), and Open AI Triton but does not specify version numbers for any of these software components.
Experiment Setup Yes The main hyper-parameters for different datasets are shown in Table 8. Other training options are listed as follows. Sequential CIFAR The data augmentation techniques include random mixup with p = 1 and α = 0.2, random cutmix with p = 1 and α = 1, random choice between the two mix methods with p = 0.5, random horizontal flip with p = 0.5, trivial augmentation, normalization, random erasing with p = 0.1, and label smoothing with the amount 0.1 [15]. The number of channels is 128. The surrogate function is the arctan surrogate function σ(x) = α 2(1+( π 2 αx)2) with α = 2. Pixel CIFAR All parameters and experimental settings are the same as Sequential CIFAR. SHD Augmentation methods include spatio jitter with var 0.55, uniform noise with number n = 35, drop event with p = 0.05, drop event chunk with p = 0.3 and max drop chunk length l = 0.02. The surrogate function is also the arctan surrogate function with α = 5. DVS-Lip The data augmentation techniques include center cropped size = 96 96, then random cropped size = 88 88, random horizontal flip with p = 0.5, 2D spatial mask with mask num = 4 and maximum length= 20, random choice between zoom in and zoom out with p = 0.5 and max scale = 26, temporal mask with mask num = 6 and maximum length = 18 [42]. The surrogate function is σ(x) = 1 1+αx2 with α = 10.