CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Networks

Authors: Yulong Huang, Xiaopeng Lin, Hongwei Ren, Haotian Fu, Yue Zhou, Zunchang Liu, Biao Pan, Bojun Cheng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on a variety of datasets demonstrate CLIF s clear performance advantage over other neuron models.
Researcher Affiliation Academia 1Function Hub, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China 2School of Integrated Circuit Science and Engineering, Beihang University, Beijing, China.
Pseudocode Yes Algorithm 1 Core function for CLIF model
Open Source Code Yes The code is available at https://github.com/Huu Yu Long/Complementary LIF.
Open Datasets Yes CIFAR-10/100 The CIFAR-10 and CIFAR-100 datasets (Krizhevsky et al., 2009)... Tiny-Image Net Tiny-Image Net contains 200 categories... DVSCIFAR10 The DVS-CIFAR10 dataset (Li et al., 2017)... DVSGesture The DVS128 Gesture dataset (Amir et al., 2017)...
Dataset Splits No For the CIFAR-10 and CIFAR-100 datasets... each dataset comprising 50,000 training samples and 10,000 testing samples. The paper does not explicitly mention a validation set split or provide specific percentages/counts for train/validation/test splits, nor does it refer to a standard split that includes validation.
Hardware Specification No No specific hardware details (such as GPU/CPU models or cloud instance types) used for running the experiments are mentioned in the paper.
Software Dependencies No The event-to-frame integration is handled with the Spiking Jelly (Fang et al., 2023) framework. No specific version numbers for software dependencies (e.g., Python, PyTorch, Spiking Jelly) are provided.
Experiment Setup Yes Unless otherwise specified or for the purpose of comparative experiments, the experiments in this paper adhere to the following settings and data preprocessing: all our self-implementations use Rectangle surrogate functions with α = Vth = 1, and the decay constant τ is set to 2.0. All random seed settings are 2022. For all loss functions, we use the TET (Deng et al., 2021) with a 0.05 loss lambda, as implemented in (Meng et al., 2023). Table 5. Training Parameters Dataset Optimizer Weight Dacay Batch Size Epoch Learning Rate CIFAR10 SGD 5e-5 128 200 0.1 CIFAR100 SGD 5e-4 128 200 0.1 Tiny Image Net SGD 5e-4 256 300 0.1 DVSCIFAR10 SGD 5e-4 128 300 0.05 DVSGesture SGD 5e-4 16 300 0.1