Advancing Training Efficiency of Deep Spiking Neural Networks through Rate-based Backpropagation

Authors: Chengting Yu, Lei Liu, Gaoang Wang, Erping Li, Aili Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments on CIFAR-10, CIFAR-100, Image Net, and CIFAR10-DVS validate that our method achieves comparable performance to BPTT counterparts, and surpasses state-of-the-art efficient training techniques.
Researcher Affiliation Academia 1 College of Information Science and Electronic Engineering, Zhejiang University 2 ZJU-UIUC Institute, Zhejiang University
Pseudocode Yes The pseudocode for rate-based backpropagation, illustrating the implementations for both rate M and rate S, is provided in Algorithm 1. Algorithm 1: Single Training Iteration of the Rate-based Backpropagation
Open Source Code Yes Our code is available at https://github.com/Tab-ct/rate-based-backpropagation.
Open Datasets Yes In this section, we conduct experiments on CIFAR-10 [37], CIFAR-100 [37], Image Net [11], and CIFAR10-DVS [41] to evaluate the proposed training method.
Dataset Splits Yes CIFAR-10 includes 60,000 images across 10 classes, with 50,000 for training and 10,000 for testing, whereas CIFAR-100 is spread over 100 classes. [...] The Image Net-1K dataset [11] comprises 1,281,167 training images and 50,000 validation images distributed across 1,000 classes
Hardware Specification Yes The experiments on CIFAR-10, CIFAR-100, and CIFAR10-DVS datasets run on one NVIDIA Ge Force RTX 3090 GPU. For Image Net, distributed data parallel processing is utilized across eight NVIDIA Ge Force RTX 4090 GPUs.
Software Dependencies No We implement SNNs training on the Pytorch [53] and Spiking Jelly [19] frameworks. (No version numbers provided for PyTorch or Spiking Jelly.)
Experiment Setup Yes We set Vth = 1, λ = 0.2, and employ the sigmoid-based surrogate function [19] for LIF neurons. Detailed setups are provided in Appendix C. [...] Table 3: Training hyperparameters. CIFAR-10/CIFAR-100/Image Net/CIFAR10-DVS: Epoch 300/300/100/300, Learning rate 0.1/0.1/0.2/0.1, Batch size 128/128/512/128, Weight decay 5e-4/5e-4/2e-5/5e-4.