Research data for paper: Efficient Event-based Delay Learning in Spiking Neural Networks

The data in this repository accompanies the paper 'Efficient Event-based Delay Learning in Spiking Neural Networks'
The data relates to 4 benchmarks:
Spiking Heidelberg Digits (SHD).Spiking Speech Commands, derived from Google Speech Commands.Yin-Yang dataset.Braille letter reading dataset.The data was generated and analysed with the code available on GitHub at https://github.com/mbalazs98/deventprop/
results.py contains all test accuracies shown in figures 4-7. The other 3 files contain the trained models that achieved the corresponding accuracies. Each zip file contains directories for each architecture:
for YY hidden layers of sizes, 5, 10, 15, 20, 25, 30 feedforward architecures with and without delays

for SHD, SSC and Braille: hidden layers of sizes 64, 128, 256, 512 and 1024, feedforward and recurrent architectures with and without delaysEach architecture directory contains 8 subdirectories for the 8 random seeds. For each connection in a given architecture the synaptic weights are in the files ending in -g.npy and if relevant, delays are provided in the architectures ending in -d.npy
To test the networks run test.py from the GitHub repository provided above, with "arguments" set to the appropriate value. To load the networks through mlGeNN, the package needs to be installed (a description of how to do this is provided at https://github.com/genn-team/ml_genn). Also see attached mlGeNN_readme.txt
Abstract:Spiking Neural Networks compute using sparse communication and are attracting increased attention as a more energy-efficient alternative to traditional Artificial Neural Networks. While standard Artificial Neural Networks are stateless, spiking neurons are stateful and hence intrinsically recurrent, making them well-suited for spatio-temporal tasks. However, the duration of this intrinsic memory is limited by synaptic and membrane time constants. Delays are a powerful additional mechanism and, in this paper, we propose an event-based training method for Spiking Neural Networks with delays, grounded in the EventProp formalism, which enables the calculation of exact gradients with respect to weights and delays. Our method supports multiple spikes per neuron and introduces a delay learning algorithm that can, in contrast to previous methods, also be applied to recurrent Spiking Neural Networks. We evaluate our method on a simple sequence detection task, as well as the Yin-Yang, Spiking Heidelberg Digits, Spiking Speech Commands and Braille letter reading datasets, demonstrating that our algorithm can optimise delays from suboptimal initial conditions and enhance classification accuracy compared to architectures without delays. We also find that recurrent delays are particularly beneficial in small networks. Finally, we show that our approach uses less than half the memory of the current state-of-the-art delay-learning method and is up to 26x faster.

本仓库配套的论文为《Efficient Event-based Delay Learning in Spiking Neural Networks》。本数据集涉及4项基准任务：脉冲海德堡数字数据集（Spiking Heidelberg Digits, SHD）、源自谷歌语音命令（Google Speech Commands）的脉冲语音命令数据集（Spiking Speech Commands）、阴阳数据集（Yin-Yang dataset）以及盲文字母识别数据集（Braille letter reading dataset）。本数据集的生成与分析代码已开源至GitHub仓库：https://github.com/mbalazs98/deventprop/。

results.py文件包含了图4至图7中展示的全部测试准确率。其余3个文件存储了取得对应准确率的训练好的模型。每个压缩包均包含对应架构的目录：针对阴阳数据集，包含隐藏层维度分别为5、10、15、20、25、30的前馈架构（带延迟与无延迟两种版本）的目录；针对SHD、脉冲语音命令（Spiking Speech Commands，简称SSC）与盲文字母识别数据集，包含隐藏层维度分别为64、128、256、512及1024的前馈与循环架构（带延迟与无延迟两种版本）的目录。每个架构目录下均包含8个子目录，对应8个随机种子的实验设置。对于指定架构中的每一组连接，其突触权重存储在后缀为-g.npy的文件中；若架构支持延迟，则延迟参数存储在后缀为-d.npy的文件中。

若需测试网络，请运行上述GitHub仓库中的test.py文件，并将"arguments"参数设置为对应取值。若需通过mlGeNN加载网络，需先安装该工具包（安装说明详见https://github.com/genn-team/ml_genn），同时可参阅附件中的mlGeNN_readme.txt文档。

摘要：
脉冲神经网络（Spiking Neural Networks, SNNs）采用稀疏通信机制进行计算，作为传统人工神经网络（Artificial Neural Networks, ANNs）的更节能替代方案，正受到越来越多的关注。尽管标准人工神经网络具备无状态特性，但脉冲神经元具备状态性，因此天生具备循环特性，使其非常适用于时空任务。然而，这种内在记忆的时长受限于突触与膜时间常数。延迟是一种极具潜力的额外机制，本文基于EventProp形式体系，提出了一种面向带延迟脉冲神经网络的基于事件的训练方法，该方法可实现关于权重与延迟的精确梯度计算。我们的方法支持神经元单次发射多脉冲，并提出了一种延迟学习算法；与此前方法不同，该算法可同样适用于循环脉冲神经网络。我们在一项简单序列检测任务以及阴阳数据集、脉冲海德堡数字数据集、脉冲语音命令数据集和盲文字母识别数据集上对所提方法进行了评估，结果表明：相较于无延迟架构，我们的算法可从次优初始条件出发优化延迟参数，并提升分类准确率。我们还发现，循环延迟在小型网络中优势尤为显著。最后，我们证明了本文所提方法的内存占用仅为当前主流延迟学习方法的一半以下，且训练速度最高可提升26倍。