基于脉冲神经网络的无线空中联邦学习

Over-the-Air Computation for Federated Learning via Spiking Neural Networks

联邦学习可以在保护数据隐私的同时,快速地从大量分布式数据中提炼智能模型,已经成为实现边缘人工智能的主流解决方案。然而,现有的联邦学习工作聚焦于在无线网络边缘部署传统的深度神经网络(如卷积神经网络等),给移动设备带来了巨大的计算负载和能量消耗。因此,提出将一种新的低消耗神经网络——脉冲神经网络,应用于联邦边缘学习中。相较于传统的深度神经网络,它训练所需的计算量和能量消耗更低。同时,为了减少通信开销,在每一轮的联邦学习训练中,提出利用空中计算技术来聚合所有局部模型的参数。整个问题是一个二次约束二次规划问题,为解决该问题,提出了一种基于分枝定界算法的算法。通过在CIFAR10数据集上的大量实验表明,该算法优于现有方法,如半正定松弛等。

Federated learning has become a mainstream solution for enabling edge AI due to its ability not only to protect data privacy,but also to rapidly refine intelligent models from massive amounts of distributed data.However,existing work on federated learning focuses on deploying traditional deep neural networks,such as convolutional neural networks,at the edge of wireless networks,which imposes huge computing load and energy consumption on mobile devices.To address this issue,a new low-cost neural network,the SNN,is applied in federated edge learning.Compared with traditional deep neural networks,SNN training requires lower computation and energy consumption.Meanwhile,in order to reduce the communication overhead,in each round of federated learning training,we propose to aggregate the parameters of all local models by using an over-the-air computation technique.The proposed problem is a quadratic constrained quadratic programming problem.To this end,an algorithm based on branch and bound is proposed.Extensive experimental results on the CIFAR10 dataset demonstrate that our algorithm outperforms existing methods,such as positive semi-definite relaxation.