Over-the-Air Split Machine Learning in Wireless MIMO Systems

Aug 5, 2026

Upload by Aurora

Split machine learning divides a neural network across multiple nodes, but frequent inter-node communication can heavily burden wireless links. Over-the-air computing enables simultaneous computation aligned with channel communication, offering a path to reduce this cost. This paper maps split ML onto wireless multiple-input multiple-output (MIMO) systems by decomposing inter-layer connections into transmitter precoding and receiver combining operations. Trained precoding/combining parameters learn while MIMO channels remain implicit, and channel reciprocity removes explicit channel estimation. Extensions to convolutional networks show strong performance under static and quasi-static memory channels via simulations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages31

Reading Time1 h 18 min

AccessFree to download

Abstract

Introduction

Motivation

Over-the-air computation for split ML

MIMO systems and channel computation

Interplay between MIMO OAC and neural networks

Over-the-Air Split Machine Learning in Wireless MIMO Systems Yuzhi Yang, Zhaoyang Zhang, Yuqing Tian, Zhaohui Yang, Chongwen Huang, Caijun Zhong, and Kai-Kit Wong Abstract In split machine learning (ML), different partitions of a neural network (NN) are executed by different computing nodes, requiring a large amount of communication among them. On the other hand, over-the-air computing (OAC) can efﬁciently implement all or part of the computation at the same time of communication. In this paper, we propose to deploy split ML in a wireless multiple-input multipleoutput (MIMO) communication system by exploiting the MIMO-based OAC capability. In particular, we show that the inter-layer connection in a NN of any size can be mathematically decomposed into a set of linear precoding and combining transformations over the MIMO channels. Therefore, the precoding matrix at the transmitter and the combining matrix at the receiver of each MIMO link, as well as the channel matrix itself, can jointly serve as a fully connected layer of the NN. In such a split ML system, the precoding and combining matrices are regarded as the parameters to be trained, while the MIMO channel matrix is regarded as unknown (implicit) parameters. By exploiting the channel reciprocity between the transmitter and the receiver and properly casting the backward propagation (BP) process through the MIMO channel, the explicit channel estimation process is eliminated, thus greatly saving the system costs and/or further improving its overall efﬁciency. Finally, we extend the proposed scheme to the widely used convolutional neural networks and demonstrate its effectiveness under both the static and quasi-static memory channel conditions with comprehensive simulations. This work was supported in part by National Key R&D Program of China under Grant 2020YFB1807101 and 2018YFB1801104, and National Natural Science Foundation of China under Grant U20A20158, 61725104 and 61631003 . Y. Yang, Z. Zhang (Corresponding Author), Y. Tian, C. Zhong and C. Huang are with the College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310007, China, and with the International Joint Innovation Center, Zhejiang University, Haining 314400, China, and also with Zhejiang Provincial Key Lab of Information Processing, Communication and Networking (IPCAN), Hangzhou 310007, China. (e-mails: fyuzhi yang, ning ming, tianyq, caijunzhong, [chongwenhuang](chongwenhuangg@zju.edu.cn)[g](chongwenhuangg@zju.edu.cn)[@zju.edu.cn](chongwenhuangg@zju.edu.cn)) Z. Yang and K. Wong are with the Department of Electronic and Electrical Engineering, University College London, WC1E 6BT London, UK. (emails: fzhaohui.yang, [kai-kit.wong](kai-kit.wongg@ucl.ac.uk)[g](kai-kit.wongg@ucl.ac.uk)[@ucl.ac.uk](kai-kit.wongg@ucl.ac.uk)) Index Terms Over-the-air computing (OAC), multiple-input multiple-output (MIMO), split machine learning, neural network I. INTRODUCTION A. Motivation In future sixth-generation (6G) wireless communication systems, human-like intelligence will be brought everywhere in networking systems [1] . The rapid development of artiﬁcial intelligence leads to booming mobile machine learning (ML) applications and requires vast data transmission. Split ML is a common method to distribute a NN to several devices to ease the computation burden on each device. Each device calculates the intermediate results in a split ML system and transmits it to the next device in order. Each device passes the intermediate gradient in backward order when casting backward propagation. The system can be applied in cloud networks, netof-things, or even the deep learning-based joint source-channel coding (JSCC) problem [2] . However, split ML requires frequent information exchange among edge devices, which increases the communication burden for wireless communications. For the speciﬁc structure of the split ML system, we can reduce the communication overhead cost by coupling c

FAQs

How does split machine learning increase communication requirements in wireless networks?

Split machine learning partitions a neural network across devices, requiring frequent exchange of intermediate results and gradients between nodes. This increases the communication burden for wireless links.

What is the key idea behind deploying split ML with over-the-air computing in MIMO systems?

The method exploits MIMO-based over-the-air computing capability by decomposing NN inter-layer connections into linear precoding and combining transformations over MIMO channels, enabling precoding/combining to act as trainable NN parameters.

How does the proposed approach avoid explicit channel estimation during training?

By leveraging channel reciprocity and properly casting the backward propagation through the MIMO channel, the scheme eliminates the explicit channel estimation process, reducing system cost and improving overall efficiency.

Aug 5, 2026

Upload by Aurora