kth.sePublications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Federated Learning Over-the-Air by Retransmissions
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Network and Systems Engineering. Digital Futures.ORCID iD: 0000-0002-5761-2580
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Network and Systems Engineering. Digital Futures.ORCID iD: 0000-0002-2764-8099
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Network and Systems Engineering. Digital Futures.ORCID iD: 0000-0001-9810-3478
2023 (English)In: IEEE Transactions on Wireless Communications, ISSN 1536-1276, E-ISSN 1558-2248, Vol. 22, no 12, p. 9143-9156Article in journal (Refereed) Published
Abstract [en]

Motivated by the increasing computational capabilities of wireless devices, as well as unprecedented levels of user- and device-generated data, new distributed machine learning (ML) methods have emerged. In the wireless community, Federated Learning (FL) is of particular interest due to its communication efficiency and its ability to deal with the problem of non-IID data. FL training can be accelerated by a wireless communication method called Over-the-Air Computation (AirComp) which harnesses the interference of simultaneous uplink transmissions to efficiently aggregate model updates. However, since AirComp utilizes analog communication, it introduces inevitable estimation errors. In this paper, we study the impact of such estimation errors on the convergence of FL and propose retransmissions as a method to improve FL accuracy over resource-constrained wireless networks. First, we derive the optimal AirComp power control scheme with retransmissions over static channels. Then, we investigate the performance of Over-the-Air FL with retransmissions and find two upper bounds on the FL loss function. Numerical results demonstrate that the power control scheme offers significant reductions in mean squared error. Additionally, we provide simulation results on MNIST classification with a deep neural network that reveals significant improvements in classification accuracy for low-SNR scenarios.

Place, publisher, year, edition, pages
Institute of Electrical and Electronics Engineers (IEEE) , 2023. Vol. 22, no 12, p. 9143-9156
Keywords [en]
Federated Learning, Over-the-Air Computation, Retransmissions
National Category
Communication Systems Signal Processing Telecommunications
Research subject
Electrical Engineering
Identifiers
URN: urn:nbn:se:kth:diva-327825DOI: 10.1109/twc.2023.3268742ISI: 001128031700032Scopus ID: 2-s2.0-85159703045OAI: oai:DiVA.org:kth-327825DiVA, id: diva2:1761071
Note

QC 20230608

Available from: 2023-05-31 Created: 2023-05-31 Last updated: 2025-03-27Bibliographically approved
In thesis
1. Function Computation via Electromagnetic Superposition: Estimation Problems
Open this publication in new window or tab >>Function Computation via Electromagnetic Superposition: Estimation Problems
2024 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

In wireless communication systems, interference is considered one of the main bottlenecks. Because all devices share the same electromagnetic spectrum, communication protocols generally attempt to separate radio resources to avoid interference. In LTE and 5G, devices are not allowed to transmit at their own behest but must receive an uplink grant that dictates which radio resources to use. In Wi-Fi 5, devices are allowed to transmit without a grant from the router, but they have to listen to the channel and wait until it is quiet before transmitting. While these interference-avoiding methods are quite useful, they face the problem of congestion. If too many devices are communicating data simultaneously, the avoidance of interference leads to insufficient spectrum for each user, and quality of service drops dramatically.

In this thesis, we study a novel form of wireless communication that takes a different approach to sharing electromagnetic spectrum. Rather than using orthogonal resources for each device, it schedules devices on the same communication resources, resulting in electromagnetic superposition of their signals. When the receiver listens, the superimposed signal will contain so much interference that it is difficult to distinguish the individual messages. However, it is possible to compute functions of the transmitted messages. Hence, this method is often referred to as Over-the-Air Function Computation (AirComp).

The challenges of AirComp are fundamentally different from those of orthogonal communication. Well-known results on, e.g., phase acquisition, forward error correction, and modulation do not map directly to the AirComp setting. Because of this, the state-of-the-art literature on AirComp usually does not guarantee error-free function computation with a vanishing probability of error but resorts to imperfect function estimation. We have dedicated this thesis to improving the state of estimation algorithms for AirComp. For example, we have developed a power control scheme that eliminates estimation bias for fast-fading channels, and we have leveraged dimensionality-reduction methods to compute certain functions without error.

Our recent work has focused on incorporating realistic assumptions concerning time synchronization and phase acquisition. In orthogonal communication methods, phase alignment is often achieved by careful correction at the receiver side, using reference signals and phase-locked loops. In AirComp, since we are interested in the coherent superposition of signals, the phase cannot be corrected at the receiver side. Transmitter-side phase correction from UEs is challenging, and therefore we have developed non-coherent AirComp methods that avoid this problem. We also specify non-coherent AirComp schemes for digital communication problems in the sparse regime, outperforming orthogonal methods.

Abstract [sv]

I trådlösa kommunikationssystem är interferens en av de mest begränsande flaskhalsarna. Eftersom alla enheter delar samma elektromagnetiska spektrum så tenderar kommunikationsprotokol att separera radioresurser för att undvika interferens. I LTE och 5G får enheterna inte ta initiativ till att sända, utan måste vänta på ett tillstånd som bestämmer vilka resurser de får använda. I Wi-Fi 5 får enheterna ta initiativ, men de måste först lyssna på kanalen och vänta tills det är tyst innan de börjar sända. Dessa interferensundvikande metoder är väldigt användbara men de påverkas oundvikligen av problem med bergänsat spektrum. Om för många enheter vill kommunicera samtidigt så kommer undvikningen av interferens leda till otillräckliga resurser för varje enhet, vilket drastiskt minskar kommunikationskvaliteten.

I den här avhandlingen studerar vi en ny metod för trådlös kommunikation som har ett annat tillvägagångssätt för att samarbeta över det elektromagnetiska spektrumet. Istället för att allokera ortogonala resurser till varje enhet så tilldelas många enheter samma kommunikationsresurser, vilket leder till elektromagnetisk superposition av deras signaler. När mottagaren lyssnar på de kombinerade signalerna, så är det så pass mycket interferens att separeringen av individuella meddelanden blir utmanande. Däremot är det möjligt att beräkna funktioner av meddelanderna, vilket är varför metoden ofta kallas Over-the-Air Function Computation (AirComp) på engelska. På svenska föreslår vi att kalla det signalfogning, för att referera till hur signaler fogas till en funktion i spektrumet.

Tekniken för signalfogning är fundamentalt annorlunda från ortogonala kommunikationsmetoder. Välkända resultat inom till exempel, fastrackning, felrättande koder och modulering fungerar inte på samma sätt för signalfogning. Pågrund av detta så brukar signalfogningslitteraturen inte garantera felfri funktionsberäkning, utan arbetar istället med estimering. Den här avhandlingen är dedikerad till att utveckla bättre estimeringsalgoritmer för signalfogning. Till exempel har vi utvecklat en effektallokeringsalgoritm för att ta bort systematiska fel under snabbfädning och vi har utvecklat kompressionsmetoder för att beräkna specifika funktioner helt utan estimeringsfel.

Våra senaste bidrag till litteraturen har fokuserat på att integrera realistiska antaganden gällande tidssynkronisering och fastrackning. Inom ortogonala kommunikationssystem så sker fastrackning ofta genom noggrann korrektion på mottagarsidan, där referenssignaler från sändaren nyttjas i regleralgoritmer för att styra mottagaroscillatorn. Inom signalfogning kan inte fasen korrigeras på mottagarsidan eftersom vi är intresserade av fasriktig superposition i det elektromagnetiska spektrumet. Samtidigt är faskorrigering på sändarsidan utmanande, och därför har vi arbetat med att utveckla signalfogningsmetoder som inte är beroende av någon faskorrigering. Vi specificerar också sådana metoder för digitala kommunikationsproblem med glesa signaler, vilket överträffar otrogonala metoder.

Place, publisher, year, edition, pages
Stockholm: Kungliga Tekniska högskolan, 2024. p. 63
Series
TRITA-EECS-AVL ; 2024:78
Keywords
Wireless Communications, Over-the-Air Computation, Compressed Sensing, Non-Coherent, Machine Learning, Histogram Estimation, Federated Learning
National Category
Telecommunications
Research subject
Electrical Engineering
Identifiers
urn:nbn:se:kth:diva-354729 (URN)978-91-8106-074-4 (ISBN)
Public defence
2024-11-04, https://kth-se.zoom.us/j/65644192644, Kollegiesalen, Brinellvägen 6, Stockholm, 13:00 (English)
Opponent
Supervisors
Note

QC 20241014

Available from: 2024-10-14 Created: 2024-10-11 Last updated: 2024-10-21Bibliographically approved

Open Access in DiVA

fulltext(439 kB)103 downloads
File information
File name FULLTEXT01.pdfFile size 439 kBChecksum SHA-512
72f379a41060974019e1c269810cf88ba65d7a4d904833184cef070de80be22e7a0747d36a7cb45c96eb1f5499802e83fe1e66182e318141e9f6fe879d48504f
Type fulltextMimetype application/pdf

Other links

Publisher's full textScopus

Authority records

Hellström, HenrikFodor, ViktóriaFischione, Carlo

Search in DiVA

By author/editor
Hellström, HenrikFodor, ViktóriaFischione, Carlo
By organisation
Network and Systems Engineering
In the same journal
IEEE Transactions on Wireless Communications
Communication SystemsSignal ProcessingTelecommunications

Search outside of DiVA

GoogleGoogle Scholar
Total: 103 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

doi
urn-nbn

Altmetric score

doi
urn-nbn
Total: 257 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf