본문

서브메뉴

Server-Side Algorithms for Communication-Efficient Federated Learning
Server-Side Algorithms for Communication-Efficient Federated Learning
Server-Side Algorithms for Communication-Efficient Federated Learning

Detailed Information

자료유형  
 학위논문 서양
최종처리일시  
20260202104719
ISBN  
9798290938646
DDC  
621.3
저자명  
Jhunjhunwala, Divyansh.
서명/저자  
Server-Side Algorithms for Communication-Efficient Federated Learning
발행사항  
[Sl] : Carnegie Mellon University, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
333 p
주기사항  
Source: Dissertations Abstracts International, Volume: 87-02, Section: A.
주기사항  
Advisor: Joshi, Gauri.
학위논문주기  
Thesis (Ph.D.)--Carnegie Mellon University, 2025.
초록/해제  
요약Recent reports estimate that nearly 70% of daily data is generated on personal user devices such as smartphones leading to increasingly decentralized data collection and storage. Federated Learning (FL) is a framework designed to train machine learning models on decentralized data distributed across a network of clients under the coordination of a central server. While FL enables distributed and privacy-preserving training, a major bottleneck arises from the need to frequently communicate high-dimensional model updates from clients to the server, especially given the limited upload bandwidth on most devices. Additionally, data on each client is typically generated independently, resulting in non-IID (non-identically distributed) data distributions which can significantly slow FL convergence and exacerbate communication inefficiency.Prior work in FL has largely focused on sophisticated client-side optimizations to address this challenge. For example, to reduce the size of updates sent by clients, popular solutions include biased compressors with error feedback, complex quantization schemes, and knowledge distillation - most of which either increase computation on resource-constrained clients and/or require clients to maintain state. Similarly, to handle data heterogeneity, common approaches include regularizing local objectives, incorporating control variates, or using contrastive losses, all of which add overhead on the client side. In contrast, our work focuses on addressing these challenges through server-side optimization algorithms, which leave local training untouched and instead modify the aggregation strategy at the server. As discussed in this thesis, there is often additional structure across client updates that can be exploited to improve aggregation. We also uncover connections between the standard FL algorithm FedAvg and classical optimization methods, which motivate our proposed techniques. Our methods not only achieve better performance than prior client-focused approaches but are also versatile enough to integrate with existing techniques, enhancing the efficiency and applicability of FL.The thesis is divided into 3 parts. In part 1) we discuss server algorithms to tackle the challenge of compression in FL by leveraging ideas based on adaptive quantization, spatial correlation and temporal correlation. In part 2) we discuss server algorithms to deal with the challenge of heterogeneous client updates by adaptively tuning the server step-size and leveraging Fisher information. Finally, in part 3) we revisit FedAvg in the context of pre-trained models and outline a server algorithm to improve aggregation for federated LoRA fine-tuning by decomposing client updates into common and client-specific components. Together, these advances provide a lightweight yet powerful framework for scalable, communication-efficient, and heterogeneity-aware FL.
일반주제명  
Electrical engineering
일반주제명  
Information technology
일반주제명  
Communication
키워드  
Algorithms
키워드  
Compression
키워드  
Federated learning
키워드  
Heterogeneity
키워드  
Pre-trained models
기타저자  
Carnegie Mellon University Electrical and Computer Engineering
기본자료저록  
Dissertations Abstracts International. 87-02A.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017358558
■00520260202104719
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798290938646
■035    ▼a(MiAaPQ)AAI32120934
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a621.3
■1001  ▼aJhunjhunwala,  Divyansh.
■24510▼aServer-Side  Algorithms  for  Communication-Efficient  Federated  Learning
■260    ▼a[Sl]▼bCarnegie  Mellon  University▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a333  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  87-02,  Section:  A.
■500    ▼aAdvisor:  Joshi,  Gauri.
■5021  ▼aThesis  (Ph.D.)--Carnegie  Mellon  University,  2025.
■520    ▼aRecent  reports  estimate  that  nearly  70%  of  daily  data  is  generated  on  personal  user  devices  such  as  smartphones  leading  to  increasingly  decentralized  data  collection  and  storage.  Federated  Learning  (FL)  is  a  framework  designed  to  train  machine  learning  models  on  decentralized  data  distributed  across  a  network  of  clients  under  the  coordination  of  a  central  server.  While  FL  enables  distributed  and  privacy-preserving  training,  a  major  bottleneck  arises  from  the  need  to  frequently  communicate  high-dimensional  model  updates  from  clients  to  the  server,  especially  given  the  limited  upload  bandwidth  on  most  devices.  Additionally,  data  on  each  client  is  typically  generated  independently,  resulting  in  non-IID  (non-identically  distributed)  data  distributions  which  can  significantly  slow  FL  convergence  and  exacerbate  communication  inefficiency.Prior  work  in  FL  has  largely  focused  on  sophisticated  client-side  optimizations  to  address  this  challenge.  For  example,  to  reduce  the  size  of  updates  sent  by  clients,  popular  solutions  include  biased  compressors  with  error  feedback,  complex  quantization  schemes,  and  knowledge  distillation  -  most  of  which  either  increase  computation  on  resource-constrained  clients  and/or  require  clients  to  maintain  state.  Similarly,  to  handle  data  heterogeneity,  common  approaches  include  regularizing  local  objectives,  incorporating  control  variates,  or  using  contrastive  losses,  all  of  which  add  overhead  on  the  client  side.  In  contrast,  our  work  focuses  on  addressing  these  challenges  through  server-side  optimization  algorithms,  which  leave  local  training  untouched  and  instead  modify  the  aggregation  strategy  at  the  server.  As  discussed  in  this  thesis,  there  is  often  additional  structure  across  client  updates  that  can  be  exploited  to  improve  aggregation.  We  also  uncover  connections  between  the  standard  FL  algorithm  FedAvg  and  classical  optimization  methods,  which  motivate  our  proposed  techniques.  Our  methods  not  only  achieve  better  performance  than  prior  client-focused  approaches  but  are  also  versatile  enough  to  integrate  with  existing  techniques,  enhancing  the  efficiency  and  applicability  of  FL.The  thesis  is  divided  into  3  parts.  In  part  1)  we  discuss  server  algorithms  to  tackle  the  challenge  of  compression  in  FL  by  leveraging  ideas  based  on  adaptive  quantization,  spatial  correlation  and  temporal  correlation.  In  part  2)  we  discuss  server  algorithms  to  deal  with  the  challenge  of  heterogeneous  client  updates  by  adaptively  tuning  the  server  step-size  and  leveraging  Fisher  information.  Finally,  in  part  3)  we  revisit  FedAvg  in  the  context  of  pre-trained  models  and  outline  a  server  algorithm  to  improve  aggregation  for  federated  LoRA  fine-tuning  by  decomposing  client  updates  into  common  and  client-specific  components.  Together,  these  advances  provide  a  lightweight  yet  powerful  framework  for  scalable,  communication-efficient,  and  heterogeneity-aware  FL.
■590    ▼aSchool  code:  0041.
■650  4▼aElectrical  engineering
■650  4▼aInformation  technology
■650  4▼aCommunication
■653    ▼aAlgorithms
■653    ▼aCompression
■653    ▼aFederated  learning
■653    ▼aHeterogeneity
■653    ▼aPre-trained  models
■690    ▼a0800
■690    ▼a0544
■690    ▼a0489
■690    ▼a0459
■71020▼aCarnegie  Mellon  University▼bElectrical  and  Computer  Engineering.
■7730  ▼tDissertations  Abstracts  International▼g87-02A.
■790    ▼a0041
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358558▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

Preview

Export

ChatGPT Discussion

AI Recommended Related Books


    New Books MORE
    Statistics for the past 3 years. Go to brief

    Подробнее информация.

    • Бронирование
    • не существует
    • моя папка
    • Первый запрос зрения
    • Non-Book Loan Application
    • Nighttime Book Loan Application
    материал
    Reg No. Количество платежных Местоположение статус Ленд информации
    TF18316 전자도서 대출가능 My Folder 부재도서신고 비도서대출신청 야간 도서대출신청

    * Бронирование доступны в заимствований книги. Чтобы сделать предварительный заказ, пожалуйста, нажмите кнопку бронирование

    Books borrowed together with this book

    Related Popular Books

    Available after logging in.