Skip to main navigation Skip to search Skip to main content

RECALL: Towards Generalized Representations in Unsupervised Federated Learning Under Non-IID Conditions

  • National Cheng Kung University

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

This study investigates the impact of Non-IID data in Federated Learning (FL), focusing on unsupervised tasks related to reconstruction, such as image restoration, image denoising, and anomaly detection. Contrary to supervised learning, where performance is heavily influenced by data label distribution, we find that unsupervised learning in FL is more significantly affected by inter-class feature diversity. Our analysis reveals that datasets with low inter-class feature diversity, wherein representations of certain classes can be generalized to others, remain largely unaffected by Non-IID settings. Motivated by this insight, we propose the Representation-level CALibration aLgorithm (RECALL), a novel approach designed to calibrate latent representations in unsupervised models towards more generalized learning. RECALL leverages insights from previous global models to enhance local model training. Our experiments demonstrate the efficacy of RECALL in improving unsupervised learning tasks within the FedAVG framework under Non-IID conditions, underscoring the predominance of inter-class diversity over label distribution in unsupervised FL.

Original languageEnglish
Title of host publicationIntelligent Information and Database Systems - 16th Asian Conference, ACIIDS 2024, Proceedings
EditorsNgoc Thanh Nguyen, Ngoc Thanh Nguyen, Richard Chbeir, Yannis Manolopoulos, Hamido Fujita, Tzung-Pei Hong, Le Minh Nguyen, Krystian Wojtkiewicz
PublisherSpringer Science and Business Media Deutschland GmbH
Pages253-263
Number of pages11
ISBN (Print)9789819749812
DOIs
Publication statusPublished - 2024
Event16th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2024 - Ras Al Khaimah, United Arab Emirates
Duration: 15 Apr 202418 Apr 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14795 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference16th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2024
Country/TerritoryUnited Arab Emirates
CityRas Al Khaimah
Period15/04/2418/04/24

Keywords

  • Federated Learning
  • Non-IID
  • Unsupervised learning

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'RECALL: Towards Generalized Representations in Unsupervised Federated Learning Under Non-IID Conditions'. Together they form a unique fingerprint.

Cite this