TY - GEN
T1 - RECALL
T2 - 16th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2024
AU - Chen, Pi Wei
AU - Chun-Wei Lin, Jerry
AU - Yeh, Feng Hao
AU - Cupek, Rafał
AU - Chen, Chao Chun
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.
PY - 2024
Y1 - 2024
N2 - 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.
AB - 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.
KW - Federated Learning
KW - Non-IID
KW - Unsupervised learning
UR - https://www.scopus.com/pages/publications/85202815039
U2 - 10.1007/978-981-97-4982-9_20
DO - 10.1007/978-981-97-4982-9_20
M3 - Conference contribution
AN - SCOPUS:85202815039
SN - 9789819749812
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 253
EP - 263
BT - Intelligent Information and Database Systems - 16th Asian Conference, ACIIDS 2024, Proceedings
A2 - Nguyen, Ngoc Thanh
A2 - Nguyen, Ngoc Thanh
A2 - Chbeir, Richard
A2 - Manolopoulos, Yannis
A2 - Fujita, Hamido
A2 - Hong, Tzung-Pei
A2 - Nguyen, Le Minh
A2 - Wojtkiewicz, Krystian
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 15 April 2024 through 18 April 2024
ER -