Institution profile

Malmö University

Academic institutioneurope · se
Official website
Research library32linked papers
Opportunities0open roles
Selected work

Representative Papers

FEAST: Federated Shared-Space Training for Resource-Heterogeneous Clients

Aug 10, 2026

This work addresses the challenges of model adaptation and low training efficiency in federated learning caused by computational heterogeneity among clients. To overcome these issues, the authors propose an elastic supernetwork training framework that jointly trains multiple subnetworks within each client’s local inference budget. The approach introduces a sub-supernetwork routing mechanism and a sparse parameter aggregation strategy to enable efficient collaborative training within a shared parameter space. Furthermore, a γ-allocation protocol is designed to decouple the confounding effects of data volume and computational budget on accuracy estimation, allowing flexible post-training deployment of subnetworks at arbitrary scales. Experiments demonstrate that the method achieves 71.06% accuracy on CIFAR-100 with only 596M MACs, significantly outperforming baseline approaches while reducing communication overhead by 6.8×.

0 citationsRead paper
Recent publications

Latest Papers

FEAST: Federated Shared-Space Training for Resource-Heterogeneous Clients

Aug 10, 2026

This work addresses the challenges of model adaptation and low training efficiency in federated learning caused by computational heterogeneity among clients. To overcome these issues, the authors propose an elastic supernetwork training framework that jointly trains multiple subnetworks within each client’s local inference budget. The approach introduces a sub-supernetwork routing mechanism and a sparse parameter aggregation strategy to enable efficient collaborative training within a shared parameter space. Furthermore, a γ-allocation protocol is designed to decouple the confounding effects of data volume and computational budget on accuracy estimation, allowing flexible post-training deployment of subnetworks at arbitrary scales. Experiments demonstrate that the method achieves 71.06% accuracy on CIFAR-100 with only 596M MACs, significantly outperforming baseline approaches while reducing communication overhead by 6.8×.

0 citationsRead paper