The article presents an extensive review of Federated Learning (FL) as applied to visual recognition, emphasizing the importance of thoughtful architectural design in achieving optimal performance. The authors argue that many existing FL solutions are tested on shallow or simple networks, which may not accurately reflect real-world applications. They demonstrate through an in-depth analysis of various cutting-edge architectures that architectural choices can significantly enhance FL systems’ performance, especially when handling heterogeneous data. The study also re-examines the inferior performance of convolution-based architectures in the FL setting and analyzes the role of normalization layers on FL performance.
Publication date: 23 Oct 2023
Project Page: https://github.com/sarapieri/fed_het.git
Paper: https://arxiv.org/pdf/2310.15165
