HDC-FL clusters classifier heads for personalized federated learning
Original title: From Classifier Heads to Encoder Prototypes: Personalized Federated Learning with Head-Derived Clustering
On 11 CIFAR setups with ResNet-18 and MobileNetV2, it leads worst-client and bottom-20% accuracy without averaging heads.
HDC-FL decides which federated clients should share representations by clustering classifier heads that shared-encoder training already produces. At refresh rounds the server groups projected heads, builds one encoder prototype per group, and returns each client a soft mixture plus its own head. No head averaging, second optimizer, or fine-tuning stage is used. Across 11 CIFAR-10 and CIFAR-100 setups it matches strong baselines on mean accuracy and ranks highest on worst-client and bottom-20% accuracy.
- Client-specific heads are never averaged or sent to other clients.
- Label-subset, Dirichlet, and rotation heterogeneity are all tested.
- A donor-swap diagnostic links head similarity to encoder compatibility.
- The link is clearest under covariate shift.
From the original article
1 min read at arxiv.orgPersonalized federated learning must decide which clients should share representations. We introduce HDC-FL (Head-Derived Clustering for Personalized Federated Learning), which reads that decision off a signal shared-encoder training already produces: the client-specific classifi- er heads.
By Amirardalan Dehghanpour · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.