Research map / Label-efficient and robust learning / Recognition and long-tail
Beyond Federated Prototype Learning: Learnable Semantic Anchors with Hyperspherical Contrast for Domain-Skewed Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FedLSA, which replaces averaged federated prototypes with learnable server-side semantic anchors (trained via a global classifier plus a separation loss) and enforces intra-class compactness for domain-skewed clients by projecting local features onto a hypersphere with a vMF-style contrastive loss.
Problem
Federated prototype learning under domain-skewed data: averaged local prototypes carry domain-specific markers, shrinking margins between global prototypes, and Euclidean local embeddings are non-uniform, hindering prototype-based contrastive learning.
Approach
Learn class anchors directly on the server through a global semantic-aware classifier and an anchor separation loss instead of averaging embeddings, and project client embeddings onto a hypersphere where a von Mises-Fisher-derived contrastive loss pulls samples toward their anchors.
Key results
- FedLSA reaches 63.56% average accuracy on Digits, the best among compared methods (FPL 62.09%, FedAvg 60.89%).
- On Office Caltech FedLSA averages 60.22%, a +7.02 gain over FedAvg, while FedProto drops below FedAvg (52.01% vs 53.20%).
- On PACS FedLSA averages 43.73%, +3.37 over FedAvg and above FedGA (42.51%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedLSA instead of FedAvg | Federated learning · Digits | average accuracy: +2.67 | Helped |
| use FedLSA instead of FedAvg | Federated learning · Office Caltech | average accuracy: +7.02 | Helped |
| use FedLSA instead of FedAvg | Federated learning · PACS | average accuracy: +3.37 | Helped |
| include LCOM (with LSEP absent and L2-norm constraint) (ablation) | FedLSA · Digits | average accuracy: 60.89 -> 61.06 | Helped |
| include LSEP (ablation) | FedLSA · Digits | average accuracy: 60.89 -> 61.21 | Helped |
| include LSEP and LCOM (ablation) | FedLSA · Digits | average accuracy: 60.89 -> 63.56 | Helped |
| include LCOM (with LSEP absent and L2-norm constraint) (ablation) | FedLSA · Office Caltech | average accuracy: 53.20 -> 59.18 | Helped |
| include LSEP (ablation) | FedLSA · Office Caltech | average accuracy: 53.20 -> 58.37 | Helped |
| include LSEP and LCOM (ablation) | FedLSA · Office Caltech | average accuracy: 53.20 -> 60.22 | Helped |
Surprising findings
- FedProto underperforms FedAvg on Office Caltech, attributed to margin decay among global prototypes under domain skew.
Limitations
- FedLSA does not achieve the best per-domain accuracy on all domains (e.g., lower on MNIST and PACS SKETCH).
- Evaluated only with a small CNN on three image benchmarks with 10 clients.
Experimental setup
Datasets: Digits (MNIST, USPS, SVHN, SYN) Office Caltech PACS
Benchmarks: Digits Office Caltech PACS
Metrics: accuracy (%) average accuracy across domains improvement over FedAvg
Baselines: FedAvg FedProx MOON FedSR FedProto FedGA FPL FedTGP
Builds on: FedProto MOON FPL FedTGP
Closest papers by meaning
- FedSA: A Unified Representation Learning via Semantic Anchors for Prototype-based Federated Learning (AAAI 2025 · similarity 0.92)
- Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains (NeurIPS 2024 · similarity 0.88)
- FedTGP: Trainable Global Prototypes with Adaptive-Margin-Enhanced Contrastive Learning for Data and Model Heterogeneity in Federated Learning (AAAI 2024 · similarity 0.87)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.87)
- Beyond Prototypes: Semantic Anchor Regularization for Better Representation Learning (AAAI 2024 · similarity 0.86)
- Resource-Aware Federated Self-Supervised Learning with Global Class Representations (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.