New label granularity skew challenge identified in federated learning

Researchers have introduced a new challenge in federated learning called label granularity skew, where clients in a hierarchical image classification task provide labels at varying levels of detail. To address this, they developed a method using a probabilistic relational neighbor classifier to generate local label hierarchies and a conditional softmax classifier that is more robust to incomplete supervision. The proposed Branch-wise Decoupled Fine-Tuning (BDFT) and its federated version, FedBDFT, showed significant improvements in robustness, with average gains of up to 56.4% on datasets like CIFAR-100 and ImageNet under severe skew. AI

IMPACT Introduces a new challenge in federated learning that may require new approaches for robust hierarchical classification.

RANK_REASON The item is a research paper detailing a new challenge and proposed solution in federated learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Branch-wise Decoupled Fine-Tuning
• CIFAR-100
• conditional softmax classifier
• FedBDFT
• federated learning
• Hierarchical image classification in the bioscience literature
• ImageNet
• label granularity skew
• probabilistic relational neighbor classifier
• TinyImageNet
• WordNet

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →