Core Concepts
MAPL jointly learns heterogeneous personalized models and a collaboration graph in a decentralized peer-to-peer setting, without relying on a central server.
Abstract
The content discusses Model Agnostic Peer-to-Peer Learning (MAPL), a novel approach for learning heterogeneous personalized models in a decentralized setting.
Key highlights:
- MAPL operates in a model heterogeneous peer-to-peer (P2P) setting, where each client has a different feature extraction backbone.
- It consists of two main modules:
- Personalized Model Learning (PML): Learns personalized models using a combination of intra-client contrastive loss and inter-client prototype alignment.
- Collaborative Graph Learning (CGL): Dynamically refines the collaboration graph based on local task similarities in a privacy-preserving manner.
- MAPL jointly optimizes the personalized models and the collaboration graph in an alternating fashion.
- Extensive experiments demonstrate that MAPL outperforms state-of-the-art centralized model-agnostic federated learning approaches, without relying on a central server.
- MAPL can effectively identify clients with similar data distributions and learn an optimal collaboration graph.
Stats
The number of clients M is varied between 10 and 20.
Each client has access to a local dataset of 300 samples per class, with varying degrees of label distribution skew and statistical heterogeneity across clients.
Client models use different feature extraction backbones, including GoogLeNet, ShuffleNet, ResNet18, and AlexNet.
Quotes
"MAPL jointly learns personalized models and a collaboration graph in a decentralized peer-to-peer setting, without relying on a central server."
"MAPL outperforms state-of-the-art centralized model-agnostic federated learning approaches in extensive experiments."