FLT

FLT models task relatedness across federated clients to improve learning and aggregation on non-independent and identically distributed (non-IID) datasets.

Key Features:

• Pretrained Encoder for Data Compression: Each client uses a pretrained, fine-tunable encoder to compress local data into latent representations.
• Data Signature Transmission: Clients transmit signatures of their compressed latent representations to the server without sharing raw data.
• Task Relatedness via Manifold Learning: The server applies manifold learning techniques to the transmitted signatures to infer task relatedness among clients.
• Generalized Federated Averaging: Aggregation uses a generalized federated averaging informed by the learned task relatedness, allowing flexible handling of heterogeneous client relationships and optional decomposition into disjoint clusters.
• Improved Performance and Fairness: Task-related aggregation improves model performance and fairness across clients in non-IID federated settings.

Scientific Applications:

• Multi-institutional bioinformatics studies: Enables collaborative model training across institutions with heterogeneous datasets from different populations or experimental conditions while preserving client privacy.
• Handling population and experimental heterogeneity: Supports learning from datasets collected across diverse populations or experimental setups to improve generalization.
• Privacy-preserving federated learning across domains: Applicable to domains beyond bioinformatics that require federated training on non-IID data distributions.

Methodology:

Clients encode local data with a pretrained, fine-tunable encoder into latent representations; clients send signatures of the compressed data to the server; the server applies manifold learning to the signatures to learn task relatedness; the server performs generalized federated averaging guided by the learned task relatedness and can decompose clients into disjoint clusters when applicable.