Fed-BioMed

The goal of Federated learning is to allow collaborative learning with decentralized data.

Healthcare is a typical application of federated learning: while hospitals across several geographical locations want to jointly train a machine learning model on the data hosted at each site, data cannot be shared between them because of privacy and security concerns. Federated learning gives us a methodological framework to train a global machine learning model, by only sharing the parameters of the models separately trained at each site. As a result, data never leaves the hospitals, and training is performed by simply aggregating models parameters to finally obtain a global model. Under certain conditions, the aggregated model faitfully represents the global variability across hospitals, and provides high generalization and robustness properties.