Federated learning gets faster when multiple organizations train AI together without sharing data

A new method for training neural networks across separate organizations makes the process more stable when each organization has different types of data. Instead of averaging all the weights from different organizations' models (which causes problems when the data doesn't match), this approach freezes the structural parts of a pretrained model and only averages the numerical fine-tuning, which reduces instability and speeds up convergence.

Federated learning solves a real problem: multiple organizations want to train a shared AI model together without exposing their proprietary data. But when each organization's data is different — different image distributions, different tasks, different quality — simple weight averaging fails. This paper shows that treating the structural and numerical parts of a neural network differently during training reduces that instability. The practical effect is that cross-organizational AI training becomes more reliable and faster, which matters for industries like healthcare, finance, and manufacturing where data sharing is restricted but collaboration would be valuable.