Researchers at the Allen Institute for AI (Ai2) have developed FlexOlmo, a system that allows AI models to be trained on private datasets without owners sharing their data.
Data owners can also remove or limit their data’s use after training is finished.
The system allows model creators and those with data to collaborate without either needing to compromise.
It works by owners of private datasets training smaller models that are then merged with a pre-trained, publicly available model.
This leads to improved performance, with the owner’s data and preferences influencing the end result, without either needing to be shared.
It could help healthcare and government data to be used in AI models without personal information being compromised.
