Researchers have introduced FlexOlmo, a framework allowing organizations to train AI models locally on their data while preventing data from leaving their networks. Traditional federated learning faces challenges due to regulations and privacy issues. FlexOlmo features a Mixture of Experts architecture, where each participant provides a trained submodel and router, enhancing task assignment efficiency. This method supports asynchronous model development, allowing organizations to contribute models at different times. Security evaluations revealed a low risk of data extraction from the final model, quantified at just 0.7% in test scenarios.
FlexOlmo enables organizations to train AI models locally on their own data and then combine those models without the data itself leaving the organization's network.
The FlexOlmo framework uses a joint anchor model that participants train locally, integrating their models into a Mixture of Experts architecture for efficient task assignment.
This approach allows organizations to join the model development process asynchronously, adding their models without needing to retrain the entire system.
Risk of data extraction from the final FlexOlmo model is very low, evaluated at only 0.7% in the test scenario conducted by researchers.
Collection
[
|
...
]