Federated Learning Using In-Space Data (FLUID )
Description
Enables neural net models to be trained using a combination of terrestrial data and space-borne data without the need to downlink or uplink data for consolidation and training in the normal manner. Instead, small components of the overall neural net model are trained in-situ with the data and transmitted for federation into a single neural net, thereby reducing data transmission demands and reducing overall latencies by several orders of magnitude.
Detailed example
For example, a neural net model to monitor lunar habitat astronauts for signs of lung toxicity due to regolith inhalation could be largely trained on Earth using similar data (e.g. data collected for volcanic ash inhaltion) and then fine-tuned with data generated in-situ on the lunar surface, without the need to transmit that data back to Earth. The FLUID arhitecture has been fully tested using the Spaceborne Computer-2 on board the ISS; in February of 2024, the world's first neural net model trained with both terrestrial and in-space data was successfully trained.
AI / analytics pattern
Classical/Predictive Machine Learning: Models trained on data to make predictions or classifications based on identified patterns or relationships.
Automation level / stage
a) Pre-deployment – The use case is in a development or acquisition status.
Expected benefit
Enables neural net models to be trained using a combination of terrestrial data and space-borne data without the need to downlink or uplink data for consolidation and training in the normal manner. Instead, small components of the overall neural net model are trained in-situ with the data and transmitted for federation into a single neural net, thereby reducing data transmission demands and reducing overall latencies by several orders of magnitude.
Controls / human review
ATO: Not reported; PIA: Not published