Updating machine learning models across devices

Abstract

A system may share updates to a machine learning model between a first device and a second device. The first device may determine one or more updatable layers of the model. The first device may train the first model to update parameters of the updatable layers in response to processing input data using the model. The first device may keep parameters of model layers other than the updatable layers constant during the training. The first device may send, to the second device, first model update data representing the updated parameters. The second device may aggregate model update data (e.g., from multiple devices), determine second model update data, and send the second model update data to the first device. The first device may update its model based on the second model update data.