Continuous, Active Federated Learning for Data Streams

Purpose and goal

Vehicles generate large amounts of data through onboard sensors such as cameras. Traditional AI training methods can entail privacy and security risks if these data are stored or transferred. Current federated learning methods often require persistent data storage. This project aims to enable continuous, active and secure federated learning from real-time data for autonomous vehicles, ensuring regulatory data compliance while optimizing the utilization of the vehicles´ own compute resources.

Expected effects and result

We expect to show that it is possible to train useful models for automated driving over data streams in a continuous and decentralized way. We expect to highlight the contribution of active learning and self supervised learning approaches to this methodology, where models that are meant to improve continuously. We expect that the results will inform regulatory discussions on data privacy, AI security, and the cost of AI deployment in the context of traffic safety for automated vehicles.

Planned approach and implementation

We will leverage our partners´ expertise in methods like federated learning and on-the-edge incremental learning to examine enablers such as active learning and self-supervised learning in the context of continuous data streams. Our approach will be closely tied to and led by domain experts for AI and software design for automated driving. We will leverage our extensive partner networks to continuously disseminate our results across a wide ecosystem of players in the automotive field and beyond.