Cooperative learning by knowledge transfer in vehicular networks for applications with user-environment interactions

Real time vehicle monitoring and driver behaviour analysis has become an essential requirement for ensuring road-safety. Given the fact that the sensors deployed in vehicles can provide helpful data in analysing the driving behaviour, Machine Learning (ML) plays a vital role in many vehicular safety applications. However, it is challenging and demanding to build ML models for all the vehicle users that would be effective in all type of road environments. It is obvious that a vehicle user's driving behaviour would vary depending on the road-environment. Also, each vehicle user will have a different driving style. Hence, the interaction between the vehicle user and the environment is an important aspect to be considered while building ML models. In this paper, we investigate a novel learning approach that specializes some of the Neural Network layers on the vehicle user, whereas the others are specific of the current environment. Furthermore, the environment-specific layers are trained cooperatively by users that are currently visiting such environment. We compared two cooperative learning solutions: one based on federated learning and the other based on gossiping. Effectiveness and efficiency of the whole approach will be assessed and the performance of the cooperative training solutions will be discussed.