Cartel : A System for Collaborative Transfer Learning at the Edge Cloud

The current approach to learn from the geodistributed data is to create a centralized, global model which is done over data collected from different datacenters transferred to the central server. The centralized model is therefore generic, in that it can provide acceptable accuracy for a wide domain of queries. The rise of Mobile Edge Computing (MEC) enables reduction in network congestion and improved performance (i.e., lower latency, increase bandwidth) of emerging applications, by bringing compute resources closer to the user, at preaggregation/aggregation points in a cellular network. In this environment, it may be beneficial to have machine learning models tailored for each individual node, enabling the MEC benefit of reducing backhaul network demand, and addressing new data privacy laws, such as the GDPR, which can prohibit data transmission across national or continental borders. This motivates the need for a collaborative learning system where: (1) each node has a personalized model, and; (2) as traffic characteristics change over time, it can request help from a logical neighbour to improve performance, with- out sharing the raw data. In this paper, we propose the system abstractions and mechanisms that realize this knowledge transfer between edge nodes collaborative environment, and build Cartel, a system for collaborative learning at the edge. Our experiments demonstrate that Cartel can improve accuracy compared to isolated learning, and reduce data transfer, training time and model size, compared to centralized learning with similar accuracy.