Federated Inference through Aligning Local Representations and Learning a Consensus Graph

Machine learning is faced with many data challenges when applied in practice. Among them, a notable barrier is that data are distributed and sharing is unrealistic for volume and privacy reasons. Federated learning is a recent formalism to tackle this challenge, so that data owners can develop a common model jointly but use it separately. In this work, we consider a less addressed scenario where a datum consists of multiple parts, each of which belongs to a separate owner. In this scenario, joint efforts are required not only in learning but also in inference. We study \emph{federated inference}, which allows each data owner to learn its own model that captures local data characteristics and copes with data heterogeneity. On the top is a federation of the local data representations, performing global inference that incorporates all distributed parts collectively. To enhance this local--global framework, we propose aligning the ambiguous data representations caused by arbitrary arrangement of neurons in local neural network models, as well as learning a consensus graph among data owners in the global model to improve performance. We demonstrate effectiveness of the proposed framework on four real-life data sets including power grid systems and traffic networks.