Learning in Next Generation (NextG) wireless systems is expected to bring about a technological and societal revolution even bigger than that which data brought to early voice-centered systems. Learning will have to be performed on data predominantly originating at edge and user devices in order to support applications such as Internet of Things (IoT), federated learning, mobile healthcare, self-driving cars, and others. A growing body of research work has focused on engaging the edge in the learning process, which can be advantageous in terms of a better utilization of network resources, delay reduction, resiliency against cloud unavailability and catastrophic failures, and increased security and privacy. Present proposed solutions, however, predominantly suffer from having a critical centralized component, typically in the cloud, that organizes and aggregates the nodes’ computations. This rigid centralized infrastructure can inhibit the full potential of resiliency and privacy in NextG systems. By relaxing the centralized infrastructure, the proposed research aims to advance Random Walk learning algorithms as the basis of a unified framework for the joint design of distributed learning and networking, with resiliency and privacy being the overarching goal.

In Random Walk learning, the model can be thought of as a “baton” that is updated and passed from one node (cloud, edge node, end-devices, etc.) in the network to one of its neighbors that is smartly chosen. This baton can be then passed to the cloud at a prescribed schedule and/or adaptively as part of the random walk, allowing thus a fluid architecture where centralization and full decentralization constitute two corner points. The proposed work will focus on major challenges and opportunities specific to the applicability of random walk learning in NextG, namely: (i) Adaptability to the heterogeneity of the data and the heterogeneity and dynamic nature of the network; (ii) Resiliency and graceful degradation in the face of failures via coding-theoretic redundancy methods; (iii) Model distribution across nodes and random walking snakes; and (iv) Privacy of the locally owned data.

This award reflects NSF’s statutory mission and has been deemed worthy of support through evaluation using the Foundation’s intellectual merit and broader impacts review criteria.