The main results of FOGHORN address fog-aided wireless systems with a focus on communications, caching, and computing. These topics are reviewed in turn in the following.
Communications
FOGHORN developed fog-aided communication protocols that introduce novel functionality splits across cloud and edge across both data and control planes. These include rate selection and scheduling at the control plane and encoding/ decoding at the data plane. Particular attention was paid to common settings with heterogeneous traffic requirements, introducing the novel concept of heterogeneous non-orthogonal multiple access (H-NOMA), as well as to multi-tenant architectures. Fundamental theoretical limits were derived for the capacity benefits of reflecting intelligence surfaces (RISs) at the edge. Furthermore, applications of the proposed frameworks were studied for exemplifying scenarios such as virtual reality social networks and satellite networks.
Moving beyond the transmission of bits, the project pioneered the framework of semantics-based multiple access protocols that leverage processing across cloud and edge via type-based modulation. Type-based multiple access protocols were designed leveraging information-theoretic principles, demonstrating significant advantages over standard bit-based methods.
Finally, the project introduced novel methods for the design of data-efficient and reliable artificial intelligence (AI)-based solutions for communications in fog-aided wireless networks. Data-efficient methods can design effective protocols and algorithms by leveraging few data samples. FOGHORN pioneered the use of meta-learning as a general framework for the optimization of data-efficient protocols, inspiring many follow-up projects.
Reliable AI methods can provide trustworthy measures of uncertainty, supporting the use of AI within larger engineered systems encompassing both classical and AI-based algorithms. FOGHORN proposed for the first time principled frameworks based on Bayesian meta-learning and conformal prediction. The project showcased their benefits for a variety of functionalities such as beamforming, decoding, channel prediction, channel generation, and localization.
Caching
FOGHORN addressed the fundamental problem of understanding optimal functional splits and operation for caching content in fog-aided wireless systems, deriving optimal placement and delivery strategies across both fronthaul and wireless, or edge, segments. The introduced information-theoretic framework allows for fractional caching, as well as for general forms of fronthaul compression and data transmission. Extensions of the theoretical framework and of the optimality results were presented to address heterogeneous content popularity, partial wireless connectivity, device-to-device transmissions, linear precoding, imperfect channel state information, secrecy constraints, and online caching. The project also pioneered the idea of soft-time-to-live caching. Furthermore, optimal placement algorithms were derived for semantic-aware caching and routing strategies that adaptively place content and select optimal routes based on content similarity.
Computing
FOGHORN contributed to the definition of optimized computing and communication strategies in fog-aided wireless networks in terms of architectures, of fundamental design methodologies and protocols, and of technological insights.
At the level of network architectures, the project introduced the concept of a drone-based mobile cloud computing system in which a moving drone is endowed with computing capabilities to offer computation offloading opportunities.
On the front of fundamental design and protocols, FOGHORN proposed novel distributed protocols based on over-the-air computing that leverage the new idea of channel-driven sampling. Through this novel approach, distributed communication systems can ensure differential privacy “for free” by leveraging channel noise as privacy mechanism, while also benefiting from the presence of channel noise as seed randomness for distributed Bayesian learning. The project also introduced new coded distributed matrix multiplication in fog networks, as well as the first distributed unlearning mechanisms in device-to-device edge networks.
Finally, FOGHORN pioneered the use of neuromorphic signal processing and learning machines at the edge for semantic-aware transmission protocols, targeting applications such as sensing and remote inference.
