AI-native Communications

In the last few years, with the advent of 5G (and beyond), communication networks are evolving from a pure communication framework to  service  enablers  in  several  different  sectors  (verticals),  such  as Industry 4.0, Internet of Things (IoT), autonomous driving, remote surgery, etc.  As key enablers of this vision, machine learning (ML)  and  artificial  intelligence  will  be  largely  exploited  in  future wireless  communication  networks,  in  order  to  build  an  effective complex system able to learn and dynamically adapt to the evolving  network  landscape.   Indeed,  the  advent  of  a  new  breed  of intelligent devices and high-stake applications foreseen in 6G have sparked a huge interest in distributed, low-latency and reliable ML, calling for a novel system design coined edge intelligence (EI), in which:  (i) training data is unevenly distributed over a large number of edge devices (including phones,  cameras, vehicles, and drones); (ii) every edge device has access to a tiny fraction of the data and training is carried out collectively and distributively;  (iii) the  inference  process  is  performed  on  the  edge  devices,  requiring not only high learning accuracy and reliability, but also a very short response time necessary for autonomous decision making in highly dynamic environments.  However, differently from cloud-based ML that has virtually infinite computing resources, EI is a nascent research field whose system design is entangled with communication and on-device resource constraints (e.g., energy and computing power).  Moreover, the process of decentralized training involves a large number of devices that are interconnected over wireless links, hindering  learning  and  adaptation  due  to  communications  under poor  channel  conditions.   As  such,  enabling  EI  introduces novel  research  problems  in  terms  of  jointly  optimizing  inference, training, communication, computation, and control under end-to-end latency, reliability, and learning performance requirements. In this context, our goal is twofold. 

Semantic and goal-oriented communications: A fundamental paradigm shift for envisioned applications is given by the semantic and goal-oriented communications approach, which goes beyond the typical bit-related metrics that are used today in system design and optimization, focusing instead on the recovery of the meaning  conveyed by the transmitted bits and/or the effective fulfillment of the tasks motivating the exchange of information. For instance, GO data compression is a key element of this new design, whose aim is to extract only relevant and useful information to the end-users/applications for serving the decision-making at the receiver with the required accuracy and respecting strict time constraints. In this context, we present an approach to semantic and goal-oriented communications building on three fundamental ideas: 1) represent data over a topological space as a formal way to capture semantics, as expressed through relations; 2) use the information bottleneck principle as a way to identify relevant information and adapt the information bottleneck online, as a function of the wireless channel state, in order to strike an optimal trade-off between transmit power, reconstruction accuracy and delay; 3) exploit probabilistic generative models as a general tool to adapt the transmission rate to the wireless channel state and make possible the regeneration of the transmitted images or run classification tasks at the receiver side.

Goal-oriented network optimization: The aim of this work is to put forward a goal-oriented (GO) system design for 6G networks, encompassing the key aspects related to architecture design and GO adaptive resource optimization. We look at the problem from a very wide and cross-layer perspective. Starting from our definition of goal effectiveness and the proposed architecture, we formally  define a communication goal as the fulfillment of a task (e.g., learning, actuation, control, etc.) with a target effectiveness level, which can be typically expressed in terms of a set of cross-layer performance measures (e.g., task accuracy, E2E latency). This approach naturally leads to a joint system optimization encompassing sensing, communication, computation, learning and control aspects, with the final aim of achieving effective GO communications with a minimum cost, e.g., in terms of energy or resource consumption. Such formulation entails striking the desired trade-off between goal effectiveness  and cost, which will be dynamically explored adapting the available degrees of freedom to cope with several sources of randomness affecting the system such as, e.g., wireless channels, data arrivals, data distribution over space and time, CPU availability, energy harvesting, etc. We will leverage  the interplay between model-based optimization (e.g., network stochastic optimization), which enable continuous learning and adaptation in random environments, and purely data-driven approaches (e.g., deep reinforcement learning) that are useful either when the environment is totally unknown, or it is too complex to be modeled.

Selected papers: