Digital Wireless Communications for Computations

In the recent years, the need of running machine learning (ML) services over wireless communication networks has promoted the design of new wireless communication protocols capable to efficiently support such ML services. In fact, in wireless networks, ML services face major challenges in terms of computation, bandwidth, scalability, privacy, and security. One proposal to overcome such challenges is Over-the-air computation (OAC), which is a known technique where wireless devices transmit values by analog amplitude modulation so that a function of these values (e.g., Federated Learning gradient aggregations) is computed over the communication channel at a common receiver. OAC dramatically reduces communication energy use, amplify spectrum efficiency of several order of magnitudes, and achieve privacy protections. The physical reason is the superposition properties of the electromagnetic waves, which naturally return sums of analog values. Consequently, the applications of OAC are almost entirely restricted to analog communication systems. However, the use of digital communications for OAC would have several benefits, such as error correction, synchronization, acquisition of channel state information, and easier adoption by current digital communication systems. In this talk, we will present a fundamentally new computing method, named ChannelComp, for performing OAC by any digital modulation. We will show how digital modulation formats allow us to compute a wider class of functions than OAC can compute, and we propose a feasibility optimization problem that ascertains the optimal digital modulation for computing functions over the-air. We will introduce the concept of channel coding for digital over the air computation. We show by analysis and simulation the superior performance of ChannelComp in comparison to OAC.