Over-the-Air Goal-Oriented Communications

TL;DR

The work investigates goal-oriented communications where the objective is edge inference rather than exact data reconstruction, leveraging programmable metasurfaces to perform computations over the wireless channel. It treats the TX–RX–MS chain as a single differentiable neural network (MINN) trained end-to-end via backpropagation, using Deep Diffractive Neural Networks (D2NNs) and Stacked Intelligent Metasurfaces (SIM) to approximate digital neural networks with energy efficiency advantages. The framework supports channel-agnostic and channel-aware transceivers, with both reconfigurable and static metasurface configurations, and demonstrates MNIST-scale edge classification with favorable accuracy-energy tradeoffs and dynamic power control. Open challenges include introducing nonlinearity, advanced architectures, theoretical guarantees, and the possibility of over-the-air training to further exploit wave-domain computation in autonomous networks.

Abstract

Goal-oriented communications offer an attractive alternative to the Shannon-based communication paradigm, where the data is never reconstructed at the Receiver (RX) side. Rather, focusing on the case of edge inference, the Transmitter (TX) and the RX cooperate to exchange features of the input data that will be used to predict an unseen attribute of them, leveraging information from collected data sets. This chapter demonstrates that the wireless channel can be used to perform computations over the data, when equipped with programmable metasurfaces. The end-to-end system of the TX, RX, and MS-based channel is treated as a single deep neural network which is trained through backpropagation to perform inference on unseen data. Using Stacked Intelligent Metasurfaces (SIM), it is shown that this Metasurfaces-Integrated Neural Network (MINN) can achieve performance comparable to fully digital neural networks under various system parameters and data sets. By offloading computations onto the channel itself, important benefits may be achieved in terms of energy consumption, arising from reduced computations at the transceivers and smaller transmission power required for successful inference.

Figures (9)

• Figure 1: A of $M=3$ layers, each with $N_m=16$ elements, that is used as a for wave-domain-based . The controllable responses $[\ddot{\boldsymbol{\phi}}_m]_n$ ($m=1,\dots,M$ and $n=1,\dots,N$) are treated as trainable weights.
• Figure 2: Illustration of techniques for transferring digital data to and from the RF domain using programmable as the first and last layers. Once training is complete, unit cells characterized by the obtained responses for intermediate layers may be manufactured to be completely passive, providing important benefits in terms of energy consumption during wave computation.
• Figure 3: The system considered in the framework incorporating either an or a device. The may include a -based controller or a basic processing unit to store or update their fixed configuration.
• Figure 4: Block diagram and computation flow of the architecture Stylianopoulos_GO where the metasurface-parametrizable channel acts as an intermediate component. Both the cases of reconfigurable and static metasurfaces are included, entailing different procedures during the forward and backward passes.
• Figure 5: Mean classification accuracy of on static fading considering and with different numbers of elements and layer, as well as two transmit values.