Semantic Over-the-Air Computation

Exploits the natural additive property of the wireless multiple-access channel (MAC) to compute a desired mathematical function (e.g., sum, average, max) directly over the air. Instead of treating simultaneous transmissions as interference to be canceled, the receiver interprets the aggregated analog waveform as the result of a nomographic function. This collapses communication and computation into a single physical-layer operation, drastically reducing latency compared to orthogonal scheduling.

Before transmission, each device passes its local data through a semantic encoder (typically a neural network) that extracts task-relevant features and discards irrelevant information. Crucially, these features are designed to be nomographically aligned—the sum of the transmitted feature vectors in the signal space directly corresponds to a meaningful aggregated semantic concept (e.g., a fused feature map for classification). This requires joint training of the encoders and the receiver's decoder.

Unifies three traditionally separate blocks—source coding, channel coding, and function computation—into a single learned operation. A deep neural autoencoder is trained end-to-end to optimize for the receiver's task accuracy (e.g., classification F1-score) rather than minimizing bit-error rate. The channel impairments (fading, noise) are treated as a non-trainable layer within the network, making the system inherently robust to the physical propagation environment.

Unlike digital systems that quantize and map features to discrete constellation points, semantic AirComp often employs analog modulation of the continuous-valued semantic feature vectors. Each device transmits a scaled version of its feature elements as amplitude-modulated symbols. The receiver's matched filtering and summation directly yields the aggregated feature vector in one shot, eliminating quantization error and enabling graceful degradation under low SNR conditions.

Devices apply complex precoding weights to their transmissions that serve a dual purpose: channel inversion for magnitude alignment at the receiver and task-aware feature scaling. This ensures that the summed signal accurately represents the desired function (e.g., a weighted average) despite heterogeneous path losses. The precoding is often optimized jointly with the semantic encoder using channel state information at the transmitter (CSIT) to minimize post-computation distortion.

Provides an inherent layer of physical-layer privacy. The receiver only observes the aggregated, superimposed waveform representing the computed function, not the individual semantic feature vectors from any single device. Recovering a specific device's raw data from the sum is mathematically ill-posed, offering a form of differential privacy without the explicit noise addition required in traditional federated learning, making it suitable for sensitive sensor networks.

A deep learning paradigm that replaces separate source and channel coding blocks with a single neural autoencoder, directly mapping source data to channel symbols. In the context of S-AirComp, JSCC enables the transmitter to encode semantic features into waveforms that are optimized for the superposition property of the multiple-access channel, ensuring that the sum received at the base station directly corresponds to the desired computation result rather than individual data streams.

An information-theoretic framework that learns a compressed, stochastic latent representation of an input that is maximally predictive of a target task while discarding irrelevant data. For S-AirComp, VIB provides the mathematical foundation for determining which semantic features to transmit:

• Compression: Strips away task-irrelevant noise before transmission
• Stochasticity: Creates robust representations that tolerate channel distortion
• Task alignment: Ensures the bottleneck preserves only features relevant to the computation function

The process of using a neural network to identify and isolate the high-level, task-relevant attributes from a raw signal, forming a compact semantic representation for transmission. In S-AirComp systems, each sensor device performs local feature extraction before transmission. The extracted features are designed such that their linear combination in the air yields the desired global function—for example, the mean, maximum, or a learned aggregation of distributed sensor readings.

A transmission paradigm where information is encoded and decoded based on its effectiveness in achieving a specific receiver task, rather than on symbol-level accuracy. S-AirComp is the ultimate expression of goal-oriented communication at the physical layer:

• The goal is the computation result, not individual data recovery
• Channel interference becomes a computational resource rather than noise
• Performance is measured by computation accuracy, not bit error rate

The optimization of the geometric arrangement of symbols in a digital modulation scheme to directly represent semantic features rather than arbitrary bit sequences. In S-AirComp, constellation points are designed so that their superposition at the receiver antenna produces a uniquely decodable computation result. This contrasts with traditional constellations optimized for minimum pairwise error probability, instead optimizing for post-computation fidelity.

A metric that quantifies the divergence between the intended meaning of a transmitted message and the meaning interpreted by the receiver, measured in task-relevant feature space. For S-AirComp systems, semantic distortion captures the error between the true function of all source data and the function computed from the received superposition. This replaces traditional MSE or BER as the primary optimization objective during end-to-end training.