Goal-Oriented Communication

The entire communication stack is jointly optimized for a specific downstream machine task (e.g., image classification, object detection) rather than minimizing symbol or bit error rate.

• Loss Function: The training objective is the receiver's task loss (e.g., cross-entropy for classification), not mean squared error of the reconstructed signal.
• End-to-End Learning: The transmitter and receiver are trained as a single neural autoencoder, learning a task-relevant latent representation.
• Result: The system learns to allocate bandwidth and power to the signal features that most impact task accuracy, ignoring irrelevant details.

Information is compressed beyond the source entropy limit by discarding data that is irrelevant to the receiver's goal, achieving a semantic rate lower than the Shannon rate.

• Information Bottleneck: The encoder extracts a minimal sufficient statistic of the input for the task, stripping away nuisance variables like background noise or lighting.
• Example: For a vehicle-to-vehicle collision avoidance task, the system transmits the position and velocity vectors of nearby cars but omits their color, make, and model.
• Efficiency Gain: This enables reliable communication at signal-to-noise ratios where classical systems fail, as the transmitted symbols carry only task-critical meaning.

The traditional separation of source coding and channel coding is replaced by a single neural encoder-decoder pair that maps source data directly to channel symbols.

• Mechanism: The encoder outputs a continuous-valued vector that is directly mapped to complex baseband symbols for transmission over the physical channel.
• Graceful Degradation: Unlike digital systems that exhibit a cliff effect, JSCC systems degrade gracefully as channel quality worsens, with task accuracy decreasing smoothly.
• Bandwidth Adaptivity: A single trained model can dynamically adjust its coding rate to match varying channel conditions without explicit rate negotiation.

The receiver's current state, context, and intent actively shape what information the transmitter sends, creating a bidirectional semantic link.

• Query-Based Transmission: The receiver sends a compact semantic query describing its information need, and the transmitter responds only with the relevant semantic features.
• Shared Knowledge Base: Both ends maintain a synchronized semantic knowledge base (SKB) of common concepts, allowing the transmitter to send lightweight pointers or indices instead of full descriptions.
• Example: In a factory digital twin, a controller requests only the temperature of a specific machine component, and the sensor transmits that single semantic feature vector.

The system is designed to be robust against semantic noise—distortions that corrupt the meaning of a message rather than its bit representation.

• Sources of Semantic Noise: Ambiguous context, mismatched background knowledge between transmitter and receiver, and adversarial perturbations.
• Mitigation: The decoder uses its semantic knowledge base and task context to resolve ambiguities and correct plausible but incorrect interpretations.
• Adversarial Training: Models are hardened against semantic adversarial attacks by training on perturbed examples designed to cause task failure, ensuring mission-critical reliability.

Traditional quality of service (QoS) metrics like bit error rate (BER) and throughput are replaced by task-effectiveness metrics.

• Semantic QoS: Defined by the accuracy, precision, recall, or F1-score of the receiver's task output, not the fidelity of the reconstructed signal.
• Semantic QoE: Measures the user's perceived quality based on successful interpretation and utility of the received meaning.
• Example: In a remote surgery application, the QoS guarantee is the mean absolute error of the robotic instrument position, not the packet delivery ratio of the haptic feedback stream.

A deep learning paradigm that replaces separate source and channel coding blocks with a single neural autoencoder. This system directly maps source data to channel symbols, optimizing for end-to-end transmission rather than intermediate bit-level fidelity. Unlike traditional separate coding, JSCC gracefully degrades under poor channel conditions and eliminates the 'cliff effect' common in digital systems.

A shared, structured repository of background knowledge, ontologies, and common sense used by both transmitter and receiver. The SKB enables disambiguation of transmitted meanings by providing a common frame of reference. Key functions include:

• Resolving ambiguous symbols using context
• Storing shared world models for compression
• Enabling zero-shot interpretation of novel messages

An information-theoretic deep learning framework that learns a compressed, stochastic latent representation of an input. The VIB objective maximizes mutual information with a target task while minimizing mutual information with the input, effectively discarding task-irrelevant data. This provides a principled mathematical foundation for semantic feature extraction and rate-distortion trade-offs.

Semantic Quality of Service (QoS) defines network guarantees by task-completion accuracy rather than bit error rate. Semantic Quality of Experience (QoE) measures perceived quality based on successful interpretation and utility of received meaning. These metrics shift evaluation from:

• Symbol-level accuracy → Task effectiveness
• Throughput → Semantic relevance
• Latency → Interpretation speed

The resilience of a semantic communication system against malicious, imperceptible perturbations designed to cause misinterpretation at the receiver. Unlike traditional jamming that targets bits, semantic attacks inject perturbations that flip the meaning of a message while preserving its surface form. Defense techniques include adversarial training and certified semantic bounds.

An architecture that partitions a deep semantic model between an edge device and a network server. The device transmits compact, intermediate semantic features instead of raw data, balancing compute load and privacy. Benefits include:

• Reduced uplink bandwidth requirements
• Privacy preservation by design
• Dynamic workload distribution based on channel conditions