Semantic Hybrid ARQ (S-HARQ)

Unlike traditional HARQ that retransmits entire code blocks, S-HARQ identifies and requests only the corrupted semantic features in the latent space. The receiver's semantic decoder provides feedback on which dimensions of the feature vector failed to meet a confidence threshold, triggering a targeted retransmission that directly repairs the meaning.

S-HARQ prioritizes retransmission based on task-criticality rather than bit-error rate. Features essential for the receiver's goal—such as object class in an image classification task—are assigned higher priority. Non-essential semantic noise is ignored, ensuring retransmission resources are allocated where they maximally impact task performance.

Each retransmission incrementally refines the receiver's semantic posterior. The decoder fuses the original corrupted feature with the newly received refinement using attention mechanisms or Bayesian updating. This allows the system to accumulate partial semantic information across multiple rounds, converging on the correct interpretation without restarting decoding from scratch.

S-HARQ is natively integrated with Joint Source-Channel Coding (JSCC) architectures. The semantic encoder and channel encoder are a single neural network, enabling the retransmission protocol to operate directly on the learned latent manifold. This tight coupling eliminates the information loss that occurs when traditional HARQ interfaces with separate source and channel codecs.

The system dynamically adjusts the amount of semantic redundancy injected into each transmission based on channel state information and the semantic importance of the content. For critical features, the encoder introduces controlled redundancy in the latent space—distinct from bit-level parity—that allows the decoder to recover meaning even under severe channel impairments without immediate retransmission requests.

When a semantic feature is corrupted, S-HARQ can leverage cross-modal correlations to recover meaning without retransmission. For example, if the audio feature for a phoneme is lost, the decoder may infer it from the accompanying visual feature of lip movement in a multimodal transmission, reducing retransmission frequency in rich media applications.

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. Unlike traditional bit-error rate, semantic distortion directly informs S-HARQ's retransmission decisions by identifying which specific feature dimensions have been corrupted and require correction.

A deep learning framework based on information theory that learns a compressed, stochastic latent representation maximally predictive of a target task while discarding irrelevant data. VIB provides the theoretical basis for S-HARQ's feature selection, enabling the system to identify which semantic features are most critical for task completion and prioritize their retransmission.

A technique that corrects transmission errors by leveraging the semantic context and meaning of received data rather than relying solely on redundant parity bits. S-HARQ extends this concept into a feedback-driven protocol where the receiver explicitly signals which semantic features were corrupted, enabling targeted retransmission that is more bandwidth-efficient than full packet repetition.

A transmission paradigm where information is encoded and decoded based on its effectiveness in achieving a specific receiver task rather than symbol-level accuracy. S-HARQ operationalizes this principle by ensuring that retransmission resources are allocated only to features that directly impact task performance, making it a task-aware reliability mechanism.

A shared, structured repository of background knowledge and ontologies used by both transmitter and receiver to interpret and disambiguate meaning. In S-HARQ systems, the SKB enables the receiver to perform predictive feature reconstruction — filling in corrupted semantic features from context without requiring retransmission, further reducing overhead.