PaperlayerView exact revision on arXiv

Source-linked AI summary

What is Semantic Communication? A View on Conveying Meaning in the Era of Machine Intelligence

Qiao Lan, Dingzhu Wen, Zezhong Zhang, Qunsong Zeng, Xu Chen, Petar Popovski, Kaibin Huang

arXiv:2110.00196v1cs.ITeess.SP

TL;DR

Rate-centric communication theory treats message meaning as largely irrelevant, motivating semantic communication for machine-intelligence systems. The paper presents SemCom principles, architectures, applications, and knowledge-graph approaches, with SplitNet and AirComp illustrating task-oriented designs.

  • Problem

    Rate-centric communication frameworks largely ignore transmitted meaning, limiting their fit for machine-intelligence systems and ambitious 6G goals.

  • Method

    The paper surveys SemCom principles, layer-coupling and end-to-end architectures, application techniques, AirComp-based designs, and knowledge-graph approaches.

  • Results

    The reviewed techniques support task-oriented communication, including improved SplitNet performance, optimal AirComp rate-distortion tradeoffs for many sensors, and reduced FL latency scaling.

  • Takeaways & Limitations

    SemCom reframes communication around conveying meaning and effectively executing computation tasks across human-machine and machine-machine systems.

  • Takeaways & Limitations

    SplitNet can converge slowly under channel perturbations and may require time-consuming retraining when radio-propagation environments change.

Abstract

from arXiv · show

In 1940s, Claude Shannon developed the information theory focusing on quantifying the maximum data rate that can be supported by a communication channel. Guided by this, the main theme of wireless system design up until 5G was the data rate maximization. In his theory, the semantic aspect and meaning of messages were treated as largely irrelevant to communication. The classic theory started to reveal its limitations in the modern era of machine intelligence, consisting of the synergy between IoT and AI. By broadening the scope of the classic framework, in this article we present a view of semantic communication (SemCom) and conveying meaning through the communication systems. We address three communication modalities, human-to-human (H2H), human-to-machine (H2M), and machine-to-machine (M2M) communications. The latter two, the main theme of the article, represent the paradigm shift in communication and computing. H2M SemCom refers to semantic techniques for conveying meanings understandable by both humans and machines so that they can interact. M2M SemCom refers to effectiveness techniques for efficiently connecting machines such that they can effectively execute a specific computation task in a wireless network. The first part of the article introduces SemCom principles including encoding, system architecture, and layer-coupling and end-to-end design approaches. The second part focuses on specific techniques for application areas of H2M (human and AI symbiosis, recommendation, etc.) and M2M SemCom (distributed learning, split inference, etc.) Finally, we discuss the knowledge graphs approach for designing SemCom systems. We believe that this comprehensive introduction will provide a useful guide into the emerging area of SemCom that is expected to play an important role in 6G featuring connected intelligence and integrated sensing, computing, communication, and control.

1 Introduction

Semantic communication broadens communication-system design beyond rate-centric reliability toward conveying meaning and achieving task effectiveness. The paper organizes this shift across H2H, H2M, and M2M contexts, emphasizing H2M and M2M in machine-intelligence systems.

  • Motivation: Shannon’s framework treated semantic aspects as irrelevant, while conventional wireless design prioritized throughput, efficiency, and latency.The paper argues that separating transmission from meaning and task effectiveness creates redundancy and difficulty handling rapidly growing data traffic.
  • SemCom framework: Semantic communication addresses Weaver’s second-level semantic problem and extends toward the third-level effectiveness problem.The semantic problem asks how precisely transmitted symbols convey desired meaning, whereas the effectiveness problem concerns how effectively received meaning achieves its purpose.
  • Communication contexts: The paper defines three communication contexts: human-to-human, human-to-machine, and machine-to-machine communications.H2H focuses on interpreting intended human meaning; H2M requires messages to be meaningful to humans and machines; M2M coordinates machines to execute computing tasks.
  • Communication contexts: H2M SemCom supports human–machine interaction, while M2M SemCom targets efficient machine coordination for distributed sensing, learning, and consensus.M2M research advocates integrated communication and computing under radio and computing-resource constraints.
  • Scope and contributions: The article surveys SemCom principles, application techniques, knowledge-graph approaches, and technologies relevant to 6G.Its emphasis is on newer H2M and M2M paradigms, while also positioning knowledge graphs across all three communication contexts.

2 SemCom Principles: Encoding, Architecture, and Design Approaches

SemCom extends conventional communication beyond symbol transmission by incorporating semantic meaning and task effectiveness into encoding, architecture, and system design. The paper presents semantic and effectiveness encoding, a layered architecture, and layer-coupling and end-to-end approaches.

  • Communication foundations: Shannon’s model reproduces messages through source encoding, channel transmission, and receiver decoding, while information-theoretic encoding focuses on message statistics rather than content.The communication chain includes a source, transmitter, channel, receiver, and destination; channel noise can distort the transmitted signal.
  • SemCom principles: SemCom addresses semantic meaning and effectiveness in addition to Shannon’s technical transmission problem.Semantic encoding targets conveying intended meaning, whereas effectiveness encoding targets delivering instructions or queries that machines can execute appropriately.
  • Effectiveness encoding: Effectiveness encoding extracts information useful for a destination machine’s task, such as discriminative features for classification or compressed gradients for distributed learning.Gradient transmission requires task-appropriate compression because mean-squared error does not directly capture gradient-direction deviation and generic quantization scales poorly with dimension.
  • Semantic encoding: Semantic transmission can judge errors by meaning rather than symbol accuracy: receiving “cattle” for “cow” is semantically closer than receiving “pig” for “big.”The example shows that a multi-letter change may preserve meaning better than a single-letter error, exposing limits of semantic-agnostic measures.
  • Semantic encoding: Practical semantic encoding includes knowledge-graph representations and learned joint semantic-channel coding that compresses messages according to semantic information.The knowledge-graph approach builds a shared background representation before encoding data, while learning-based schemes jointly train semantic and channel encoders.
  • Architecture and design approaches: The proposed Semantic Layer interfaces with application data, sensors, and actuators, while lower radio-access layers continue optimizing rate, reliability, and latency.Layer coupling exchanges semantic controls with radio-access layers; fully integrated SplitNet can improve end-to-end performance, whereas layer coupling offers backward compatibility, modularity, and flexibility.

3 Human-to-Machine Semantic Communications

H2M SemCom conveys messages understandable by both humans and machines, enabling dialogue and machine assistance across several application areas.

  • H2M SemCom transmits messages that humans and machines can understand for dialogue or machine assistance.
  • The paper examines human-machine symbiosis, recommendation, human sensing and care, and VR/AR as representative H2M application areas.

3.1 SemCom for Human-Machine Symbiosis

Human-machine symbiosis uses semantic encoding to capture meaningful human sensing data or knowledge in compact representations that support complementary cooperation and machine assistance.

  • Human-machine symbiosis establishes complementary cooperation in which humans and machines jointly perform difficult tasks and improve one another’s capabilities.
  • Semantic encoding maps human sensing data or knowledge into low-dimensional vectors while preserving semantic meanings or latent features.Examples include embedded knowledge in text and object-background boundaries in images.
  • Semantic Encoding by LSA: LSA extracts semantic information by representing documents in a document-term matrix, applying SVD, and projecting them into a k-dimensional semantic space.The paper denotes the decomposition as X = UΣV^T and the reduced semantic space as U_k.
  • Semantic Encoding by LSA: LSA-based semantic encoders can be integrated into either layer-coupling or SplitNet SemCom architectures.One SplitNet design integrates semantic and channel coding through a split DNN whose transmitter performs LSA.
  • Semantic Encoding by BERT: BERT uses transformer-based self-attention and masked-language-model training to learn contextual representations from whole input sentences.
  • Applications: Human-machine symbiosis applications include AI-assisted systems, interactive machine learning, and worker-AI collaboration.Examples span chatbots, healthcare assistance, human-in-the-loop learning, content moderation, and remote surgery.

3.2 SemCom for Recommendation

Recommendation-oriented SemCom distills user preference information for efficient wireless transmission and supports collaborative, content-based, demographic, and hybrid filtering approaches.

  • A recommendation SemCom system semantically encodes user information at an edge device and transmits it to support preference prediction.
  • Collaborative Filtering: Collaborative filtering extracts rating information from historical user activities to remove redundancy and compress data for transmission.
  • Collaborative Filtering: The server uses Pearson correlation between users’ jointly rated items to measure preference similarity and recommend items across similar users.
  • Collaborative Filtering: Increasingly sparse user-item matrices make similar-user clustering difficult because insufficient ratings are available.Proposed remedies include SVD, non-negative matrix factorization, clustering, and probability matrix factorization.
  • When explicit ratings are unavailable, low-dimensional item-rating encoders or high-rate access can reduce the transmission burden of preference inference.The alternatives include semantic subspace extraction, eMBB, and targeted mmWave deployment.
  • Recommendation filtering also includes content-based, demographic, and hybrid methods, with applications in social networks, healthcare, travel, media, and dating.

3.3 SemCom for Human Sensing and Care

Human sensing-and-care SemCom monitors health and movement through biomedical sensing, extracting useful signal statistics and prioritizing reliable, real-time delivery for abnormalities.

  • Human sensing-and-care uses sensors around or on humans to transmit health and movement data for server-based analysis and decision making.
  • Biomedical Semantic Encoding: Biomedical semantic encoding estimates informative signal statistics from periodic and predictable signals using time-domain or frequency-domain techniques.
  • Biomedical Semantic Encoding: ECG encoding can detect R peaks and QRS-wave characteristics to support heart-rate characterization and biological-activity analysis.
  • Abnormality detection requires real-time and highly reliable transmission, making layer-coupling preferable to a complex DNN-based design in this setting.
  • Elderly monitoring systems use multiple sensors to track activities such as cooking, dining, and sleeping, including applications for people with dementia.

3.4 SemCom for VR/AR

SemCom for VR/AR supports immersive human experiences by combining real-time multimedia processing with edge-assisted semantic encoding and transmission. These systems require high rate and low latency, with MEC offloading computation-intensive tasks near users.

  • VR replaces the physical world with a virtual one, whereas AR augments real objects with computer-generated perceptual information.
  • AR semantic encoding recognizes and tracks user-relevant physical objects before projecting icons, characters, or information onto them.The implementation cooperates between a device and an MEC server processing locally recorded raw video.
  • VR/AR connectivity requires high rate and low latency for real-time multimedia collection, processing, and transmission.Realistic FoV simulation requires at least 1 Gb/s, while immersive interaction requires motion-to-photon latency below 15 ms.
  • MEC addresses VR/AR demands by offloading tracking, mapping, and recognition and caching multimedia content at edge servers near users.
  • Layer-coupling can assign different DIIs to heterogeneous kinesthetic information, while collaborative systems can exploit PAI from AI models.
  • SemCom for VR/AR remains relatively new, despite extensive broader VR/AR application literature and recent wireless-networking studies.

4 Machine-to-machine Semantic Communication

M2M SemCom connects and coordinates machines to execute computation tasks effectively in wireless networks, especially across IoT applications. Its tasks span sensing, analytics, learning, reasoning, decision making, and actuation.

  • M2M SemCom efficiently connects multiple machines so they can execute a specific computation task in a wireless network.
  • Typical M2M SemCom applications involve sensing, data analytics, learning, reasoning, decision making, and actuation.

4.1 Distributed Learning

Distributed-learning SemCom targets the communication bottleneck created by uploading model updates from many devices. It combines local effectiveness encoding, aggregation-oriented modulation and multi-access, and importance-aware resource management.

  • Distributed learning trains an AI model from data and computation resources distributed across many mobile devices.
  • Federated learning protects mobile-data ownership by uploading local model updates or stochastic gradients instead of raw data.The server aggregates local updates and broadcasts the updated global model until convergence.
  • A ResNet-50 model contains 25.6 million parameters, or 1638.4 million float64 bits, making high-dimensional updates a communication bottleneck.
  • Effectiveness Encoding: Local effectiveness encoding updates the broadcast global model through multiple mini-batch gradient-descent steps before uploading the local model or gradient.The gradient is computed from a predefined loss function using back-propagation.
  • Effectiveness Encoding: Gradient compression retains large-magnitude elements or randomly deletes parameters, reducing communication overhead and potentially avoiding over-fitting.
  • Effectiveness Modulation and Multi-access: AirComp uses linear analog modulation and waveform superposition to aggregate distributed sensing or gradient signals over the air.For Gaussian sources and noise, it achieves the optimal rate-distortion tradeoff for a large number of sensors.
  • Effectiveness Radio Resource Management: Importance-aware RRM allocates more resources to training data with higher importance while preventing unimportant data from occupying channels.A DII can combine data importance and channel state, with scheduling designed to maximize the DII.

4.2 Split Inference

Split inference partitions inference between a resource-constrained device and an edge server, transmitting reduced-dimension features instead of redundant raw data. Its design balances local computation, communication overhead, model placement, and application-specific connectivity requirements.

  • Split inference offloads much of a trained model’s inference task from a mobile device to an edge server hosting a large-scale model.
  • Effectiveness Encoding: Effectiveness encoding extracts reduced-dimension features or feature maps from high-dimensional raw data, removing redundancy without compromising inference performance.Feature extraction can use PCA or neural-network representations.
  • Effectiveness Encoding: Feature-importance awareness assigns heterogeneous importance levels to features and guides quantization, resource allocation, and transmission order.More important features can receive higher resolution, more bandwidth or time slots, and earlier transmission.
  • SplitNet: In SplitNet, an effectiveness encoder extracts features, while an effectiveness decoder performs inference after channel decoding.Autoencoder-based channel components map features to analog symbols and reconstruct features for the inference decoder.
  • SplitNet: Moving the split point left simplifies on-device encoding but increases server decoding complexity, creating a computation-communication tradeoff.Resource-constrained devices therefore favor split points closer to the model input.
  • Mission-critical split-inference applications require URLLC, while other applications may prioritize close-to-human machine vision instead of latency.Remote autonomous-driving inference is expected to require 1 ms latency and near-100% communication reliability.

4.3 Distributed Consensus

Distributed consensus SemCom uses effectiveness encoding to transmit only state information needed for agents to reach agreement, reducing communication overhead in applications such as vehicle platooning and blockchains.

  • Effectiveness encoding extracts and transmits semantic information from agent states based on the consensus algorithm and its objective.
  • Vehicle Platooning: Vehicle platooning requires connected vehicles to share local states and coordinate braking and acceleration, but transmitting raw sensor data is impractical.An autonomous vehicle can generate several gigabytes of sensor data per second.
  • Vehicle Platooning: Platooning effectiveness encoders extract information essential for convergence, while encoder complexity and output size trade off against algorithm sophistication.Predicted trajectories can be compressed, and CNN encoders can extract traffic lights, lanes, and obstacles from video.
  • Vehicle Platooning: Vehicle platooning requires URLLC connectivity, and information latency is a more relevant coordination metric than conventional over-the-air latency.
  • Blockchains: Blockchain consensus can reduce communication overhead through effectiveness-based resource allocation, including PBFT schemes that restrict communication to inner-layer transactions.

4.4 Machine-vision Cameras

Machine-vision camera networks face severe communication bottlenecks from high-rate video and large camera populations. Effectiveness encoding addresses this by identifying regions of interest and allocating resources according to sensing relevance.

  • Machine-vision networks generate large raw-data volumes because 1080P video contains two million pixels per frame and can reach 60 frames per second.The resulting data rate can reach 100Mbps.
  • Effectiveness Encoding: Effectiveness encoding detects regions of interest in visual data and trims videos or images before streaming them to edge or cloud servers.
  • Effectiveness Encoding: A lightweight on-camera CNN scans image regions, records spatial ROIs, and retains frames containing at least one interesting object.
  • Resource Management: Effectiveness-aware resource management prioritizes sensing data associated with ROIs rather than delivering video bits indiscriminately.More bits can be allocated to high-resolution quantization of ROI pixels and fewer bits to background pixels.

5 KG based Semantic Communications

Knowledge graphs provide structured semantic representations that support efficient, robust, and task-aware SemCom across H2H, H2M, and M2M settings. They enable inference, error handling, semantic encoding, and knowledge sharing.

  • A knowledge graph represents entities and their relations in a semantic space, supporting interpretation and inference over facts.The paper models facts as triples consisting of a head node, relation, and tail node.
  • KG-based SemCom uses graph representations to improve communication efficiency and robustness, including error correction, behavioral constraints, and knowledge sharing.
  • KG Theory: Missing edges can make plausible facts absent from a finite KG, so scoring functions estimate fact plausibility using distance-based or semantic-similarity designs.
  • KG Theory: KG construction can use rule-based or data-driven approaches trained with loss functions such as margin-based, logistic-based, or cross-entropy-based objectives.
  • H2H SemCom: In H2H SemCom, KG knowledge can be injected into semantic encoders through entity embeddings mapped from input tokens.
  • H2M SemCom: In H2M SemCom, structured knowledge helps machines recognize entities and relations, supporting logical reactions in question answering, dialogue, recommendation, and robotics.
  • M2M SemCom: For M2M SemCom, dynamic KGs can represent connected-vehicle states and relationships, while semantic cores can translate messages across heterogeneous machine languages.
  • KG-Assisted SemCom: KGs used in autoencoders can improve decoding of semantic meanings distorted by communication channels and enhance SemCom performance when combined with source information.

6 Towards 6G Semantic Communication

6G is framed as a setting where SemCom supports connected intelligence, expansive services, and integrated communication, sensing, control, and computing. Future applications impose increasingly complex semantic representation and interaction requirements.

  • Future Services: Immersive XR depends on interpreting human movements, surroundings, intentions, and preferences so machines can generate desired content.
  • Future Services: Holographic and all-sense communication extend SemCom toward remote presence and multisensory interaction, while diverse sensing signals increase semantic representation complexity.
  • Future 6G applications require SemCom advances, while breakthroughs in 6G technologies may make those advances possible.
  • 6G Requirements: 6G aims for almost limitless connectivity, including bit rates up to 1 Tbps, latency below 100 microseconds, and massive device densities.
  • 6G Requirements: 6G research emphasizes comprehensive AI and a library of SemCom techniques that can support diverse optimization tasks and heterogeneous network configurations.
  • The paper positions SemCom as a potential primary area for 6G innovation and connects it with the convergence of physical and virtual worlds.
Loading 2110.00196v1…