6G Networks: Beyond Shannon Towards Semantic and Goal-Oriented Communications

Motivation

• Although 5G remains in early deployment, future individual and societal needs are expected to require B5G and ultimately standardized 6G technologies.Some solutions may extend 5G, while others require a radical technological change.Source: §1 · p. 2

Background

• The paper motivates this vision against the Shannon-based communications tradition, whose advances include MIMO, waveform design, interference mitigation, and key 5G architectural technologies.These technologies include traffic engineering, NFV, SDN, and network slicing.Source: §1 · p. 2

Motivation

• 6G must move beyond Shannon’s technical communication problem because expanding services and data-rate demands will encounter spectrum, energy, and other resource bottlenecks.Higher carrier frequencies can provide wider bandwidth but introduce blocking, atmospheric absorption, and reduced power-amplifier efficiency.Source: §1 · pp. 2–3

Paradigm shift

• The proposed paradigm prioritizes successful task execution and semantic effectiveness over error-free symbol delivery, focusing on what information must be transmitted rather than only how.This follows Shannon and Weaver’s distinction between technical, semantic, and effectiveness problems.Source: §1 · pp. 3–4

Proposed vision

• Semantic and goal-oriented communication can discard irrelevant data, reducing bandwidth, delay, and energy while preserving information needed to convey meaning or fulfill a predefined goal.The paper presents this as a central design strategy for effective and sustainable 6G networks.Source: §1 · pp. 3–4
• The paper envisions 6G networks incorporating semantics and effectiveness aspects, inspired by brain-like learning, rapid decision-making, and sustainable energy consumption.The network should support cognition and effective task execution among distributed nodes.Source: §1 · pp. 3–4

Explainability

• Explainable machine learning is important because deep neural networks can perform well without making clear why their outputs follow from inputs or learned weights [5].Explainability should relate input features to outputs, although an in-depth review is outside this paper’s scope.Source: §1 · p. 4

2.1. 6G Use Cases and KPIs

• 6G KPIs will combine 5G continuity with new measures for intelligent, computation-oriented and sustainable networks, including decision reliability, goal-related energy use and learning reliability.Only locally relevant KPI subsets need be achieved, with a targeted 10–100-fold improvement across generations and a possible 100-fold capacity increase.Source: §2.1 · pp. 6, 9, 10
• 6G use cases span extended reality, holographic and multi-sense communication, industrial automation, autonomous systems, massive sensing, intelligent production and three-dimensional mobility services.Holographic links may require several Tb/s per link, while communicate-and-compute services require stringent latency, jitter and reliability; peak mobility may reach 1000 Km/h.Source: §2.1 · pp. 6, 8, 9

2.2. 6G New Services

• 6G new services comprise MMTCxDI for pervasive distributed intelligence, GeMBB for on-demand rate-hungry applications, URLLCCC for edge computation and end-to-end control, and semantic services connecting natural and artificial intelligences.Semantic services support H2H, H2M and M2M knowledge sharing, while URLLCCC extends reliability and latency requirements to computation measures such as learning accuracy and correct classification probability.Source: §2.2 · pp. 11–12

• Legacy wireless generations prioritized reducing reception uncertainty through higher capacity, reliability, lower latency, and broader bandwidth, while energy consumption increasingly constrained practical operation.Since 4G, research has approached Shannon limits while optimizing available resources.Source: §3 · p. 12
• 5G established communicate-and-compute systems in which edge-cloud support links sensing, computing, control, and actuation in a continuous loop.This architecture lays foundations for intelligent machines.Source: §3 · pp. 12–13
• 6G will require advances that make knowledge and decision-making network commodities as information accumulates faster than natural or artificial intelligence can filter, transmit, and process it.The envisioned shift changes conventional notions of knowing, learning, guessing, and discovering.Source: §3 · p. 13
• Under a degrowth principle, 6G should deliver more intelligent mobile services through qualitatively better resource use rather than an endless race for capacity, infrastructure, or energy.The vision may retain millimeter-wave technologies while complementing them with sub-Terahertz and Visible Light Communications, and draws on Shannon and Weaver’s three communication levels.Source: §3 · p. 13

Motivation

• 6G must address technical transmission, semantic interpretation, and effective goal accomplishment because semantic and effectiveness concerns are increasingly central to networked human–machine services.Shannon formalized the technical problem while deliberately leaving semantic and effectiveness aspects aside; the proposed evolution incorporates all three levels, as also suggested in.Source: Level A. How accurately can the… · pp. 13–14

Three-level architecture

• The proposed 6G architecture adds semantic and effectiveness levels above a virtualized technical stack, enabling cross-layer coordination and intelligent allocation of network and computation resources.The effectiveness level orchestrates lower layers to meet service KPIs, while use cases without semantics may connect it directly to the technical level.Source: Level A. How accurately can the… · pp. 14–15

Semantic communications

• Semantic communication targets correct interpretation of concepts rather than error-free bit recovery, exploiting shared prior knowledge and formal languages to correct symbolic errors.Shared knowledge may include human language or entities and logical rules, with interactions between natural and artificial intelligence also relevant.Source: Level A. How accurately can the… · pp. 15–16

Goal-oriented communications

• Goal-oriented communication focuses system design on accomplishing a joint goal within time and resource constraints, allowing irrelevant information to be omitted.The effectiveness level uses semantic information when appropriate and orchestrates network, computation, control, and actuation resources.Source: Level A. How accurately can the… · p. 16

Online learning-based communication and control

• Online machine learning can adapt traffic, coding, decoding, and scheduling from network monitoring while communication infrastructure distributes learning tools under stringent delay constraints.This requires joint orchestration of computation, communication, storage, and control resources.Source: Level A. How accurately can the… · p. 16

4. Semantic Communications

• Semantic information concerns a message’s meaning, not merely the probabilistic structure of its encoding symbols.A message communicates semantically when its destination recovers meaning equivalent to the source’s and increases its knowledge; semantic content depends on source and destination knowledge.Source: §4 · p. 17, §4.1 · p. 18

4.1. Semantics and knowledge representation systems

• Knowledge representation systems encode domain concepts, relations, facts, rules, and constraints, while reasoning engines answer application-domain questions despite incomplete knowledge and time-limited inference.Graph-based knowledge representations use nodes for entities and edges for relations; incompleteness can make answers unavailable or only reliable to a certain degree.Source: §4.1 · p. 19

4.2. Semantic source and channel coding

• Semantic coding maps messages to potentially multiple equivalent symbol sequences or resolves many-to-one ambiguity, then decodes for meaning rather than exact symbol recovery.Syntactic errors arise from channel noise or interference, whereas semantic errors can result from knowledge-base differences or misinterpretation; semantic decoding may recover correct content despite some symbol errors by exploiting language rules.Source: §4.2 · pp. 21–24

Example of application to text transmission

• At low SNR, DeepSC significantly outperforms conventional coding schemes and an E2E DNN on semantic sentence similarity over AWGN and Rayleigh channels.DeepSC jointly learns semantic and channel encoding to minimize semantic error while reducing transmitted symbols; the comparisons are reported in Fig. 4.Source: Example of application to text… · p. 25, Figure 4 · p. 26

4.3. Feedback and cross-layer interaction in semantic communications

• Semantic feedback can request message retransmission or a different formulation when meaning is unclear, while cross-layer feedback can reduce source data rates when context enables prediction.This interaction can preserve recovery accuracy while saving energy or bandwidth.Source: §4.3 · pp. 26–27

4.4.3. Holographic communications

• Semantic gains require additional receiver computation and delay, motivating predictive architectures that transmit mainly deviations from expected sensory inputs.Shared knowledge enables error correction and omission of predictable details, but receiver complexity remains a significant limitation.Source: §4.3 · p. 27, §4.4.3 · p. 29

4.4.1. Speech signals

• Speech communication can use speech recognition, NLP correction, and contextual reconstruction to compensate bit errors or recover speech lost during deep fades.These mechanisms reduce pressure on forward error correction and can avoid retransmissions.Source: §4.4.1 · p. 28

4.4.2. Video streaming

• Video interpreters can segment scenes and predict missing frames, reconstructing semantically equivalent content during channel fades while saving transmit power or bandwidth.The reconstructed video need not equal the transmitted video if the overall event flow remains unchanged.Source: §4.4.2 · p. 28

4.4.3. Holographic communications

• In holographic communications, shared digital models can reconstruct missing scene views in real time, reducing the data rate needed for the desired quality of experience.This approach exploits common knowledge between source and destination.Source: §4.4.3 · pp. 28–29

Motivation

• Goal-oriented communication transmits only information strictly relevant to fulfilling a specified goal, evaluating performance by goal effectiveness given communication resources.The framework supports trade-offs such as power consumption versus delay and accuracy versus delay, including edge-learning applications,, [62], [63].Source: §5 · pp. 29–30

Framework

• The proposed framework adds feedback from the decision maker to the source encoder for classification or parameter learning from observations.The encoder maps measurements X to transmitted data Z while satisfying service requirements such as decision accuracy and delay, potentially minimizing energy.Source: §5 · pp. 30–31

Goal-oriented encoding

• Source encoding is optimized for learner accuracy and decision time rather than reconstruction distortion, seeking a maximally compressed representation that preserves recovery of the target parameter.This differs from conventional lossless or lossy compression, which prioritizes reconstructing observed data.Source: §5 · p. 31

Sufficient statistics

• Minimal sufficient statistics can reduce transmitted bits substantially without sacrificing inference, because their entropy may be much smaller than the entropy of the original data.For parameter estimation, sending f(X) instead of X preserves all information relevant to recovering the parameter vector.Source: §5 · p. 32

Information bottleneck

• When minimal sufficient statistics are difficult to obtain, the information bottleneck method balances information loss about the target parameter against the number of encoded bits.Small β favors compression at the expense of learning accuracy, whereas large β favors accuracy at the expense of compression; iterative optimization is possible when probability densities are known.Source: §5 · p. 32

Adaptive implementation

• Auto-encoders can approximate the information-bottleneck solution when data statistics are unknown, while online adaptation of β can balance energy consumption, end-to-end delay, and learning accuracy.The adaptation depends on channel state and the learner’s achieved accuracy.Source: §5 · p. 33

Goal-oriented edge learning

• At the same E2E delay, lowering target accuracy from 94% to 91% achieves substantial energy savings, while ensemble learning improves the overall energy-delay-accuracy trade-off.The comparison evaluates ensemble learning with two learners against single-learner operation under two different energy-and-E2E-delay objective combinations.Source: Example of application: Edge online… · p. 34
• The method dynamically adjusts quantization bits, transmit power, and edge-server CPU scheduling to minimize energy and classification accuracy under average E2E-delay constraints.It uses feedback from receiver-side classification performance to adapt source encoding over a fading wireless channel.Source: Example of application: Edge online… · pp. 33–34
• Classifier-output entropy estimates confidence and switches online between parallel SVM and MLP learners when classification correctness cannot be directly evaluated.Low entropy indicates strong class preference, whereas high entropy indicates classifier confusion.Source: Example of application: Edge online… · p. 34

DNN-based edge inference

• Splitting a DNN after its compressive initial layers can reduce transmitted bits by performing feature extraction at the source while retaining roughly the same overall performance [70].The approach relies on observed mutual-information reduction and progressive geometric clustering during training.Source: Example of application: Edge online… · pp. 34–35

Video analytics

• Goal-based video analytics can lower camera-to-edge data rates by filtering irrelevant frames through staged background, target-object, and object-count screening.FFS-VA is a three-stage pipelined system for multistage video analytics, potentially operating in real time.Source: Example of application: Edge online… · p. 35

Semantic and goal-oriented communication

• Combining semantic and goal-oriented communication can transmit only information strictly relevant to goal achievement, reducing data without affecting goal-accomplishment accuracy or reliability.This extends the edge-learning example beyond quantization-based source adaptation.Source: Example of application: Edge online… · p. 35

6. Online Learning-Based Communication and Control

• Merging machine learning with semantics can improve learning through semantic features, context-based disambiguation, and robustness against adversarial attacks while identifying relevant information more efficiently.The network is envisioned as AI-native, with intelligence at the edge supporting both machine learning for network functions and communication for distributed learning,, [73],.Source: §6 · pp. 35–36

6.1. Holistic Orchestration of C4 Resources

• Joint communication, computation, caching, and control orchestration is required because delay-critical service time couples these resources, whose limited edge availability demands coordinated optimization.Online convex optimization addresses settings where the system state and objective are not fully known at each slot, while proactive caching also supports software needed to run applications near users.Source: §6.1 · pp. 36–37

6.2. Machine learning for wireless networks

• Online learning intertwines training and testing as wireless channels evolve, using reinforcement learning or stochastic optimization to adapt actions from current observations.Wireless learning can use simulators to generate large labeled datasets, while graph and topological models extend learning beyond pairwise relations for complex network data.Source: §6.2 · pp. 38–39

6.3. Federated learning

• Federated learning preserves privacy by exchanging local parameter estimates rather than data, converging to the global optimum under convexity conditions but requiring multi-task variants for heterogeneous devices.Channel, behavioral, and model heterogeneity can affect aggregation accuracy and convergence time; different device groups may require distinct estimates wk,.Source: §6.3 · pp. 39–40

6.4. Intelligent content delivery at the edge

• Proactive edge caching predicts demand to optimize cache deployment, content placement, and routing, while semantic delivery additionally requires migrating knowledge bases and virtual machines near users.Distributed learning and content delivery face bandwidth, straggler, privacy, and security bottlenecks; coded distributed computing injects structured computational redundancy to alleviate some of them.Source: §6.4 · pp. 40–41

6.5. Semantic machine learning

• Semantic machine learning addresses brittle correlation-based decisions by incorporating external knowledge, context, logical rules, and reasoning into learning processes.6G semantic networks can place knowledge bases across network nodes, enabling more efficient and reliable learning through semantic interpretation, transfer learning, and knowledge sharing.Source: §6.5 · pp. 41–42

Vision

• Semantic and goal-oriented communications could improve effectiveness and reliability without necessarily increasing bandwidth or energy by identifying information strictly necessary for meaning or goal accomplishment.The approach selectively transmits, processes, infers, and remembers information relevant to goals defined by interacting parties.Source: §7 · pp. 42–43

Sustainability

• The vision makes sustainability a central property of future networks by replacing the trend of continually allocating more energy or bandwidth for increasingly sophisticated services.Source: §7 · p. 43

Challenges

• Implementing this approach requires distributed computing mechanisms that learn meaning from data, use knowledge representation, and identify strictly relevant information for goal-oriented communications.Source: §7 · p. 43

Learning

• Semantic aspects can shift learning from purely inductive strategies toward interplay between inductive and deductive mechanisms, using abstract models to guide subsequent learning.This combines learning from examples with model building in a manner compared with human-brain learning.Source: §7 · p. 43