Semantic Communications: Theory & Applications

• Semantic communications are defined as transmitting intended meanings instead of raw bits, emphasizing context and task relevance through AI.
• Architectures integrate deep learning-based semantic encoding, joint semantic-channel coding, and adaptive resource management for optimized performance.
• Practical implementations demonstrate improved video resolution, machine translation, and autonomous driving while reducing bandwidth usage.

Semantic communications represent a paradigm shift from the Shannon doctrine of symbol-level fidelity to the task-oriented and meaning-centric transmission of information. This discipline focuses on conveying the semantic content—the intention, context, or actionable meaning—underlying the source signal, rather than preserving its precise syntactic form. Fueled by advances in deep learning and AI, semantic communications leverage knowledge bases, adaptive coding, distributed computing, and reasoning frameworks to achieve efficiency and expressiveness beyond traditional communication systems. Semantic metrics such as semantic entropy, semantic similarity, and semantic rate supersede bit error rate as the main evaluation criteria, guiding design choices for both theoretical modeling and deployed architectures.

1. Theoretical Foundations of Semantic Communications

Semantic communications generalize Shannon’s framework to meaning-level metrics, superseding raw symbol or bit recovery. The semantic random variable $S$ encodes the “intended meaning,” and the central objective is to maximize $I(S;\hat S)$ or minimize semantic distortion $d_{\mathrm{sem}}(S,\hat S)$ at the receiver (Qin et al., 2021).

Semantic entropy quantifies uncertainty at the meaning level: $H_{\mathrm{sem}}(S) = -\sum_{s} p(s)\log p(s)$ This leads to definitions of semantic rate $R_s = H_{\mathrm{sem}}(S)/n$ and semantic channel capacity: $C_{\mathrm{sem}} = \max_{P_X} I(S;Y)$ where $S = f_{\mathrm{sem}}(X)$ is the semantic embedding produced by the transmitter.

Rate-distortion for semantics extends classical theory: $$R(D) = \min_{P(\hat S|S)} I(S;\hat S) \quad\text{s.t.}\quad \mathbb{E}[d_{\mathrm{sem}}(S,\hat S)] \le D$$ In “From Philosophical Conceptions Towards a Mathematical Framework,” the semantic capacity is shown to include an additional $H(X|S)$ term under physical channel noise, allowing message-packing density beyond Shannon’s $I(X;Y)$ bound when only meaning recovery is required (Gholipour et al., 2 May 2025).

Semantic distortion metrics are application-dependent; for classification tasks, cross-entropy $d_{\mathrm{cls}}(L,\hat L)$ is appropriate, while perceptual differences for generative tasks may be captured via KL divergence on distributional representations (Wu et al., 2024).

This theoretical expansion enables the quantification of semantic efficiency, trade-offs between compression and task effectiveness, and the development of capacity limits adapted to meaning-centric transmission.

2. System Architectures and Computational Frameworks

End-to-end semantic communication stacks integrate knowledge bases, semantic encoders/decoders, joint semantic-channel coding, and resource-adaptive protocols. Figure 1 in “Computing Networks Enabled Semantic Communications” illustrates a modern three-tier architecture:

• Cloud Tier: Trains deep models (semantic samplers, JSCC codecs), hosts “heavy” inference (e.g. transformer-based reasoning, super-resolution), and manages centralized storage.
• Edge Tier: Performs model partitioning, fast feedback, low-latency inference, and intermediates coordination between resource-constrained devices and cloud.
• End Tier: Implements lightweight semantic sampling and compressed JSCC coding, with offloading to cloud/edge as dictated by task computational/latency requirements (Qin et al., 2023).

Coordination occurs via a global resource pool, with cross-layer managers performing scheduling, offloading, and computational power perception. Semantic codecs adaptively incorporate real-time computing-network status vectors $c$ as side-input, allowing all key components—$f_{\mathrm{sem}}(x,c)$, $f_{\mathrm{ch}}(s,c)$, $g_{\mathrm{ch}}(y,c)$, $g_{\mathrm{sem}}(s,c)$—to optimize based on available resources.

Hybrid architectures such as HybridBSC facilitate deployment within existing infrastructure by embedding semantic information into conventional bit-level streams, with semantic insertion/extraction executed via transforms (DWT, DCT, SVD) and CNN-based semantic encoders/decoders (Yu et al., 2024).

Explainable frameworks enforce interpretable semantic representations, e.g., via $\beta$-VAE-based disentangled encoding, feature selection, and robust semantic channel modeling, remaining compatible with bit-level systems and supporting real-time edge deployment (Ma et al., 2023).

3. Key Technologies: Semantic Sampling, Coding, and Reasoning

Semantic Sampling and Reconstruction

Two-stage semantic sampling maximizes task-relevant content under sampling budgets. Programmable sensors activate binary masks $m(\phi)$ optimized via loss minimization: $$\min_{\phi}\; \mathbb E_{x, y_{\mathrm{task}}}\!\Bigl[\ell_{\mathrm{task}}(y_{\mathrm{task}}, g_\theta(S_\phi(x))) + \lambda\,\|S_\phi(x)\|_0\Bigr]$$ Here, $S_\phi(x)=m(\phi)\odot x$ denotes masked sampling; $g_\theta$ reconstructs the signal; $\lambda$ controls fidelity vs. sample count (Qin et al., 2023).

Downstream, learned low-pass filtering and downsampling preserve essential semantics while enabling compact transmission: $x_{\mathrm{LR}} = D\bigl(h_\phi * S_\phi(x)\bigr)$ Training minimizes a reconstruction loss (MSE, perceptual, or task-driven) with regularization.

At the receiver, task-oriented reconstruction comprises super-resolution, bidirectional RNN propagation, and optical-flow refinement for video applications.

Semantic-Channel Coding (JSCC)

Semantic–channel coding jointly maps semantic features to channel symbols. End-to-end architectures optimize dual losses: $$\mathcal{L}(\theta) = \mathbb{E}_{x}[\ell_{\mathrm{sem}}(f_{\mathrm{sem}}(x),g_{\mathrm{ch}}(H(f_{\mathrm{ch}}(f_{\mathrm{sem}}(x))))) + \lambda \ell_{\mathrm{rec}}(x, g_{\mathrm{sem}}(g_{\mathrm{ch}}(\cdot)))]$$ Information-theoretic performance is governed by achievable rate–distortion formulas and semantic capacity bounds, adapted via side-information reflecting available computing resources.

Explicit semantic base architectures represent atomic “Sebs” $b_k$, organizing them as a poset to enable hierarchical, intent-specific, explainable coding. KB update mechanisms allow for robust adaptation to changing scenarios without retraining large NNs (Wang et al., 2024).

Knowledge-graph and reasoning frameworks embed relational semantics as triplet graphs, using energy-based or margin-ranking methods and lifelong learning to address evolving semantic environments and hidden-entity inference (Liang et al., 2022).

4. Resource Allocation, Optimization, and Multi-User Access

Semantic-aware resource allocation leverages joint optimization over transmission (power $p_i$, bandwidth $B_i$) and computational ($f_i$, offloading $\alpha_i$) resources: $$\min_{\{p_i, f_i, \alpha_i\}} \sum_{i=1}^K(E_i^{\mathrm{comp}}(f_i,\alpha_i) + E_i^{\mathrm{tx}}(p_i))$$ subject to delay, semantic rate, and power/frequency constraints (Qin et al., 2023). Centralized methods include Lagrangian dual-decomposition; distributed methods employ multi-agent RL, e.g. MAPPO.

In heterogeneous multi-user networks, semantic and bit users are jointly scheduled using hybrid OMA–NOMA strategies. Rate regions are defined as: $$\mathcal{R} = \{(R_s,R_b): \exists (W,p_s,p_{b,N},p_{b,O}) \text{ satisfying constraints}\}$$ Numerical results demonstrate semi-NOMA strictly expands achievable regions compared to pure OMA/NOMA, particularly under channel asymmetry (Mu et al., 2022).

Opportunistic semantic–bit switching further improves ergodic rates for secondary users under primary user constraints.

5. Practical Applications and Prototype Demonstrations

Semantic communications have demonstrated significant gains in diverse applications:

• Video Super-Resolution: End devices downsample and select key frames, cloud-based RNN upsamplers reconstruct detailed output, delivering higher PSNR/SSIM at reduced bpp (Qin et al., 2023).
• Neural Machine Translation: DeepSC encoders offload semantic symbols to edge, with joint optimization yielding 20–30% user-side energy savings.
• IoV Autonomous Driving: Semantic segmentation and graph-based feature packing enable low-latency, accurate decision support and traffic management (Ye et al., 3 Mar 2025).
• Edge-Cloud Deployment: Hybrid and explainable frameworks (HybridBSC, $\beta$-VAE) provide real-time, resource-efficient semantic transmission on wireless platforms (Ma et al., 2023, Yu et al., 2024).
• World Model-Aided Video: WFM-based semantic prediction, segmentation-assisted partial transmission, and active scheduling reduce bandwidth by 50–80%, maintaining visual and semantic fidelity (Jiang et al., 27 Oct 2025).

6. Security, Interpretability, and Standardization

Security in semantic communications entails unique challenges—attackers target semantic information and ML models. Techniques include:

• Information bottleneck penalization of sensitive semantic leakage: $\min I(X;Z) - \beta I(Y;Z) + \gamma I(T;Z)$
• Adversarial training, model watermarking, and physical-layer secrecy via semantic-aware beamforming (Yang et al., 2023).

Interpretability is advanced by explicit semantic bases and disentangled $\beta$-VAE codebooks, allowing inspection and rational resource allocation for robustness and adaptation (Wang et al., 2024, Ma et al., 2023).

Networking principles such as separation of concerns are reconciled via standardized semantic interfaces, embedding models, task descriptions, and semantic SLAs, supported by network control plane extensions for semantic resource negotiation (Lampin et al., 25 Feb 2025).

7. Open Problems and Future Directions

Key research directions include:

• Formal semantic information theory: establishing universal metrics, distortion-cost regions, and capacity limits under joint communication-computation constraints (Shao et al., 2022, Gholipour et al., 2 May 2025).
• Multimodal semantic fusion: integrating semantic bases across image, text, audio, and sensor modalities (Wang et al., 2024, Ahmed et al., 13 Jun 2025).
• Scalable KB management: distributed and federated updates, privacy-preserving techniques, and robust alignment against semantic noise (Ye et al., 3 Mar 2025).
• Semantic resource allocation: real-time RL-based scheduling, semantic prioritization, and energy-latency-task trade-offs.
• Robustness and security: adversarial resilience, semantic-layer encryption, feature selection under privacy constraints.
• Standardization: interoperable APIs, task-driven codebook optimization, and cross-layer semantic signaling primitives (Lampin et al., 25 Feb 2025).

Semantic communications thus transcend bit-level constraints by optimizing for meaning under practical system, resource, application, and security requirements, positioning the discipline as foundational to future 6G intelligent networks.