Beyond Shannon: Transmitting Meaning

Claude Shannon's 1948 Mathematical Theory of Communication established the foundational framework for all modern wireless systems. Shannon's channel capacity theorem defines the maximum rate at which bits can be reliably transmitted over a noisy channel:

`

C = B · log2(1 + SNR)

`

This framework treats all bits equally — a bit representing a critical safety message has the same transmission cost as a bit representing a redundant background pixel. Shannon explicitly noted this limitation: his theory addressed the technical problem of reproducing symbols, not the semantic problem of conveying meaning.

Semantic communication breaks this paradigm. Instead of faithfully reproducing every transmitted bit, the receiver only needs to extract the intended meaning (or complete a downstream task) from the received signal. This shift can reduce transmitted data by 10–100× while preserving task-relevant information.

Weaver's Three Levels of Communication

Warren Weaver, in his 1949 commentary on Shannon's work, identified three levels of communication problems:

Traditional 5G NR operates exclusively at Level A: the PHY and MAC layers are optimized to minimize BER/BLER without any awareness of what the bits represent. Semantic communication operates at Level B (transmit meaning) or Level C (transmit only what the receiver needs to complete a specific task).

The Compression Opportunity

Consider transmitting a 1080p video frame of a highway intersection for autonomous vehicle monitoring:

The semantic and task-oriented approaches transmit orders of magnitude less data because they discard information irrelevant to the receiver's task. The tradeoff: they cannot reconstruct the original video frame. If the receiver only needs to make driving decisions (Level C), this is acceptable — even optimal.

Joint Source-Channel Coding (JSCC) with Deep Learning

Classical wireless systems use a separation architecture: source coding (compression) and channel coding (error protection) are designed independently. Shannon's separation theorem proves this is optimal in the limit of infinite block length. In practice, for finite block lengths and semantic objectives, joint source-channel coding (JSCC) can outperform separation.

Architecture Overview

A JSCC semantic communication system consists of:

`

Source → [Semantic Encoder] → [Channel Encoder] → Channel → [Channel Decoder] → [Semantic Decoder] → Task Output

(DNN) (DNN) (DNN) (DNN)

`

In a deep learning implementation, all four blocks are replaced by a single end-to-end trained neural network:

`

Source → [Encoder DNN (jointly learned)] → Channel → [Decoder DNN (jointly learned)] → Task Output

`

The encoder maps the source directly to channel symbols (complex-valued vectors transmitted over the air), and the decoder maps received symbols directly to the semantic output. No explicit source coding or channel coding stage exists — the network jointly learns compression and error protection.

Key Advantage: Graceful Degradation

Traditional separate coding exhibits a cliff effect: below a threshold SNR, the channel decoder fails catastrophically, and the output is useless. JSCC systems degrade gracefully — as SNR decreases, the semantic output quality diminishes smoothly rather than falling off a cliff. This property is documented in multiple studies and is particularly valuable for mobile channels with rapid SNR fluctuations.

DeepSC: The Reference Framework

The Deep Semantic Communication (DeepSC) framework, developed at Tsinghua University, is the most widely cited semantic communication architecture. DeepSC uses a Transformer-based encoder-decoder for text transmission:

DeepSC Architecture

Worked Example 1: DeepSC Compression vs Traditional

Compare DeepSC with a traditional Turbo-coded 16-QAM system for transmitting a 20-word English sentence at SNR = 5 dB:

`

Traditional system (Level A):

Average English word: ~5 characters = 40 bits (ASCII)

20 words: 800 bits

Turbo code rate 1/2: 1,600 coded bits

16-QAM (4 bits/symbol): 400 channel symbols

DeepSC (Level B):

Semantic encoder output: ~64 complex symbols (learned representation)

Channel symbols: 64 (no separate channel coding — jointly learned)

Compression ratio: 400 / 64 = 6.25×

`

At SNR = 5 dB, DeepSC achieves a BLEU score of 0.85 (semantic similarity metric) while the traditional system achieves BLEU of 0.72 (due to residual bit errors after Turbo decoding that corrupt word boundaries). At SNR = 0 dB, the gap widens: DeepSC BLEU = 0.68, traditional BLEU = 0.15 (Turbo code cliff effect).

Worked Example 2: Bandwidth Savings for Video Surveillance

A smart city deploys 1,000 cameras streaming 1080p/30fps video to a central traffic management AI. Compare bandwidth requirements:

`

Traditional (H.265 + 5G):

Per camera: 4 Mbps (H.265 medium quality)

Total: 1,000 × 4 Mbps = 4 Gbps aggregate uplink

Task-oriented semantic (transmit scene graph only):

Per camera: semantic encoder extracts objects, positions, velocities

Encoded scene graph: ~2 KB per frame

Per camera: 2 KB × 30 fps × 8 = 480 kbps

Total: 1,000 × 480 kbps = 480 Mbps aggregate uplink

Bandwidth reduction: 4,000 / 480 = 8.3×

If only anomaly events need reporting (task Level C):

Per camera (average): 50 kbps (transmit only when events detected)

Total: 1,000 × 50 kbps = 50 Mbps

Bandwidth reduction: 80×

`

The 80× reduction for task-oriented communication transforms the infrastructure requirement from a dedicated fiber backhaul to a standard 5G connection per camera cluster.

Metrics: Beyond BER and BLER

Semantic communication requires new performance metrics because BER/BLER (Level A metrics) do not capture meaning preservation:

The shift from Level A to Level B/C metrics has profound implications for PHY layer design. A semantic-aware scheduler might deprioritize retransmissions for data that the receiver's AI model can reconstruct from context, while prioritizing novel or surprising information that changes the semantic state.

Task-Oriented Communication

Task-oriented communication (Level C) represents the most extreme form of semantic communication. The transmitter and receiver share a common task objective (e.g., classify an image, control a robot), and only information relevant to that task is transmitted.

Architecture

`

Sensor Data → [Task Encoder] → Channel → [Task Decoder] → Task Decision

Extracts only Directly outputs

task-relevant classification/

features control signal

`

The task encoder is trained end-to-end with the task decoder and a differentiable channel model. The encoder learns to extract and transmit only the features that the decoder needs for the specific task, discarding everything else.

Feature Compression Performance

Challenges and Limitations

Training Data and Generalization

Semantic encoders trained on specific data distributions (e.g., English text, highway scenes) fail to generalize to out-of-distribution inputs. A DeepSC model trained on news articles performs poorly on medical texts. This lack of generalization is documented extensively in the literature and represents the most significant barrier to deployment.

Potential solutions include:

• Foundation model encoders (GPT-class models as semantic extractors) — but these are computationally expensive for real-time PHY processing
• Few-shot adaptation of semantic codecs to new domains
• Modular architectures that separate domain-specific semantic extraction from domain-agnostic channel adaptation

Standardization Gap

No 3GPP specification addresses semantic communication. The concept sits at the intersection of PHY, application layer, and AI — a cross-layer design that does not fit cleanly into the current 3GPP stack architecture (per TS 38.300 overall description). Potential standardization paths include:

The application-layer approach — where semantic encoding/decoding is performed above the NR stack — requires no specification changes and can be deployed today. However, it cannot exploit cross-layer optimization (e.g., semantic-aware HARQ, priority scheduling). The ITU-R IMT-2030 framework (ITU-R M.2160) includes AI-native air interface as a design principle, which could eventually encompass semantic PHY.

Real-World Research

Tsinghua University — DeepSC and Extensions

The Tsinghua University research group (Prof. Zhijin Qin, Prof. Geoffrey Ye Li) developed the original DeepSC framework and has extended it to:

• DeepSC-ST (speech-to-text): semantic communication where the transmitter sends speech and the receiver outputs text, achieving 10× compression vs Opus + ASR pipeline
• DeepSC-VQA (visual question answering): transmitter sends an image, receiver answers questions about it without reconstructing the image
• Multi-user DeepSC: NOMA-based semantic multiple access for up to 8 users sharing the same channel resources

Published results show DeepSC outperforming traditional separation-based systems by 3–8 dB in required SNR for equivalent semantic accuracy across AWGN, Rayleigh, and Rician fading channels.

Samsung Research — Task-Oriented Video

Samsung Research published work on task-oriented video communication for autonomous driving, where a roadside camera transmits only driving-relevant features to connected vehicles:

• Object detection accuracy: 94.2% mAP using task-oriented encoding vs 95.1% mAP using full H.265 video — only 0.9% accuracy loss
• Bandwidth reduction: 50× compared to H.265 at equivalent detection performance
• Latency: <5 ms end-to-end (encoder + channel + decoder) on edge GPU vs ~30 ms for H.265 encode/decode + YOLO inference

Samsung presented these results at IEEE ICC 2024 and has filed patents covering task-oriented encoding for V2X standardization in 3GPP SA1 (service requirements).

3GPP Outlook: Potential Rel-20+ Study Items

While no formal study item exists for semantic communication in 3GPP, multiple RAN1 contributions (from Samsung, Huawei, ZTE, and Qualcomm) have proposed:

• AI/ML-based JSCC as an extension of the TR 38.843 study on AI for NR air interface
• Semantic-aware scheduling where the MAC layer uses content importance metrics from higher layers
• Task-oriented HARQ that skips retransmission of semantically redundant data

The earliest formal study item is anticipated for Release 20 (3GPP work plan ~2027), with normative specifications possible in Release 21–22 (2029–2031), aligning with the 6G deployment timeline.

Key Takeaway: Semantic communication shifts the design objective from faithfully reproducing bits (Shannon's Level A) to preserving meaning (Level B) or completing tasks (Level C). Joint source-channel coding with deep learning — exemplified by Tsinghua's DeepSC framework — achieves 6–100× compression over traditional separation-based systems while maintaining semantic fidelity. Samsung's task-oriented video demonstrates 50× bandwidth reduction with under 1% accuracy loss for object detection. The concept is technically viable today at the application layer, but native PHY integration awaits 6G standardization in the 2029–2031 timeframe, driven by ITU-R IMT-2030's AI-native air interface vision.