As wireless communications approach the Shannon physical-layer capacity limit, researchers are looking toward semantic communications for 5G and future 6G networks. However, a critical distinction must be made: simply using Deep Neural Networks (DNNs) for feature extraction—a process often called "neural coding"—is not equivalent to true "semantic coding." This paper from researchers at the Beijing University of Posts and Telecommunications proposes a standardized workflow to bridge this gap.
The Problem with Neural Coding
Current "semantic" systems often rely on DNNs to extract statistically sound features. While these features reduce dimensionality, they often lack context-awareness and fail to capture "semantic equivalence." For instance, a model might fail to recognize different dog breeds as the same high-level concept. True semantic coding requires moving beyond raw feature extraction to a representation that understands meaning and context, allowing for more compact and general representations.
A Standardized Semantic Coding Workflow
The authors propose a three-stage framework designed to achieve higher transmission efficiency and better visual realism, particularly at low bit rates. This workflow decouples the semantic representation from the physical transmission constraints.
The workflow consists of three main stages:
- Tokenization: Raw data is decomposed into discrete machine-processable units (tokens). Using attention-aided models like Vision Transformers (ViT), the system identifies regions of high importance.
- Reorganization: This is the core innovation of the paper. Instead of maintaining a high number of tokens (e.g., the 196 tokens standard in a ViT-B/16 model), the system uses similarity-based consolidation (such as bipartite soft matching) to merge semantically related tokens. This reduces the representation to a much smaller set of "concepts" (e.g., 10–30 tokens).
- Quantization: These compact tokens are mapped to discrete codes from a codebook, further reducing bandwidth and enabling compatibility with auto-regressive generators.
Key Insights: "One Code, One Concept"
The ultimate goal of this workflow is a "One Code, One Concept" state, where a single latent code represents a distinct semantic concept, drastically reducing redundancy. This approach also addresses the "Consistency-Realism Trade-off." In human-centric communication, semantic coding allows for better control over how much the output looks like the specific input (consistency) versus how natural it appears (realism).
Furthermore, the architecture is decoupled. By extracting general semantic representations first, the same transmitted data can serve different purposes at the receiver. A generative model can reconstruct a high-fidelity image for a human user, while a simple classifier can perform machine tasks directly from the semantic tokens.
Results and Impact
In a case study conducted over an Additive White Gaussian Noise (AWGN) channel, the proposed method demonstrated significant advantages over traditional neural coding baselines:
- Efficiency: The system successfully reduced the token count from 196 to just 10–30 while maintaining satisfactory reconstruction quality.
- Performance: The semantic coding system outperformed baselines in both PSNR (fidelity) and FID (realism) across all tested Channel Bandwidth Ratios (CBR), with the most pronounced gains at very low bit rates.
- Speed and Accuracy: The system achieved an inference throughput of 524 images per second (FP16), which is nearly 1.5x faster than a vanilla ViT-B/16 model, with only a negligible (5%) drop in Top-1 accuracy.
Conclusion
As generative AI continues to evolve, semantic coding is poised to become the primary catalyst for generative communication. By focusing on sparse semantic anchors rather than raw bits, this standardized workflow enables a future where receivers can "imagine" details based on core meanings, ensuring robust communication even in bandwidth-constrained environments.
Summary updated