Not All Symbols Are Equal: Importance-Aware Constellation Design for Semantic Communication

2026-05-14 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Communication Systems & Signal Processing · Depth: Expert, quick

Summary

A novel joint semantic-physical layer framework for goal-oriented semantic communication systems is proposed, addressing the issue of critical symbols being equally vulnerable to channel errors as irrelevant ones in existing decoupled approaches. This framework integrates a vector quantized-variational autoencoder for discrete latent concept extraction, a semantic criticality indicator (SCI) to score concept relevance, and a deep reinforcement learning agent for dynamic transmission subset selection based on channel conditions. At the physical layer, a learned semantic-aware M-QAM constellation assigns symbol positions based on joint co-occurrence statistics and SCI scores, diverging from standard M-QAM's uniform spacing and Gray coding. The system introduces semantic symbol vulnerability (SSV) and semantic protection probability (SPP) metrics, demonstrating near 100% SPP for the proposed constellation across 4-QAM to 1024-QAM, compared to 50% for standard constellations, achieving a 21:1 compression ratio with semantic quality above 0.9 on MNIST, Fashion-MNIST, and FSDD datasets.

Key takeaway

For AI Scientists designing communication systems where data importance varies, you should consider integrating semantic criticality directly into physical layer constellation design. This approach significantly enhances the protection of task-critical information against channel errors, achieving near 100% semantic protection probability and high compression ratios, outperforming traditional Gray-coded constellations. Evaluate your system's semantic symbol vulnerability and protection probability to ensure robust goal-oriented transmission.

Key insights

Joint semantic-physical layer design protects critical information by assigning symbol positions based on semantic importance.

Principles

• Gray-coded constellations are suboptimal for non-uniform semantic importance.
• Semantic criticality should guide physical layer mapping.

Method

A VQ-VAE extracts concepts, an SCI scores relevance, and a DRL agent selects subsets. A learned semantic-aware M-QAM constellation then maps symbols based on SCI and co-occurrence.

In practice

• Achieves 21:1 compression ratio.
• Maintains semantic quality >0.9.
• Works across MNIST, Fashion-MNIST, FSDD.

Topics

• Semantic Communication
• Constellation Design
• Deep Reinforcement Learning
• Vector Quantized-Variational Autoencoder
• Semantic Criticality Indicator

Best for: AI Scientist, Research Scientist

Related on AIssential

• Text-as-Signal: Quantitative Semantic Scoring with Embeddings, Logprobs, and Noise Reduction — Text can be transformed into operational, continuous semantic data points for AI engineering via embeddings and logprob scoring.
• Semantic drift and semantic integrity: Stewarding meaning in the age of AI — AI exponentially accelerates semantic drift, demanding a shift to continuous semantic integrity to govern system meaning.
• TONIC: Token-Centric Semantic Communication for Task-Oriented Wireless Systems — Align wireless communication with token-level semantics for foundation models, optimizing task performance over bit fidelity.
• The Evolution of Meaning: NLP’s Long March from Syntax to Semantics — Meaning is a dynamic process requiring active computation and external anchoring, not an inherent property of a word.
• The Mathematical Foundations of Intelligence [Professor Yi Ma] — Intelligence can be formalized through parsimony and self-consistency, enabling principled AI architecture design.
• Structure Before Collapse: Transient semantic geometry in next-token prediction — Language models transiently learn semantic structure from one-hot labels before representations collapse to a symmetric, non-semantic state.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.