Introduction
Traditional communication theory, established by Claude Shannon, focuses on the reliable transmission of bits over noisy channels. However, the emerging paradigm of Semantic Communications (SemCom) shifts this focus toward the transmission of the meaning or task-relevant information inherent in data. While SemCom offers significant potential for bandwidth savings and improved robustness, it relies heavily on Deep Learning (DL) architectures.
A major hurdle in the practical deployment of SemCom is the "black box" nature of DL models. Unlike classical systems, where closed-form expressions relate Signal-to-Noise Ratio (SNR) to Bit Error Rate (BER), SemCom lacks tractable mathematical models to predict end-to-end performance metrics, such as image similarity (Multi-Scale Structural Similarity Index, MS-SSIM) or inference accuracy. This paper introduces the Alpha-Beta-Gamma (ABG) formula, a novel empirical model designed to bridge this gap and enable precise resource allocation.
System Model and the ABG Formula
The authors consider a typical semantic communication system consisting of a semantic encoder, a quantizer, a modulator, a physical channel, and their corresponding receiver-side components.
To model the performance of this system, the authors propose the ABG formula: $$\phi(\rho, n_b) = \alpha(n_b) - \frac{\gamma}{1 + (\beta\rho)^\tau}$$ where $\phi$ represents the performance metric (e.g., MS-SSIM), $\rho$ is the transmission SNR, and $n_b$ is the number of quantization bits. The parameter $\alpha(n_b)$ represents the performance upper bound, while $\beta, \gamma,$ and $\tau$ are fitting parameters specific to the neural network architecture. This formula allows engineers to treat complex neural networks as tractable functions for optimization.
Universal Performance Modeling
One of the most significant findings is the universal applicability of the ABG formula. The study evaluated the model across various modalities and architectures using the CIFAR-10 dataset for images and the European Parliament Proceedings Parallel Corpus for text.
As shown in the empirical results, the ABG formula (red line) closely matches the performance of diverse architectures, including Convolutional Neural Networks (CNNs), SCUNet, Vision Transformers (ViT), and Swin Transformers. The Sum of Squared Errors (SSE) for these fits ranged from $10^{-5}$ to $10^{-6}$, indicating exceptional accuracy across different model complexities.
Key Insights and Optimization
- Performance Saturation: The model captures the reality that SemCom performance increases rapidly with SNR initially but eventually plateaus at a level $\alpha$. This plateau is determined by the model's architecture and quantization level, suggesting that increasing power beyond a certain point yields diminishing returns.
- The Quantization Bottleneck: The upper bound $\alpha$ is strictly tied to the number of quantized bits ($n_b$). By modeling this relationship, designers can effectively trade off source coding (quantization) against channel coding (power).
- Resource Allocation: By making the "black box" tractable, the authors solved three critical optimization problems:
- Adaptive Power Control: Adjusting power in real-time to maintain a target Quality of Service (QoS) over fading channels.
- Energy Efficiency Maximization: Finding the optimal balance between reconstruction quality and power consumption using a Dinkelbach-type algorithm.
- Multi-user Fairness: Optimizing resource allocation in multi-user OFDMA systems to ensure the best performance for the "worst-off" user.
Results and Conclusion
The proposed framework demonstrates substantial practical benefits. In simulations of random fading channels, adaptive power control based on the ABG formula achieved a 0% outage rate, whereas traditional fixed-power schemes frequently dropped below quality thresholds. Furthermore, in multi-user scenarios, the proposed allocation scheme improved the worst-case MS-SSIM from 0.818 to 0.924, providing a visible boost in image reconstruction quality.
While the ABG formula requires pre-calculating parameters for specific models, it provides the first rigorous mathematical foundation for resource allocation in semantic communications. Future work will focus on exploring more complex interference environments and developing "few-shot" parameter estimation techniques to reduce the empirical data requirements.
Summary updated