Dynamic Relative Representations for Goal-Oriented Semantic Communications

Simone Fiorellino, Claudio Battiloro, Emilio Calvanese Strinati, Paolo Di Lorenzo

Introduction

As the telecommunications industry transitions toward 6G, the paradigm is shifting from traditional bit-centric communication to Semantic Communication (SemCom). Unlike classical systems that prioritize the error-free delivery of every bit, SemCom focuses on transmitting the "meaning" or the essential features required to complete a specific task. This approach promises significant improvements in efficiency and latency by transmitting only the data relevant to the receiver's goal. However, a significant barrier to the widespread adoption of SemCom is the challenge of model heterogeneity and the resulting "semantic noise."

The Challenge of Semantic Misalignment

In a typical SemCom setup, a transmitter uses a Deep Neural Network (DNN) to extract features, and a receiver uses another DNN to interpret them. A major hurdle arises when these devices use different architectures—for instance, a transmitter using a Vision Transformer (ViT) and a receiver using a MobileNet. Because these models have different internal "languages" or latent spaces, the receiver cannot interpret the transmitter's features without joint retraining. In dynamic edge environments where devices frequently join or leave the network, retraining the entire system is computationally prohibitive and practically impossible. This misalignment is often referred to as Semantic Noise.

Proposed Framework: Relative Representations

To address this, the authors introduce a framework based on Relative Representations (RelReps). Instead of transmitting absolute feature vectors (the raw outputs of a DNN), the transmitter calculates the similarity between the data and a shared set of "anchors"—representative data samples known to both the transmitter and receiver.

This similarity vector is what is actually transmitted. Because the relative relationships between data points tend to remain consistent across different well-trained models (even those with different architectures), this approach enables "zero-shot stitching." It allows a receiver to interpret a transmitter's data even if they have never been trained together, effectively turning the latent space into a universal coordinate system.

Dynamic Resource Allocation via Lyapunov Optimization

Beyond semantic alignment, the paper addresses the physical constraints of edge computing. The authors employ Lyapunov stochastic optimization to manage the complex trade-offs between power consumption, latency, and task accuracy. The system dynamically adjusts several parameters in real-time:

  • Communication Parameters: Uplink data rates ($R_t$).
  • Computation Resources: CPU frequency ($f_t$).
  • Learning Resources: Selection of the encoder ($E_t$) and the number of anchors ($A_t$).

The number of anchors serves as a "knob" for semantic compression: fewer anchors reduce bandwidth requirements but may lower accuracy, while more anchors provide a more detailed description at the cost of higher resource usage. The optimization framework provides simple closed-form solutions, making it ideal for implementation on resource-constrained edge devices without requiring prior knowledge of wireless channel statistics.

Experimental Evaluation

The study evaluated the framework using the CIFAR-10 dataset for image classification and a variety of pre-trained encoders from the ImageNet library (Timm models), including MobileNetV3, RexNet, and Vision Transformers (ViT).

Key Findings

  1. Encoder Agnosticism: The system demonstrated robust zero-shot performance. A decoder trained on a MobileNet encoder could successfully classify images sent by a much larger Vision Transformer (ranging from 1.5M to 98M parameters) without any retraining of the receiver.
  2. Power-Latency Efficiency: The results revealed a clear Pareto frontier. For example, by relaxing the average latency constraint from 0.03s to 0.06s, the system could reduce average power consumption by nearly 90% (from ~0.1W to ~0.01W), depending on the accuracy target.
  3. Stability and Constraint Satisfaction: The dynamic algorithm successfully maintained instantaneous latency below the maximum tolerable threshold while meeting long-term average accuracy targets.

Conclusion and Future Work

This research provides a viable path for heterogeneous AI models to communicate efficiently in 6G environments. By leveraging Relative Representations, the framework overcomes the need for joint training, while the Lyapunov-based optimization ensures energy and latency efficiency.

Future research will focus on optimizing anchor selection—moving from random sampling to choosing more "prototypical" samples—and exploring more complex tasks such as object detection and semantic segmentation in real-time video streams. Additionally, investigating alternative similarity metrics beyond cosine similarity may further enhance the robustness of latent space alignment.

Summary updated

BibTeX
@inproceedings{fiorellino2024dynamic,
  author = {Simone Fiorellino and Claudio Battiloro and Emilio Calvanese Strinati and Paolo Di Lorenzo},
  title = {{Dynamic Relative Representations for Goal-Oriented Semantic Communications}},
  booktitle = {EUSIPCO},
  year = {2024},
}
Figure