Dynamic Relative Representations for Goal-Oriented Semantic Communications

TL;DR

The paper tackles semantic mismatches in goal-oriented semantic communications by introducing Relative Representations (RelRep) that enable latent-space alignment and zero-shot stitching across heterogeneous encoders. It couples RelRep with a Lyapunov stochastic optimization framework to dynamically allocate rate, CPU frequency, and learning resources (encoders and anchors) in order to minimize long-term power while meeting latency and accuracy targets. Key contributions include formal RelRep construction $r_xi(E,A) = (sim(e_xi, e_a1), ..., sim(e_xi, e_a|A|))$, a tractable per-slot optimization with closed-form updates for $R_t^*$ and $f_t^*$, and numerical validation showing favorable energy-delay-accuracy trade-offs. The approach enables encoder-agnostic, robust, energy-efficient edge SemCom in dynamic 6G scenarios with heterogeneous devices and links, by shifting semantic understanding workload to the receiver and using dynamic resource control at the edge.

Abstract

In future 6G wireless networks, semantic and effectiveness aspects of communications will play a fundamental role, incorporating meaning and relevance into transmissions. However, obstacles arise when devices employ diverse languages, logic, or internal representations, leading to semantic mismatches that might jeopardize understanding. In latent space communication, this challenge manifests as misalignment within high-dimensional representations where deep neural networks encode data. This paper presents a novel framework for goal-oriented semantic communication, leveraging relative representations to mitigate semantic mismatches via latent space alignment. We propose a dynamic optimization strategy that adapts relative representations, communication parameters, and computation resources for energy-efficient, low-latency, goal-oriented semantic communications. Numerical results demonstrate our methodology's effectiveness in mitigating mismatches among devices, while optimizing energy consumption, delay, and effectiveness.

Figures (4)

• Figure 1: The edge device (TX) encodes data according to the optimal encoder-anchor set pair $(E^*,\mathbb{A}^*)$. The RX is equipped with relative decoders trained on shared anchor sets to interpret the transmitted relative representations.
• Figure 2: Performance of zero-shot stitching using different pre-trained encoders and varying anchor set sizes on the CIFAR-10 classification task. The dashed line is end-to-end performance.
• Figure 3: Long-term average power vs $\bar{L}$, for different $\bar{G}$.
• Figure 4: Latency and accuracy vs $t$, $\bar{L}=0.04\ s$ and $\bar{G}=0.8$.