Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Transformers for Green Semantic Communication: Less Energy, More Semantics

Shubhabrata Mukherjee, Cory Beard, Sejun Song

TL;DR

The paper addresses the lack of universal metrics tying semantic fidelity to energy usage in SemCom by introducing Energy-Optimized Semantic Loss (EOSL), a multi-objective criterion that guides transformer-based encoder/decoder selection. EOSL jointly accounts for semantic loss, channel loss, and energy expenditures, formalized as $EOSL = \sum_{j=1}^{n} {\{ \lambda_{sm}(1 - S_{sm_j}(M_i, M_p)) + \lambda_{lch}(L_{ch_j}) + \lambda_{e_c}(\frac{E_{c_j}}{E_{c,max}}) + \lambda_{e_s}(\frac{E_{s_j}}{E_{s,max}}) \}}$, and is used for model selection rather than training. Through experiments with image-to-text and text-to-image tasks, the authors benchmark CPU/GPU energy across five encoders and demonstrate that EOSL-guided selection can achieve up to 90% energy savings and about 44% improvement in semantic similarity during inference, with decoders consuming substantially more energy than encoders. The work lays the groundwork for greener SemCom architectures by enabling energy-aware inference-time model selection and cross-modal semantic optimization across resource-constrained platforms.

Abstract

Semantic communication aims to transmit meaningful and effective information, rather than focusing on individual symbols or bits. This results in benefits like reduced latency, bandwidth usage, and higher throughput compared with traditional communication. However, semantic communication poses significant challenges due to the need for universal metrics to benchmark the joint effects of semantic information loss and practical energy consumption. This research presents a novel multi-objective loss function named "Energy-Optimized Semantic Loss" (EOSL), addressing the challenge of balancing semantic information loss and energy consumption. Through comprehensive experiments on transformer models, including CPU and GPU energy usage, it is demonstrated that EOSL-based encoder model selection can save up to 90% of energy while achieving a 44% improvement in semantic similarity performance during inference in this experiment. This work paves the way for energy-efficient neural network selection and the development of greener semantic communication architectures.

Transformers for Green Semantic Communication: Less Energy, More Semantics

TL;DR

The paper addresses the lack of universal metrics tying semantic fidelity to energy usage in SemCom by introducing Energy-Optimized Semantic Loss (EOSL), a multi-objective criterion that guides transformer-based encoder/decoder selection. EOSL jointly accounts for semantic loss, channel loss, and energy expenditures, formalized as , and is used for model selection rather than training. Through experiments with image-to-text and text-to-image tasks, the authors benchmark CPU/GPU energy across five encoders and demonstrate that EOSL-guided selection can achieve up to 90% energy savings and about 44% improvement in semantic similarity during inference, with decoders consuming substantially more energy than encoders. The work lays the groundwork for greener SemCom architectures by enabling energy-aware inference-time model selection and cross-modal semantic optimization across resource-constrained platforms.

Abstract

Semantic communication aims to transmit meaningful and effective information, rather than focusing on individual symbols or bits. This results in benefits like reduced latency, bandwidth usage, and higher throughput compared with traditional communication. However, semantic communication poses significant challenges due to the need for universal metrics to benchmark the joint effects of semantic information loss and practical energy consumption. This research presents a novel multi-objective loss function named "Energy-Optimized Semantic Loss" (EOSL), addressing the challenge of balancing semantic information loss and energy consumption. Through comprehensive experiments on transformer models, including CPU and GPU energy usage, it is demonstrated that EOSL-based encoder model selection can save up to 90% of energy while achieving a 44% improvement in semantic similarity performance during inference in this experiment. This work paves the way for energy-efficient neural network selection and the development of greener semantic communication architectures.
Paper Structure (8 sections, 9 equations, 6 figures, 2 tables)

This paper contains 8 sections, 9 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Energy Optimized Semantic Loss
  4. Semantic Encoder and Decoder design
  5. Experimental results
  6. Image-to-text encoding and EOSL-based model selection
  7. Image-to-text encoding and text-to-image decoding
  8. Conclusion and future work

Figures (6)

  • Figure 1: The basic blocks of a semantic communication
  • Figure 2: Evolution of complexity and training requirement of models
  • Figure 3: Resource utilization during inference
  • Figure 4: Encoder and decoder block design using image/text transformation and effect of semantic noise
  • Figure 5: Changes of EOSL with the probability of bit error rate when using different values of $\lambda_{sm}$, $\lambda_{lch}$, $\lambda_{e_c}$, $\lambda_{e_s}$
  • ...and 1 more figures