Deep Learning Based Superposition Coded Modulation for Hierarchical Semantic Communications over Broadcast Channels

TL;DR

This work addresses hierarchical semantic communications over Gaussian broadcast channels where receivers request correlated but different semantic content. It introduces DeepSCM, a deep learning-based superposition coded modulation architecture that encodes a basic EFV $U_1$ and an enhanced EFV $U_2$, then decorrelates to form a refinement $R$ for a two-layer superposition broadcast, shaped by a power allocation factor $\alpha$ and realized with probabilistic modulation via Gumbel-Softmax. A three-stage training procedure aligns the transmitter and two receivers to deliver both observable sources and hierarchical semantics, with losses that balance semantic accuracy and pixel-level image fidelity. Experiments on CIFAR-100 show that DeepSCM outperforms non-superposition coded modulation and conventional SSCC baselines, particularly under large channel disparity and with higher-order modulation, while approaching the single-receiver performance upper bound and offering robustness to channel variation.

Abstract

We consider multi-user semantic communications over broadcast channels. While most existing works consider that each receiver requires either the same or independent semantic information, this paper explores the scenario where the semantic information desired by different receivers is different but correlated. In particular, we investigate semantic communications over Gaussian broadcast channels where the transmitter has a common observable source but the receivers wish to recover hierarchical semantic information in adaptation to their channel conditions. Inspired by the capacity achieving property of superposition codes, we propose a deep learning based superposition coded modulation (DeepSCM) scheme. Specifically, the hierarchical semantic information is first extracted and encoded into basic and enhanced feature vectors. A linear minimum mean square error (LMMSE) decorrelator is then developed to obtain a refinement from the enhanced features that is uncorrelated with the basic features. Finally, the basic features and their refinement are superposed for broadcasting after probabilistic modulation. Experiments are conducted for two-receiver image semantic broadcasting with coarse and fine classification as hierarchical semantic tasks. DeepSCM outperforms the benchmarking coded-modulation scheme without a superposition structure, especially with large channel disparity and high order modulation. It also approaches the performance upperbound as if there were only one receiver.

Figures (12)

• Figure 1: Framework of the proposed DeepSCM scheme.
• Figure 2: Constellations after superposition with various values of $\alpha$. The red diamonds indicate the inner constellation multiplied by the PAF, and the blue dots indicate the resulted super-constellation. All symbols are assumed to be uniformly distributed in this figure with $P=1$. Notably, (b) and (e) present rectangular 16QAM and 64QAM, respectively.
• Figure 3: NN architecture of the first JCM block.
• Figure 4: Performance of different schemes at varying transmission rates, with 4QAM$\times$16QAM super-constellation.
• Figure 5: Performance of different schemes at varying transmission rates, with 4QAM$\times$4QAM super-constellation.

Theorems & Definitions (1)

• Proposition 1