Semantic Communications for Multi-View Generation

Semantic communication has shown great potential in efficiently accomplishing intelligent tasks. However, existing systems usually make a single-view decision in a deterministic manner while ignoring the intrinsic semantic noise caused by ambiguity. This may result in the misunderstanding during the exchange of semantic information. In this letter, we propose a novel probabilistic uncertainty coding (PUC) architecture to achieve versatile multi-view semantic communications with multiple outputs. Specifically, each feature region is characterized as a Gaussian distribution whose variance represents semantic uncertainty. Then, we leverage the Gaussian information bottleneck theory to jointly optimize the rate-distortion tradeoff between the semantic informativeness and inference performance. Moreover, we provide a semantic similarity metric to evaluate the accuracy of multi-view semantic communication. Considering the cross-modal task of semantic knowledge transfer, simulation results show that PUC outperforms conventional deterministic method and traditional separate transmissions. Due to its semantic-aware capability, the proposed PUC can make diverse yet plausible predictions in a highly reliable and low-latency manner.