Conceptual Learning and Causal Reasoning for Semantic Communication

Semantic communication is a paradigm shift toward meaning-oriented communication that is largely enabled by artificial intelligence technologies. One approach to semantic communication that leverages this intelligence is based on the theory of conceptual spaces. However, learning these conceptual spaces in an automated fashion is a major challenge. Moreover, the true intelligence of semantic communication approaches is still limited, and some have suggested casual reasoning as a key to unlocking this higher-level intelligence. To address these challenges, in this paper, we present two major innovations for semantic communications with conceptual spaces. First, we introduce the use of the variational autoencoder as well as the p-Wasserstein distance metric from optimal transport theory for learning the domains of the conceptual space model of semantics. These additions simplify the training process and allow for greater control, and experiments on a sample dataset show that they additionally improve the quality of the learned domains. Second, we present for the first time, a mechanism for carrying out causal reasoning over semantic information modeled by a conceptual space with respect to some communication goal, and show how this mechanism can be used to reduce communication bandwidth. Simulations comparing the proposed method to baseline communication techniques demonstrate accuracy improvements of up to 40% on classification tasks and providing similar performance to a traditional approach with a 99% reduction in rate.