Classification-Driven Discrete Neural Representation Learning for Semantic Communications

Semantic communications is a key enabler of the Internet of Things (IoT). By focusing on the semantic meaning of data rather than bit-level recovery, it allows intelligent agents to communicate necessary information at much lower rates. A promising technique for semantic communications is discrete neural representation learning (DNRL). The main idea is to learn discrete symbols from low-level, high-dimensional sensory data, such that each symbol is grounded to a meaningful pattern in the sensory domain. This article proposes a DNRL scheme that integrates three mechanisms into a coherent framework: 1) contrastive learning; 2) sparse coding; and 3) neural index quantization. The proposed scheme is applied to public image data sets for lossy image compression with a downstream classification task. Results show that the proposed approach produces a highly compact continuous latent representation and a semantic discrete representation, with marginal degradation to the classification accuracy. The interpretability and consistency of the learned subsymbolic discrete representations are validated by experiments of neural-net dissection, neural-net visualization, and MaxAmp-K classification test, a concept that we propose to evaluate classification performance of extremely compressed signals. Finally, the discrete representations are shown to be useful in rate-adaptive distributed sensing applications at the low-to-medium signal-to-noise ratios (SNRs).