In recent decades, communication technology has advanced at unprecedented speed. A key breakthrough is semantic communications—a shift from transmitting raw data to conveying semantic meaning. For example, in image transmission, meaning takes priority over pixel-level accuracy. By integrating user tasks into the communication process, semantic communications improve both efficiency and user experience.
While deep learning has accelerated progress, a transition from analog to digital modulation is essential for compatibility with modern infrastructure. Yet, current digital semantic communication systems still lack an effective digitization mechanism.
Recently, a pair of researchers led by Dr. Dong Jin Ji, an Associate Professor at the Department of Semiconductor Engineering at Seoul National University of Science and Technology, Republic of Korea, have proposed ConcreteSC, a novel digital semantic communication framework that foregoes massive codebooks via temperature-controlled concrete distributions. Their innovative findings were made available online and have been published in the journal IEEE Wireless Communications Letters on 19 June 2025.
Dr. Ji highlights the novelty of their approach, stating "Unlike vector quantization (VQ)—a state-of-the-art digitization technique that suffers from channel noise and codebook divergence during training—our framework offers a fully differentiable solution to quantization, allowing end-to-end training even under channel noise. Notably, due to the nature of the ConcreteSC that directly generates the required bitstream, it is possible to train a multi-feedback-length model pair with a relatively simple masking scheme."
The team carried out simulations on ImageNet under Rayleigh and Rician fading to test the performance of their innovation. They remarkably found that ConcreteSC consistently outperforms VQ-based baselines in terms of structural similarity index and peak signal-to-noise ratio parameters.
A key aspect of this research is that the proposed ConcreteSC can be seamlessly integrated into other semantic communication frameworks, effectively quantizing the codewords. It results in an overall better quantization quality while also decreasing the computational complexity greatly compared to conventional schemes. In ConcreteSC, computational complexity scales linearly with bit length, helping it mitigate the exponential complexity of codebooks.
In this way, ConcreteSC is robust and flexible, and also lowers overhead for semantic communication in next-generation wireless systems. "Sixth-generation wireless communication systems, where semantic communication technologies are expected to be one of the key enablers, are expected to be a major area of application for our technology. These include smart factories, where ultra-dense machine-type communications are prevalent," points out Dr. Ji.
Furthermore, a fully autonomous factory system that does not require communication cables to each equipment but has AI built into every small component could be enabled by this technology. Lastly, an all-around lifecare system for seniors and toddlers alike, with low-power IoT devices that monitor the health of every individual around—all of these devices are expected to run on large AI models, which would be impossible without advanced semantic communication technology similar to the one presented in this work.
