Current communication systems, that aims to accurately reconstruct a transmitter (TX)’s message at the receiver (RX) without considering message meaning, has served us well for decades. However, their limitations start to become apparent when faced with the challenge of reliably transmitting massive data or meeting extremely stringent transmission requirements of intent-based systems, within which the message intent and its system impact must be considered. Under this knowledge- and reasoning-driven semantic communication (SC) paradigm, TXs and RXs can exploit their common knowledge to proactively anticipate and correct errors, generate content, operate more effectively under intermittent channels, and reduce communication overhead. While SCs was first proposed in 1949, it remained largely unexplored for decades due to a lack of advanced artificial intelligence (AI) techniques (e.g., deep learning), computing resources, and compelling applications. The objective of this project is to design a novel wireless SC framework via developing novel adaptive, generalizable, and reasonable semantic information models in a structured format with data entities and their relationships, as well as jointly optimizing semantic information generation and transmission over resource-constrained wireless networks while considering knowledge of both TX and RX. The project's broader significance and importance are contributing towards transforming a communication problem from a reconstruction problem where the RX is a passive node and the TX takes full control of the transmission and manipulation of the message, into a symmetric system in which the RX can regenerate the original content with instilled capability of reasoning and processing multi-modal data. The project develops a comprehensive educational plan that includes new AI native communication system course materials, and hands-on activities using the designed software tools and testbeds.
This research project establishes the fundamental theoretical and practical scientific foundations of SC networks via (i) designing novel semantic information model and generation methods to precisely represent source data, (ii) developing robust and efficient wireless SC systems by jointly optimizing semantic information generation and transmission over a noisy, wireless resource constrained link, and (iii) introducing a novel learning framework for semantic information generation, resource management, and user association. The developed models and methods are tested through an integrative experimental implementation over software simulations, prototype evaluations, and real-world wireless testbeds, coupled with an advanced SC application on mixed reality.