The evolution of artificial intelligence applications has exponentially increased the demand for efficient and low-latency transmission of large data volumes, posing significant challenges to conventional wireless communication systems. While conventional systems, known for simplicity and elegance, have been extensively studied with a focus on reliable bit sequence transmission without considering the content semantics and communication objectives, they can yield sub-optimal results in practical scenarios characterized by finite block lengths and rapidly varying wireless channel conditions. Next-generation communication systems aim to address these limitations by incorporating message semantics and specific communication objectives into their design, a paradigm known as semantic communication. One of the most promising advancements in this area is the deep joint source and channel coding (DeepJSCC), a pioneering deep learning-based technique that combines source compression and error correction into a unified process. This approach allows end-to-end optimization and promises significant improvements in bandwidth efficiency and reliability, thereby challenging Shannon’s separation theorem.

This thesis delves into the transformative potential of DeepJSCC in the realm of semantic communications, with a particular focus on developing practical DeepJSCC-based transceivers for image transmission and edge inference tasks. By exploring the DeepJSCC for image transmission across three canonical channel models: orthogonal frequency division multiplexing (OFDM), multiple-input and multiple-output (MIMO), and feedback channels, this thesis showcases DeepJSCC's superiority, adaptability, efficiency, and robustness in diverse semantic communication scenarios. Furthermore, this thesis extends the application of DeepJSCC to edge inference tasks, showcasing its ability to optimize different communication objectives. The comprehensive evaluation of the proposed DeepJSCC frameworks across diverse tasks and scenarios underscores its versatility, establishing it as a highly promising alternative to traditional separation-based algorithms and a potential solution for next-generation semantic communication systems.