Revisiting Compression and Communication through the Lens of Generative AI

The classical insight from information theory that prediction, learning, and compression are intimately connected has gained renewed relevance in light of recent advances in generative artificial intelligence. Large language models and diffusion processes have demonstrated remarkable capabilities in capturing and synthesizing complex data distributions across modalities including language, audio, and images. These developments prompt a re-examination of fundamental problems in compression and communication, particularly in the context of joint source-channel coding. The signal processing, communications, and information theory communities are increasingly investigating how generative models can be leveraged not only to model sources more effectively, but also to reimagine encoders, decoders, and estimation algorithms. In this talk, I will survey recent approaches at this intersection, focusing mostly on algorithmic innovations.