Learning to Compress and Communicate
The focus of this work package is to communicate and compress data and semantics using AI technologies in the context of distributed systems. The team’s research is aimed at finding new fundamental performance bounds on lossy and lossless compression and to develop algorithms attaining these fundamental bounds.
Outputs to date: Actions Speak Louder Than Words: Rate-Reward Trade-off in Markov Decision Processes This paper explores how intelligent agents (like robots or software systems) can communicate with each other by actions (agent sends messages through its actions in a shared environment). They propose a new method called Act2Comm, which teaches agents how to act in a way that both achieves goals and communicates effectively.
The group is also furthering research in Deep Joint Source Channel Coding to use generative diffusion models in the encoding and reconstruction during image transfer.
