about

I work on control theory for systems that learn. The question I keep returning to is what happens to a closed-loop guarantee once the controller is distributed across a system too large for any one part to observe.

System level synthesis makes that tractable by specifying the closed-loop response directly and recovering the controller from it. Locality constraints then turn a problem that grows with the whole system into one that grows with a neighbourhood.

The same parametrization describes architectures that came out of machine learning. A residual stream carries state forward and a layer acts on it, which is a dynamical system with a different vocabulary.

My research career has spanned France, the UK, Japan and the US. Previously, I used RL to improve adaptive perception and foveated sensing in embodied robotic agents. My ideas included developing adaptive sensors that dynamically switch modalities based upon prediction errors, and control systems that take input from those dynamic sensors.

The throughline from that work to this one is decision-making under partial observation.