Research
I look for geometric structure in learned representations that we can put to work in three ways: as a prior that makes learning better, as a verifier at inference time, and as a diagnostic that flags anomalies. Most of what I do sits in two lines of work.
World models and physical reasoning
Neural networks see the world through their representations. I ask what those representations encode about physics, and how to read it out without changing the model in the process.
- Is physical plausibility written into the geometry of frozen image features? GeoPhys
- Do real and generated videos trace different paths through a perceptual representation? ReStraV
- When is a successful linear probe evidence of a linear representation? Linear probe
- Does adapting a model to read out its physics change the physics it represents? Observer Effect
- How do we keep video world models from losing what changes between frames? MotionJEPA
- Can procedural data give vision transformers a better start? Procedural Core
Distributed learning
Many models, many owners, and data that never leaves home. I work on how such models can still learn together when their data barely overlap, and how to keep that collaboration small, fast and private.
- How can federated training converge when the clients’ data barely overlap? FedLEx
- Can a single round of communication be enough when every client sees a single class? FederatedFactory
- How do we prune federated models without losing what each client knows? AutoFLIP
- How do we keep collaborative optimisation private and secure? PPFLex