Research
I am interested in what science costs, and in whether that cost can be studied rather than simply absorbed.
Where this comes from
I came to machine learning from particle physics, where a single measurement takes years, the instrument is as much an engineering problem as a scientific one, and the expensive part is deciding what to measure next. That shapes what I find interesting now. I care less about whether a system can produce a correct answer than about what it spent getting there, and whether the path it took would hold up if you ran it again.
What I find interesting
Ideas are cheap and getting cheaper. Experiments are not. Any account of scientific progress that ignores what an experiment costs is describing a different activity.
Two systems can reach the same conclusion by completely different routes — one reproducible, one lucky. The trajectory is where that difference lives, and it is usually thrown away. How you choose to represent it decides what you are able to compare afterwards.
Most of what I build is measurement infrastructure. I think this is where the field is currently under-invested, and I would rather work on the part that is under-invested.
In rare-event searches most of the work is understanding your own instrument well enough to trust a handful of events. I have never stopped thinking that way about machine-learning systems.
Where I think this goes
If scientific trajectories can be captured well enough to evaluate, they can eventually be learned from. I am more confident about the first half of that sentence than the second, which is why my work currently sits on the measurement side of it.