Bayesian
What a fitted weight is entitled to mean, and what the features have to look like before it means anything.
Sparse concept atoms · one project not yet named
For about ten years I've been asking a version of the same question, first in philosophy and then in physics: how does meaning come out of mechanism, and how would you know if it had?
The answer I've settled into is that structure lives in the data, and intelligence is whatever extracts it. That puts the constraint in a slightly unfashionable place. How clever the model is matters less than whether the measurement got made at all, and whether it got made somewhere rich enough that the structure survived contact with the experiment.
Most of what I've worked on has been messy on purpose for that reason. Stories instead of word lists. Actual birdsong. Patients at home rather than in a clinic room.
to now
-
I wrote my undergraduate philosophy thesis on AlphaGo, mostly while the matches against Lee Sedol were happening.
-
Then two years at Penn on birdsong with Vijay Balasubramanian and Eve Armstrong, reconstructing state spaces from recordings to see whether the dynamics carried anything people were throwing away by treating song as acoustics.
-
–
At Berkeley I built encoding models that predict cortical responses from language model features, using [the actual stimulus — story listening, which stories, how many hours] rather than controlled stimuli.
-
Now
Now I build characterization instruments.
Research
Closed loop
Recording and stimulation in one circuit, measured in people’s homes rather than a clinic.
Parkinson’s, measured at home
Cognition
What cortex does with quantity, meaning and song, measured from natural stimuli.
Number, space and time · Birdsong in state space
Contact
[invitation line — yours]
Otherwise: camping, mountains, ballet, and I still read philosophy.