Research
Everything, grouped into the three areas the work falls into, each opened by what the area means to me.
Cognition
Whether two kinds of thought share machinery in the brain is usually tested with one small experiment per kind, followed by a check for overlap between the regions each activates.
That design has a flaw it cannot correct for. Each experiment arrives with its own task and its own confounds, so any overlap found between them reflects how the two experiments were built as much as it reflects the brain.
A stimulus that already contains all of them removes the need for that comparison. Someone telling a story counts things, places them and times them continuously, in one stream of words, so a single recording and a single model can produce a map for each, and those maps can be compared directly. What it demands in return is a model whose features are separable to begin with, which is the problem the work under Bayesian addresses.
Closed loop
Deep brain stimulation is set in a clinic over an afternoon and then runs unchanged for months, in a disease whose symptoms move over the course of a day. Closing that gap means letting the stimulator respond to the person, and that has to start with measuring what the person is doing.
A clinic visit cannot supply that measurement. The behaviour worth responding to happens at home across hours, and the neural signal you would drive the stimulator from is itself partly a consequence of the stimulation already being delivered. Recording is therefore the first problem: implanted electrodes, wearables and video running at once in somebody’s house, and synchronized closely enough that they can be compared.
Once they can, the question is whether behaviour predicts the neural signal well enough to be worth decoding from.
Bayesian
Every model here is fit to predict something measured. This area is about what you may then conclude from its parameters.
A model can predict well and still support no conclusion. When its features are correlated, many different settings of the weights produce the same prediction, so the weights a fit returns are one choice among many that the data cannot separate. Accuracy does not reveal this, because accuracy is identical for all of them.
This is a property of the basis, not of the fit. If the features are sparse and non-negative, with one interpretable direction per coordinate, the weights become identifiable, and reading them off is a measurement rather than an interpretation of an arbitrary point.
