I had two major conversations with research-heros of mine at NeurIPS. Quick outline of what we talked about.
Anti-Toy Models
One conversation involved disliking toy models for neuroscience. I agreed - but with the caveat that we should love toy models for clinical layers. Including behavior and clinician decision-making.
Why? Because those models are flat out being used right now to change decision-making. Any downstream measurement of that decision-making will carry information about the toy model with it. The fact that interventions don’t fail spectacularly is a huge hint!
So while toy models of brain layer networks is imho doomed to fail, toy models of behavior layer are much more promising and a veritable proving ground for geometric ML. Start from there and “reverse engineer” into the brain from there!!
Clinician is Controller
A physicist I respect (in the science) is starting to be more vocal against the AI-hype. And the mathematical approaches they’re implementing are awesome - I’m just a bit jealous.
I asked them: “why aren’t you doing control theoretic analyses in human patients with DBS?” The response clarified to me that I didn’t miss my boat at all: they said “it’s hard to run experiments in patients”.
Wild.
Because their base model was still reductionist. In order to implement control theoretic analyses, which can be framed as hypotheses, they couldn’t fathom an alternative to careful isolation and measurement. So the first step of finding ways to do experiments in patients needed to happen before they could tackle the cool question.
That’s absurd!
The DBS is the experiment - just a multivariate one. The clinician is the controller that adjusts the inputs based on models and measurements - let’s formalize what they’re doing and move forward from there!
Much, much more to say on this elsewhere.