Paul Thompson

OK, here's today's video. Guided tour of diffusion MRI + tractometry in lots of brain diseases - a bit scary in front of an audience of diffusion MRI experts who know tons more about this than I do ! youtube.com/watch?v=i2jHFm…
(start at 15 min to skip the ENIGMA intro)

Some fun puzzles [1] for a lecture on DMD analysis of brain networks.
1. Can you make a Davis-Kahan-like finite-sample law for the phase of a complex DMD eigenvector, with the spectral gap replaced by a quantity that accounts for non-normality?
2. Can you propose a data sufficiency law that would help us to predict a finite-sample phase-identifiability horizon for a specific pair of networks or regions
[1] there is actually a 6-way intersection in Beverly Hills, near Franklin Canyon Lake I think. You have to work out whose turn it is to go, and not just wait. But it is clearer to me if the 3rd network goes up and down (like 3D chess). Maybe this happens in the brain, as there are (at least) 3 principal gradients.

🔥Can diffusion MRI make a "temperature map" in the brain of an unborn fetus? What about in adults?
🔥Apparently yes [1] (I had not thought about this before).
🔥Water diffusion increases with temperature, so [1] T= [2256.74 / ln(4.39221/D) ] - 273.15, where D is the CSF diffusion coefficient in units of 10^{-3} {mm^2/s}, T is in °C
🔥SO a 1°C temperature change corresponds to ~2.35% change in CSF diffusivity.
!n 11 preterm neonates, estimates agreed with axillary temperatures to within ~1°C
🔥Apparently this has been used in adults too (and in Alzheimer's +Parkinson's - ventricular temperature correlated with DTI-ALPS in controls (r=0.44), but not significantly with amyloid PET) [2].
🔥They probably have to be careful about CSF pulsation, which they mention.
[1] *Disclaimer, this is not our paper and I don't know the authors onlinelibrary.wiley.com/doi/10.1002/jm…
[2] onlinelibrary.wiley.com/doi/10.1002/jm…

🔥There is an interesting Japanese area of Los Angeles where you can find puzzle books in Japanese.
🔥These puzzles are interesting because the book is in Japanese, so, unless you understand Japanese, you have to figure out the rules of the puzzle, and then in a second step, you have to solve the puzzle based on guessing what the rules of the puzzle are. So you may be unable to solve the puzzle because you have guessed the rules wrongly, or you did guess the rules* but you can't solve it.
🔥This is a lot like Tab-PFN, which is a very bizarre (fairly new) machine learning method that can figure out what problem it is solving when it sees some data. Rather than train it to solve some specific problem, it can figure out what type of problem it is solving from the structure of the data, and then go ahead and solve it. This seems impossible but you can set it up using amortised Bayesian inference (watch videos by Frank Hutter and his team in Freiburg). AutoGluon is a related method.
🔥Now I think this can be quite promising for us doing brain-related AI, if we train a PFN with lots of spectral priors from the zeta laws across spectral slopes.
*when I got home I looked up what the puzzle is and it turns out it's a popular one called nonograms [1] so now we have some good weekend reading [2]. Non is apparently not "9" or from the Greek word νόμος but named after a Japanese woman who won a competition in Tokyo to design pictures using windows of a skyscraper (where you turn selected windows on or off to produce a recognizable image on the building)
[1] sciencedirect.com/science/articl…
[2] arxiv.org/pdf/2507.07283

Interesting inverse problem*
*it says it should only take 46 minutes to do