Friday, May 4, 2007

Neuroscience is hard... Let's go shopping!

Classic quote: "Thus, although neuroanatomical information will be central to understanding how the brain processes stimuli and forms representations, our current knowledge of neuroanatomy is sufficient to constrain neither the problem of binding nor its solution." - Adina L. Roskies, writing in the introduction to a special issue of Neuron on the binding problem.

I recently read (well, listened to, at any rate) Surely You're Joking Mr. Feynman and Further Adventures of a Curious Character, both semi-autobiographical works by famed physicist Richard Feynman. Although these books are mostly about Feynman's extracurricular adventures, he does touch upon his philosophy of scientific pursuits. I was struck by what he described as the radical honesty good scientists must bring to their work. Feynman claims that it is not sufficient to simply present all of the details of your work, he asserts that the good scientist must lay bare all of the potential flaws in their theories and experiments. He particularly warns against scientists lying to themselves and failing to recognize weaknesses in their work. This idea initially intimidated me. I make a point of remaining convinced that I am on the cusp of a (or perhaps THE) great discovery regarding how the brain works. It makes going into lab more fun, it seems mostly harmless, and some days I'm even convinced that it's true. However, none of my work can stand up to this sort of merciless intellectual assault. Yes, there are some interesting ideas in there that bear a passing resemblance to the experimental data, but I can't quite jam my higher-level ideas into a neuron by neuron, receptor by receptor, and ion by ion map of the biophysics of the brain. I can't pretend to have read even 1% of the literature on these detailed phenomena. Moreover, it is a truism of neuroscience that for every paper there is an opposite (but not necessarily equal) paper claiming a mutually exclusive result. Creating a model which matches a set of consistent, correct data is hard. Creating a model which is compatible with an undifferentiated mass of mutually contradictory data, half of which is necessarily wrong, is by definition impossible. I think the real answer is that the brain is a messier affair than particle physics. Cloud chambers are neat and sterile. The brain is squishy an amorphous, and you need to peer into it through the tiniest of tubes (generally, a bunch of solid wires). In this sense, neuroscience is much more akin to statistical mechanics than to particle physics. Asking how the brain works is like asking for a detailed description of the turbulence behind a large truck. It's possible to write down rules describing the interactions at the smallest scale, and it's possible to make some hand-wavy measurements of phenomena at the largest scales, but ne'er the twain shall meet.


eric said...

Speaking of stat mech, you may wish to check out some of Sethna's work on "sloppy models" which are (mostly biological models) with many unknown parameters:

Sethna Phys Rev E 68, 021904 (2003)

I think it's on the arXiv too.

He takes the approach that if one considers the ensembles (in te stat mech sense) that are possible in a model, rather than a much more constrained system of ODEs, one gains a better understanding of the system, or at least one that is not limited as much by the sloppiness of the empirical data. It may be that this is appropiate anyway, because all of the "rate reactions" etc. are really statistical quantities anyway.

I think that this is the kind of scrutiny that Feynman is talking about. Not "does my model agree with all the data" (some of which is probably wrog anyway, or at least being wrongly interpreted as you said), but "what are the failures of my general approach -- is it over-specific? What big assumptions does it make?" etc.

Also keep in mind that while feynman was a fucking genius, the rest of us have to be content with making some progress, any progress. There's a difference between intellectual honesty and brilliance: sometimes we cannot forsee the potential failures of our approach.

eric said...

Oops, here's the full article citation:

K.S. Brown and Sethna, J.P." Statistical mechanical approaches to models
with many poorly known parameters" PHYS REV E 68, 021904 (2003)