Slashdot Mirror
Sand in the Brain: A Fundamental Theory To Model the Mind
An anonymous reader writes "In 1999, the Danish physicist Per Bak proclaimed to a group of neuroscientists that it had taken him only 10 minutes to determine where the field had gone wrong. Perhaps the brain was less complicated than they thought, he said. Perhaps, he said, the brain worked on the same fundamental principles as a simple sand pile, in which avalanches of various sizes help keep the entire system stable overall — a process he dubbed 'self-organized criticality.'"
Dear fellow scientists, admire us for the 1% of the cases when things like "oh i have a very simple theory about this" are brilliant and dont hate us for the 99% of the cases where this is just idiotic and arrogant.
http://xkcd.com/793/
The really interesting thing will be when Randall does a comic about how you can get easy upvotes for "oblig xkcd" posts.
The objective reality is that this process has been observed to happen in the brain. Repeatedly; consensually; experientially.
The open question, at least for me, is, is there any reason to think that this is the only, or even the primary, mode of neural operation?
Sand will indeed avalanche following the power law when it's poured on top of itself. But it does something completely different when it is suspended in turbulent water, or melted into glass, or just sitting there on the beach (seems to have an affinity for the inside of bathing suits as I recall, though it's been a while.)
Perhaps avalanche at criticality is "the" answer. But I think we're quite some distance from declaring that particular win. I'm all for the exploration, though.
I've fallen off your lawn, and I can't get up.
http://www.smbc-comics.com/?id=2556
Ok, well... my understanding of it is that nature is made up by random events. If those events were all there were, you'd get white noise. A perfectly even randomness. However, nature also has laws. With regard to sand, there's gravity, and slope, friction, etc... and that means these randomly falling grains of sand, on the macro scale, end up forming patterns. These patterns end up being very complex but predictable with statistics. Understanding a dune from the point of view of a grain of sand is nearly impossible. You just need to know the rules the system is following and then you can make accurate macro-scale predictions without having to compute every grain of sand in the dune.
The arguments made its way into nearly every branch of science now. Our attempts at brute forcing nature, and trying to connect the sub-atomic scale with the macro scale have mostly failed. But it now seems that maybe nature doesn't work that way. Nature seems more to work based on sets of probabilities, and particles seem to work more like "attributes" than matter. So perhaps the brain works like this to. It's a collection of chaos, bound by rules. Those rules cause the microscopic chaos to form patterns on the macro scale.
The theory is an overarching idea of how the brain works and best makes sense when compared to other theories. One (not this theory) way to think of the brain is that it is like a computer, with specialized areas which each calculate for specific functions and having a whole mess of complicated parts that evaluate against each other and somehow all work together. This theory instead sees the areas not as having logically complicated interlocking parts, but as each part having a sort of pile where if enough stuff (hormones, electric potential, etc.) is piled onto it, it performs its action. Often this action will include piling more stuff onto other areas piles and then resetting to a baseline.
This theory better explains how the brain can operate in a logical, deterministic fashion while allowing for easy error correction. A computer-like brain would continue to use bad data and damaged instructions could cause whole parts of the brain to fail permanently. The "piles of sand" resetting to a baseline model would accept the bad data once, reset to base, and move on. Damaged instructions (mislinked neurons, brain damage, etc.) could continue to send to much or too little "stuff" to other nodes or wrong nodes, but a system which monitors what is considered "normal" and resets to such will eventually be able to re-normalize every node that isn't directly damaged.
The pendulum regarding self-organized criticality is beginning to swing back in the other direction: many researchers now believe it's being over-applied, and the "power law" distributions that people see for natural phenomenon that are "evidence" of S.O.C. have been shown to not actually obey power laws (it's really easy to make these kinds of mistakes when you make your graphs on log-log scales). Sorry if that was a bit dense, but the long and short of it is that not everything that is being touted as an example of self-organized criticality likely is. For instance, the Bak–Tang–Wiesenfeld sandpile (Bak being the one from TFA)? Turns out it's a HORRIBLE model for how real sandpiles behave.
A lot of the above really needs citations, but I'm too tired and lazy, sorry. To "back this up," let me just say that I have a Ph.D in physics, specializing in nonlinear dynamics, and the above comes from a graduate-level course I took from a professor who knows her shit.
The linked article was horribly written. I'll give a shot at trying to explain it (or rather, a really, really simplified version).
Two of the fundamental problems that neural circuits must solve are the noise-saturation dilemma and the stability-plasticity dilemma. The first is best explained in the context of vision. Our visual system is capable of detecting contrast (ie. edges) over a massive range of brightness, spanning a space of about 10^10. Given that neurons have limited firing rates (typically between 0 and 200hz), there needs to be some normalization criteria that allows useful contrast processing over massive variations in absolute input (more on this later). The stability-plasticity dilemma is that the brain needs to be sufficiently flexible to learn based on a single event (let's say, touching a hot stove is a bad idea), but once learned memories have to be sufficiently stable to last the rest of a creatures' life span.
The stability-plasticity dilemma implies that neural circuits must operate in at least two (as I said, very simplified) distinct states, a "resting" or "maintenance" state, and a "learning" state, and that there is a phase-transition point in between them. Furthermore, these states need to have the following properties regarding stability:
1) the learning state must collapse into the maintenance state in the absence of input (otherwise you get epilepsy).
2) reasonable stimulation (input) during the resting state must be able to trigger a phase change into the learning state (or you become catatonic).
Many circuits/mechanisms have been proposed to explain how the brain solves these dilemmas. Most of them involve the definition of a recurrent neural network using some combination of gated-diffusion and oscillatory dynamics to fit well known oscillatory and wave-based dynamics that have been recorded in neural circuits. Some of these models employ intrinsic learning using a learning-rule (ie. self-organized maps) while others are fit by the researcher. One key point about this class of models (as opposed to the TFA approach) is that they have a macro-circuit architecture specified by the modeler. Typically these models are at least somewhat sensitive to parametric perturbation.
TFA describes another approach, which comes out of research on cellular automata done by Ulam, von Neumann, Conway and Wolfram. This approach posits that parametric stability and macro-circuit organization is only loosely important so long as the system obeys a certain set of rules regarding local interaction (could also be through of as micro-circuit) because it will self-organize to a point of 'critical stability'. In the the two-state model described above, this approach predicts that neural circuits are always at a state of 'critical stability' where maintenance occurs through frequent small perturbations or avalanches, and any new input will trigger a large avalanche, causing learning. Bak has proposed this as a general model of neural circuit organization. One trademark of these type of models is that they show 'scale free' or 'power law' behavior, where the size of an event is inversely proportional to its frequency by some exponential function. Some recent data has shown power-law dynamics in neural populations (a lot of other data doesn't show power-law dynamics).
One big problem with the critical stability hypothesis is that it doesn't deal well with the noise-saturation dilemma: it needs to cause the same general size of avalanche whether it's hit by one grain of sand, or 10^10 grains of sand.
None of this is particularly new, neural-avalanches (albeit in a different context) were postulated in the early 70s. Could some systems in the brain exploit self-organized criticality? Sure, but there is a lot of data out there that's inconsistent with it being the primary method of neural organization.
Actually, since neurons have functional homeostatic pruning and nonlinear membrane responses, there are quite large values of zero when we're recording firing rate.
Look up "boids". Each critter has a field of view and a current direction. It only responds to what it sees in that field-of-view. If other critters start running, it starts running too. If they stop, it stops. With fish, the minute one turns, there is a flash of light. That instructs all the others to turn as well, providing the flash is bright enough. Maybe it takes two or more.
Vintage computer adverts: http://www.vintageadbrowser.com/computers-and-software-ads
Can We stop artificially dividing the brain and the Mind, please? The two ideas are likely the same and their distinction speaks of lingering medieval mysticism.
See that? I even put in the gratuitous capitalization!