Book Review: The Root of Thought

by Scott Alexander4 min read22nd Jul 201091 comments

59

Book ReviewsNeuroscience
Frontpage

Related to: Brain Breakthrough! It's Made of Neurons!

I can't really recommend Andrew Koob's The Root of Thought. It's poorly written, poorly proofread, lacking much more information than is in the Scientific American review, and comes across as about one part neuroscience to three parts angry rant. But it does present an interesting hypothesis and an interesting case study on a major failure of rationality.

Only about ten percent of the brain is made of neurons; the rest is a diverse group of cells called "glia". "Glia" is Greek for glue, because the scientists who discovered them decided that, since they were in the brain and they weren't neurons, they must just be there to glue the neurons together. Since then, new discoveries have assigned glial cells functions like myelination, injury repair, immune defense, and regulation of blood flow: all important, but mostly things only a biologist could love. The Root of Thought argues that glial cells, especially a kind called astrocytes, are also important in some of the higher functions of thought, including memory, cognition, and maybe even creativity. This is interesting to neuroscientists, and the story of how it was discovered is also interesting to us as aspiring rationalists.

Glial cells involved in processing

Koob's evidence is indirect but suggestive. He points out that more intelligent animals have a higher astrocyte to neuron ratio than less intelligent animals, all the way from worms with one astrocyte per thirty neurons, to humans with an astrocyte: neuron ratio well above one. Within the human brain, the areas involved in higher thought, like the cortex, are the ones with the highest astrocyte:neuron ratio, and the most down-to-earth, like the cerebellum, have barely any astrocytes at all. Especially intelligent humans may have higher ratios still: one of the discoveries made from analyzing Einstein's brain was that he had an unusually large number of astrocytes in the part of his brain responsible for mathematical processing. And learning is a stimulus for astrocyte development. When canaries learn new songs, new astrocytes grow in the areas responsible for singing.

Astrocytes have a structure especially suited for learning and cognition. They have their own gliotransmitters, similar in function to neurotransmitters, and they communicate with one another, sparking waves of astrocyte activity across areas of the brain. Like neurons, they can enter an active state after calcium release, but unlike neurons, which get calcium only when externally activated, astrocytes can fill with calcium either because of external stimuli or when their own calcium stores randomly leak out into the cell, a process which resembles the random, unprovoked nature of thought during sensory deprivation and dreaming.

Astrocytes also affect and are affected by neurons. Each astrocyte "monitors" thousands of synapses, and releases calcium based on the input it receives. Output from astrocytes, in turn, affects the behavior of neurons. Astrocytes can take up or break down neurotransmitters, which changes the probability of nearby neurons activating, and they can alter synapses, promoting some and pruning others in a process likely linked to long-term potentiation in the brain.

Although it wasn't in the book, very recent research shows a second type of glial cell, the immune-linked microglia, play a role in behavior that may be linked to obsessive-compulsive disorder; a microglia-altering bone marrow transplant cures an OCD-like disease in mice.

By performing computations that influence the firing of neurons, glial cells at the very least play a strong supporting role in cognition. Koob goes way beyond that (and really beyond what he can support) and argues that actually neurons play a supporting role to glia, being little more than the glorified wires that relay astroglial commands. His argument is very speculative and uses words like "could" a lot, but the evidence at least shows that glia are more important than a century of neurology has given them credit for.


We don't know how much we don't know about cognitive science

Previous Less Wrong articles, for example Artificial Addition, have warned against trying to replicate a process without understanding it by copying a few surface features. One of the most popular such ideas is to replicate the brain by copying the neurons and seeing what happens. For example, IBM's Blue Brain project hopes to create an entire human brain by modeling it neuron for neuron, without really understanding why brains work or why neurons do what they do1.

We've made a lot of progress in cognitive science in the past century. We know where in the brain various activities take place, we know the mechanisms behind some of the more easily studied systems like movement and perception, and we've started researching the principles of intelligence that the brain must implement to do what it does. It's tempting to say that we more or less understand the brain, and the rest is just details. One of the take-home messages from this book is that, although cognitive scientists can justifiably be proud of their progress, our understanding still hasn't even met the low bar of being entirely sure we're even studying all the right kinds of cells, and this calls into question our ability to meet the higher bar of being able to throw what we know into a simulator and hope it works itself out.

A horrible warning about community irrationality

In the late 19th century, microscopy advanced enough to look closely at the cellular structure of the brain. The pioneers of neurology decided that neurons were interesting and glia were the things you had to look past to get to the neurons. This theory should have raised a big red flag: Why would the brain be filled with mostly useless cells? But for about seventy five years, from the late 19th century to the mid to late 20th, no one seriously challenged the assumption that glia played a minor role in the brain.

Koob attributes the glia's image problem to the historical circumstances of their discovery. Neurons are big, peripherally located, and produce electrical action potentials. This made them both easy to study and very interesting back in the days when electricity was the Hot New Thing. Scientists first studied neurons in the periphery, got very excited about them, and later followed them into the brain, which turned out to be a control center for all the body's neurons. This was interesting enough that neurologists, people who already had thriving careers in the study of neurons, were willing to overlook the inconvenient presence of several other types of cells in the brain, which they relegated to a supporting role. The greatest of these early pioneers of neurology, Santiago Ramon y Cajal, was the brother of the neurologist who first proposed the idea that glial cells functioned as glue and may have (Koob theorizes) let familial loyalty influence his thinking. The community took his words as dogma and ignored glia for years, a choice no doubt made easier by all the exciting discoveries going on around neurons. Koob discussed the choice facing neuroscientists in the early 20th century: study the cell that seemed on the verge of yielding all the secrets of the human mind, or tell your advisor you wanted to study glue instead. Faced with that decision, virtually everyone chose to study the neurons.

There wasn't any sinister cabal preventing research into glia. People just didn't think of it. Everyone knew that neurons were the only interesting type of cell in the brain. They assumed that if there was some other cell that was much more common and also very important, somebody would have noticed. I've read neuroscience books, I read the couple of paragraphs where they mentioned glial cells, and I shrugged and kept reading, because I assumed if they were hugely important somebody would have noticed.

The heuristic, that an entire community doesn't just miss low-hanging fruit, is probably a good one and as many people have pointed out the vast majority of people who think they've found something that the scientific community has missed are somewhere between wrong and crackpot. Science is usually pretty good at finding and recognizing its mistakes, and even in the case of glial cells they did eventually find and recognize the mistake. It just took them a century.

One common theme across Less Wrong and SIAI is that there are some relatively little-known issues that, upon a moderate amount of thought, seem vitally important. And one of the common arguments against this theme is that if this were true, surely somebody would have noticed. The lesson of glial cells is that sometimes this just doesn't happen.

Related: Glial Cells: Their Role In Behavior, Underappreciated Star-Shaped Cells May Help Us Breathe, Glial Cells Aid Memory Formation, New Role For Supporting Brain Cells, Support Cells, Not Neurons, Lull Brain To Sleep

59

91 comments, sorted by Highlighting new comments since Today at 12:11 PM
New Comment
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

Glial cell are actually about 1:1. A few years ago a researcher wanted to cite something to back up the usual 9:1 figure, but after asking everybody for several months nobody knew where the figure came from. So, they did a study themselves and did a count and found it to be 1:1. I don't have the reference on me, it was a talk I went to about a year ago (I work at a neuroscience research institute).

I have asked a number of neuroscientists about the importance of glia and have always received the same answer: the evidence that they are functionally important is still "very weak". They might be wrong, but given that some of these guys could give hour long lectures on exactly why they think this, and know the few works that claim otherwise... I'm inclined to believe them.

1Alan11yThis new finding may be correct, but the old dictum about "nullius in verba" still makes sense.

One of the most popular such ideas is to replicate the brain by copying the neurons and seeing what happens. For example, IBM's Blue Brain project hopes to create an entire human brain by modeling it neuron for neuron, without really understanding why brains work or why neurons do what they do.

No, the Blue Brain project (no longer affiliated with IBM, AFAIK) hopes to simulate neurons to test our understanding of how brains and neurons work, and to gain more such understanding.

If you can simulate brain tissue well enough that you're reproducing the actual biological spike trains and long-term responses to sensory input, you can be pretty sure that your model is capturing the relevant brain features. If you can't, it's a pretty good indication that you should go study actual brains some more to see if you're missing something. This is exactly what the Blue Brain project is: simulate a brain structure, compare it to an actual rat, and if you don't get the same results, go poke around in some rat brains until you figure out why. It's good science.

I have a new idea for AI: glial networks!

Be sure to add some extra complexity so that you can get more emergence out!

5xamdam11yWe can optimize the emergence/complexity ratio with an evolutionary algorithm... Sorry Chalmers.
3Will_Newsome11yGlial networks would totally work. If you have enough of them, eventually you'll get Boltzmann AI.

I think that science usually works a little differently. People do not choose what they are going to investigate by what is not boring or is a hot topic. Very often they look for (metaphorically) a chink where they can put a crowbar in and open a crack to see some new knowledge. It was a lot easier to study neurons than glia - they stained well, their activity could be measured (a bit) without opening the skull, their electrical potentials could be measured, in some animals they were extremely large etc. Glial cells were not that forthcoming with their secrets and so they had to wait. That glial cells are not glue or just support has been known for at least a decade but what they might be doing was not (and still isn't) easy to discover. They are not boring - they involve the regulation of calcium ions and calcium ions are very definitely not boring to anyone interested in cellular communication.

The other big motivator is what the grant money is following.

A further note on staining: pioneer neurobiologist Ramon y Cajal got a lot of mileage out of a staining technique which, for reasons he didn't understand, only stained a small fraction of neurons. Bingo: instead of getting a dense thicket, you get some beautiful branching structures to draw. If his technique had picked out individual astrocytes instead, perhaps glial cells would have gotten more attention.

the choice facing neuroscientists in the early 20th century: study the cell that seemed on the verge of yielding all the secrets of the human mind, or tell your advisor you wanted to study glue instead.

Very plausible! But I would also like to know how this stable state eventually got upset. Who were the first people to study glial cells? Did they have any distinctive characteristics (personality/educational background/institutional affiliation/...) or did anything in particular happen that prompted them to take an interest in glue? It seems that if we could replicate that miracle, it could be very beneficial for science.

astrocytes can fill with calcium either because of external stimuli or when their own calcium stores randomly leak out into the cell, a process which resembles the random, unprovoked nature of anything that's random.

Copyedit: Extra space in "astrocyte: neuron" in third paragraph.

4Vladimir_Nesov11yYou can use private messages to report typos.
3Mass_Driver11yAnd should, unless they're the sort of typos that might seriously confuse someone before it gets fixed.
[-][anonymous]11y 5

Thanks for this. I suspect there's quite a bit of value in every so often discarding a basic working assumption, and seeing where it takes you. Of course, do this too often and you won't make any progress at all, so it's an interesting problem.

Also in the vein of "things we think we know about the brain that might be completely wrong", there's good evidence that our model of nerve propagation (electrical signals mediated by ion channels) isn't actually how the brain works - see this paper, and a good explanation of it here and here

5NancyLebovitz11yDo you have any heuristics for identifying basic working assumptions?
6[anonymous]11yA good one might be "the things you would have to mention first if you were explaining your field/problem/whatever to a person with no knowledge of it". For example, if you're explaining A.I. to someone, something that will come up very quickly is that human-level intelligence is almost certainly not the maximum possible, due to the constraints of biology and evolution. What then, would be the result if you ignore or reverse this, and act as if human-level intelligence IS the maximum possible?

That will get you to the category of reversing your conscious premises, but possbily not to examining something like "glial cells are too boring to bother with".

0Mass_Driver11yWhy not?
0[anonymous]11yGood point - that sort of assumption is much harder to isolate. I'll have to mull this over for a while.
3SilasBarta11yUnfortunately, what I've found to be commonplace (and am writing an article on) is that the very same people who don't have a deep understanding (and thus don't know what assumptions interplay with their work) are also the ones who are incapable of explaining the field to outsiders.
2Cyan11yI hope your article mentions and/or addresses the notion that deep understanding is necessary but not sufficient for the capability of explaining the field to outsiders.
3SilasBarta11yYes, I plan to critique that notion. The reason why I don't buy it is that I take a "deep understanding" to be a Level 2 [http://lesswrong.com/lw/1yq/understanding_your_understanding/] understanding, in which you recognize far-reaching inferential flows between that field and the rest of your knowledge. This would mean there are arbitrarily many inferential paths you can take from your understanding of the field to the nepocu (nearest point of common understanding) between you and the outsider. If one path isn't clicking, then you should be able to "fall back a rank" to the grounding concepts, if the listener had a tenuous grasp on those, or simply take another path. To the extent that you can't do this, that causes me to call into question just how well you understand. The caveat, of course, is that no matter how good your understanding, it can become time-consuming if the inferential distance is great, or you can't get immediate feedback about which concepts are confusing. As for the article, its current status is that I've dropped the idea of listing my whole back of tricks in one article (it got really long), and the first one will just focus on (what I consider to be) the critical connection between understanding and explanation capability, and on the importance of locating the nepocu (defined above, and yes I've already been told it's a flaky term).
2Cyan11yYou're assuming that the explainer knows enough about explaining to try to identify the nepocu and to solicit feedback about which concepts are confusing.
0SilasBarta11yBecause it's a good assumption. Explaining is nothing but tracing out your own internal model's inferential relationships between the concepts. The only bar to this would be not knowing it. So I don't see what kind of "explaining skill" there is that goes above and beyond that. Soliciting feedback, for its part, is but a matter of asking, "do you understand [link in my ontology]?" and/or watching and listening for when they say the don't understand. (And I take it you don't find the term "nepocu" to be particularly annoying?)
5Mass_Driver11yNo, I think Cyan is right. Have you read Eliezer's "A Technical Explanation of Technical Explanation?" You may wish to write "A Lay Explanation of Lay Explanation." I would certainly read and probably vote up such an article.
1SilasBarta11yI don't see, though, how I'm describing a different kind of explanation, or a distinctly lay one. The explanation standards I'm giving are what you would need to give for a technical explanation as well, in the case where your listener starts from a point of less knowledge about your field (i.e. a far nepocu). The technical explanation only differs in terms of its greater detail (afaict -- you may mean something else); it doesn't change in type.
4Cyan11yThis is exactly the point under dispute. I'm open to evidence on this point, but you'll have to do better than flat assertion. As a skilled explainer, you may not be aware of some things you do automatically that do not come naturally to less skilled explainers -- even those who are competent within their domains of expertise. I am indifferent to the term "nepocu". Obviously some kind of abbreviation is necessary.
2SilasBarta11yAre you sure you're not overparsing me there? The part you truncated is the key point, that to explain, you need only trace out your internal ontology. To reject my position, it would have to be possible for someone both to actually know the connection to the nepocu, and be unable to articulate the inferential connection. There are certainly people who have a "Chinese room" understanding, allowing them to deftly match outputs with the right input, and thus meet the standard "expert" threshold. But this is only a level 1 understanding. I do appreciate your input, though, about what I should include.
4Cyan11yIt's entirely possible. I suppose I'd dispute that, then. It seems to me that to explain skillfully, you need to have not just a grasp of your internal ontology, but also a reasonably accurate map of your conversant's internal ontology. One could, in their own head, recognize far-reaching inferential flows between their field of expertise and the rest of their knowledge, and yet fail to recognize that the task of explaining essentially lies is seeking the nepocu and going from there. Level 2 understanding is a property of one individual's internal ontology; seeking the nepocu is in the same class as understanding the typical mind fallacy and the problem of expecting short inferential distances, these being concerned with the relationship between two distinct internal ontologies. But it seems premature to go on with this discussion until you've made the post. I'm happy to continue if you want to (there's no shortage of electrons, after all), but if the post is near completion, it probably makes more sense to wait until it's done.
0SilasBarta11yOkay, point taken. In any case, it would be hard for me to simultaneously claim that understanding necessarily enables you to explain, and that I have advice that would enable you to explain if you only have an understanding. On the other hand, the advice I'm giving is derided as "obvious", but, if it's so obvious, why aren't people following it? But someone doesn't really need to recognize the difference between their own internal ontology and someone else's. In the worst case, they can just abandon attempts to link to the listener's ontology, and "overwrite" with their own, and this would be the obvious next step. In my (admittedly biased) opinion, the reason people don't take this route is not because this would take too long, but because the domain knowledge isn't even well-connected to the rest of their own internal ontology. (Also, this is distinct from the "expecting short inferential distances" problem in that people don't simply expect it short, but that they wouldn't know what to do even if they knew it were very long.) I still think advice would be helpful at this stage. I'll send you what I have so far, up to the understanding / nepocu points.
2RichardKennaway11yI disagree. Your internal model cannot be copied into anyone else's head just by expounding it. To explain something successfully -- that is, to get someone else to understand something -- you have to take account of the state of the person you are explaining it to. An explanation that one person finds a model of clarity, another may find tedious and confusing. (I have seen both reactions to Eliezer's article on Bayes' theorem [http://yudkowsky.net/rational/bayes].) When I am assisting students in a computer laboratory, and a student indicates they have a problem, the question I ask myself when I listen to them is "what information does this student need, and not have?" That is what I seek to provide, not a dump of my own thought processes around the subject. I generally get favourable feedback, so I think I'm onto something here. As a general rule, explanations share this property with software: until you have tried it and seen it work, you do not know that it works.
2SilasBarta11yI agree with and practice all of that, so I was oversimplifying with the part you quoted. I should probably have said something more like, "Explaining starts from tracing out your internal model's inferential relationship between the concepts, and proceeds by finding how it can connect to -- and if necessary, correct -- the listener's ontology."
3billswift11yLook for things that don't look important at first glance. Another example from science history is that proteins were originally considered the likely material of heredity, nucleic acids were overlooked because they were thought to be "too simple" in structure.

I bet that if glial cells had been baptized "mesosynaptic cells" instead, they'd have been studied much more.

Somewhat related: today I found a bug in the software I'm working on that's due to the fact that a variable's name doesn't correspond to what it actually means, which means that a change that "looked right" actually screwed everything up.

Thanks for the interesting article.

and regulation of blood flow: all important, but mostly things only a biologist could love.

I'd argue that people who like designing computer architectures should be interested in this as well.

Ignoring glia seems to me to have been an (mis-)application of assuming the simplest explanation consistent with the facts, when people weren't in a position to fully explain the brain. I.e. people knew that you needed neurons to explain brain function, but because they couldn't predict how the brain functioned, they didn't kno... (read more)

7RobinZ11yI'd agree - I think the reasonable position at this point is to say that we shouldn't privilege the hypothesis [http://wiki.lesswrong.com/wiki/Privileging_the_hypothesis]. Most of the argumentation along those lines that I have seen cited seems to be permissive, rather than compelling, towards the claim.
-12bogus11y
2sharpneli11yConsidering that quantum physics is turing complete (unless it's nonlinear etc) any quantum effects could be reproduced with classical computation. Therefore the assumption that cognition must involve quantum effects implicitly assumes that quantum physics is nonlinear or one of the various other requirements. In this light the first question that ought to be asked from persons claiming quantum effects on brain is: What computation [performed in brain] requires basically infinite loops completed on finite time and based on what physics experiment they believe that quantum effects are more than turing complete.
0whpearson11yI think the brain is probably ultimately computable by a classical computer and yet quantum computing in the brain might be significant. Here are couple of the potential problems we'll have if the brain relies on quantum effects. 1) Difficulty in replacing bits of the brain functionally. If consciousness is some strange transitory gestalt quantum field; then you would need to to make a brain prosthesis that had the same electromagnetic properties as a neuron. Which might be quite hard. 2) A harder time simulating brains/doing AI: You might have to up the date you expect Whole Brain Emulations to become available (depending upon when we expect quantum computers to be useful).
1JoshuaZ11yI'm having trouble parsing your above comment. Are the points labeled 1 and 2 arguments for the presence of quantum computing in the brain or consequences of that belief?
0whpearson11ySorry consequences. I'll edit for clarity.
0sharpneli11yQuantum computing in the brain might be happening, but if we want to understand conciousness it is irrelevant (Unless conciousness is noncomputable where it becomes a claim about quantum physics yet again). It's as relevant as details about transistors or vacuum tubes are for understanding sorting algorithms. Naturally when considering brain prostheses or simulating a brain the actual method with which brain computes is relevant.
0whpearson11yWhoever said that this conversation was about understanding consciousness? Personally I think that that topic is a tarpit, which I prefer to ignore until we know how the brain works.
1sharpneli11yI merely wished to clarify the difference between conciousness and how it is implemented in the brain. I had no intention of implying that it was part of the discussion. On retrospect the clarification was not required. It's just way too common for the two issues to get mixed up, as can be seen on the various threads.