Note: This was intended to be a brief note. Instead it developed into a monster post that took me two weeks of stolen moments to write. It’s very light on non-blog references but they exist. Nevertheless, it is still easy to find a number of oversimplifications, conflations, and other imperfections below. The general thrust of the argument however remains.
How Far Can You Trust a Neuroscientist?
Image via Wikipedia
A couple of days ago I watched a TED talk called the Linguistic Genius of Babies by Patricia Kuhl. I had been putting it off, because I suspected I wouldn’t like it but I was still disappointed at how hidebound it was. It conflated a number of really unconnected things and then tried to sway the audience to its point of view with pretty pictures of cute infants in brain scanners. But all it was, is a hodgepodge of half-implied claims that is incredibly similar to some of the more outlandish claims made by behaviorists so many years ago. Kuhl concluded that brain research is the next frontier of understanding learning. But she did not give a simple credible example of how this could be. She started with a rhetorical trick. Mentioned an at-risk language with a picture of a mother holding an infant facing towards her. And then she said (with annoying condescension) that this mother and the other tribe members know something we do not:
What this mother — and the 800 people who speak Koro in the world — understand that, to preserve this language, they need to speak it to the babies.
This is garbage. Languages do not die because there’s nobody there to speak it to the babies (until the very end, of course) but because there’s nobody of socioeconomic or symbolic prestige children and young adults can speak the language to. Languages don’t die because people can’t learn them, they die because they have no reason (other than nostalgia) to learn them or have a reason not to learn them. Given a strong enough reason they would learn a dying language even if they started at sixteen. They just almost never are given the reason. Why Kuhl felt she did not need to consult the literature on language death, I don’t know.
Patricia Kuhl has spent the last 20 years studying pretty much one thing: acoustic discrimination in infants (http://ilabs.washington.edu/kuhl/research.html). Her research provided support for something that had been already known (or suspected), namely that young babies can discriminate between sounds that adults cannot (given similar stimuli such as the ones one might find in the foreign language classroom). She calls this the “linguistic genius of babies” and she’s wrong:
Babies and children are geniuses until they turn seven, and then there’s a systematic decline.
First, the decline (if there is such a thing) is mostly limited to acoustic processing and even then it’s not clear that the brain is the thing that causes it. Second, being able to discriminate (by moving their head) between sounds in both English and Mandarin at age 6 months is not a sign of genius. It’s a sign of the baby not being able to differentiate between language and sound. Or in other words, the babies are still pretty dumb. But it doesn’t mean they can’t learn a similar distinction at a later age – like four or seven or twelve. They do. They just probably do it in a different way than a 6-month old would. Third, in the overall scheme of things, acoustic discrimination at the individual phoneme level (which is what Kuhl is studying) is only a small part of learning a language and it certainly does NOT stop at 7 months or even 7 years of age. Even children who start learning a second language at the age of 6 achieve a native-like phonemic competence. And even many adults do. They seem not to perform as well on certain fairly specialized acoustic tests but functionally, they can be as good as native speakers. And it’s furthermore not clear that accent deficiencies are due to the lack of some sort of brain plasticity. Fourth, language learning and knowledge is not a binary thing. Even people who only know one language know it to a certain degree. They can be lexically, semantically and syntactically quite challenged when exposed to a sub-code of their language they have little to no contact with. So I’m not at all sure what Kuhl was referring to. François Grosjean (an eminent researcher in the field) has been discussing all this on his Life as Bilingual blog (and in books, etc.). To have any credibility, Kuhl must address this head on:
There is no upper age limit for acquiring a new language and then continuing one’s life with two or more languages. Nor is there any limit in the fluency that one can attain in the new language with the exception of pronunciation skills.
Instead she just falls on old prejudices. She simply has absolutely nothing to support this:
We think by studying how the sounds are learned, we’ll have a model for the rest of language, and perhaps for critical periods that may exist in childhood for social, emotional and cognitive development.
A paragraph like this may get her some extra funding but I don’t see any other justification for it. Actually, I find it quite puzzling that a serious scholar would even propose anything like this today. We already know there is no critical period for social development. Well, we don’t really know what social development is, but there’s no critical brain period to what there is. We get socialized to new collective environments throughout our lives.
But there’s no reason to suppose that learning to interact in a new environment is anything like learning to discriminate between sounds. There are some areas of language linked to perception where that may partly be the case (such as discriminating shapes, movements, colors, etc.) but hardly things like morphology or syntax, where much more complexity is involved. But this argument cuts both ways. Let’s say a lot of language learning was like sound development. And we know most of it continues throughout life (syntax, morphology, lexicon) and it doesn’t even start at 6 months (unless you’re a crazy Chomskean who believes in some sort of magical parameter setting). So if sound development was like that, maybe it has nothing to do with the brain in the way Kuhl imagines – although she’s so vague that she could always claim that that’s what she’d had in mind. This is what Kuhl thinks of as additional information:
We’re seeing the baby brain. As the baby hears a word in her language the auditory areas light up, and then subsequently areas surrounding it that we think are related to coherence, getting the brain coordinated with its different areas, and causality, one brain area causing another to activate.
So what? We know that that’s what was going to happen. Some parts of the brain were going to light up as they always do. What does that mean? I don’t know. But I also know that Patricia Kuhl and her colleagues don’t know either (at least not in the way she pretends). We speak a language, we learn a language and at the same time we have a brain and things happen in the brain. There are neurons and areas that seem to be affected by impact (but not always and not always in exactly the same way). Of course, this is an undue simplification. Neuroscientists know a huge amount about the brain. Just not how it links to language in a way that would say much about the language that we don’t already know. Kuhl’s next implied claim is a good example of how partial knowledge in one area may not at all extend to knowledge in another area.
What you see here is the audio result — no learning whatsoever — and the video result — no learning whatsoever. It takes a human being for babies to take their statistics. The social brain is controlling when the babies are taking their statistics.
In other words, when the children were exposed to audio or video as opposed to a live person, no effect was shown. At 6 months of age! As is Kuhl’s wont, she only hints at the implications, but over at the Royal Society’s blog comments, Eric R. Kandel has spelled it out:
I’m very much taken with Patricia Kuhl’s finding in the acquisition of a second language by infants that the physical presence of a teacher makes enormous difference when compared to video presence. We all know from personal experience how important specific teachers have been. Is it absurd to think that we might also develop methodologies that would bring out people’s potential for interacting empathically with students so that we can have a way of selecting for teachers, particularly for certain subjects and certain types of student? Neuroscience: Implications for Education and Lifelong Learning.
But this could very well be absurd! First, Kuhl’s experiments were not about second language acquisition but sensitivity to sounds in other languages. Second, there’s no evidence that the same thing Kuhl discovered for infants holds for adults or even three-year olds. A six-month old baby hasn’t learned yet that the pictures and sounds coming from the machine represent the real world. But most four-year olds have. I don’t know of any research but there is plenty of anecdotal evidence. I have personally met several people highly competent in a second language who claimed they learned it by watching TV at a young age. A significant chunk of my own competence in English comes from listening to radio, audio books and watching TV drama. How much of our first language competence comes from reading books and watching TV? That’s not to say that personal interaction is not important – after all we need to learn enough to understand what the 2D images on the screen represent. But how much do we need to learn? Neither Kuhl nor Kandel have the answer but both are ready (at least by implication) to shape policy regarding language learning. In the last few years, several reports raised questions about some overreaching by neuroscience (both in methods and assumptions about their validity) but even perfectly good neuroscience can be bad scholarship in extending its claims far beyond what the evidence can support.
The Isomorphism Fallacy
This section of the post is partly based on a paper I presented at a Czech cognitive science conference about 3 years ago called Isomorphism as a heuristic and philosophical problem.
The fundamental problem underlying the overreach of basic neuroscience research is the fallacy of isomorphism. This fallacy presumes that the same structures we see in language, behavior, society must have structural counterparts in the brain. So there’s a bit of the brain that deals with nouns. Another bit that deals with being sorry. Possibly another one that deals with voting Republican (as Woody Allen proved in “Everyone Says I Love You“). But at the moment the evidence for this is extremely weak, at best. And there is no intrinsic need for a structural correspondence to exist. Sidney Lamb came up with a wonderful analogy that I’m still working my way through. He says (recalling an old ‘Aggie‘ joke) that trying to figure out where the bits we know as language structure are in the brain is like trying to work out how to fit the roll that comes out of a tube of tooth paste back into the container. This is obviously a fool’s errand. There’s nothing in the tooth-paste container that in any way resembles the colorful and tubular object we get when we squeeze the paste container. We get that through an interaction of the substance, the container, external force, and the shape of the opening. It seems to me entirely plausible, that the link between language and the brain is much more like that between the paste, the container and their environment than like that between a bunch of objects and box. The structures that come out are the result of things we don’t quite understand happening in the brain interacting with its environment. (I’m not saying that that’s how it is, just that it’s plausible.) The other thing to lends it credence is the fact that things like nouns or fluency are social constructs with fuzzy boundaries, not hard discrete objects, so actually localizing them in the brain would be a bit of a surprise. Not that it can’t be done, but the burden of evidence of making this a credible finding is substantial.
Now, I think that the same problem applies to looking for isomorphism the other way. Lamb himself tries to look at grammar by looking for connections resembling the behavior of activating neurons. I don’t see this going anywhere. George Lakoff (who influenced me more than any other linguist in the world) seems to think that a Neural Theory of Language is the next step in the development of linguistics. At one point he and many others thought that mirror neurons say something about language but now that seems to have been brought into question. But why do we need mirror neurons when we already know a lot of the immitative behaviors they’re supposed facilitate? Perhaps as a treatment and diagnostic protocol for pathologies but is this really more than story-telling? Jerome Feldman described NTL in his book “From Molecule to Metaphor” but his main contribution seems to me lies in showing how complex language phenomena can be modelled with brain-like neural networks, not saying anything new about these phenomena (see here for an even harsher treatment). The same goes for the Embodied Construction Grammar. I entirely share ECG’s linguistic assumptions but the problem is that it tries to link its descriptive apparatus directly to the formalisms necessary for modeling. This proved to be a disaster for the generative project that projected its formalisms into language with a imperfect fit and now spends most of its time refining those formalisms rather than studying language.
So far I don’t see any advantage in linking language to the brain in either the way Kuhl et al or Feldman et al try to do it (again with the possible exception of pathologies). In his recent paper on compositionality, Feldman describes research that shows that spacial areas are activated in conjunction with spatial terms and that sentence processing time increases as the sentence gets removed from “natural spatial orientation”. But brain imaging at best confirms what we already knew. But how useful is that confirmatory knowledge? I would argue that not very useful. In fact there is a danger that we will start thinking of brain imaging as a necessary confirmation of linguistic theory. Feldman takes a step in this dangerous direction when he says that with the advent of new techniques of neuroscience we can finally study language “scientifically”. [Shudder.]
We know there’s a connection between language and the brain (more systematic than with language and the foot, for instance) but so far nobody’s shown convincingly that we can explain much about language by looking at the brain (or vice versa). Language is best studied as its own incredibly multifaceted beast and so is the brain. We need to know a lot more about language and about the brain before we can start projecting one into the other.
And at the moment, brain science is the junior partner, here. We know a lot about language and can find out more without looking for explanations in the brain. It seems as foolish as trying to illuminate language by looking inside a computer (as Chomsky’s followers keep doing). The same question that I’m asking for language was asked about cognitive processes (a closely related thing) by William Uttal in The New Phrenology who’s asking “whether psychological processes can be defined and isolated in a way that permits them to be associated with particular brain regions” and warns against a “neuroreductionist wild goose chase” – and how else can we characterize Kuhl’s performance – lest we fall “victim to what may be a ‘neo-phrenological’ fad”. Michael Shremer voiced a similar concern in the Scientific American:
The brain is not random kludge, of course, so the search for neural networks associated with psychological concepts is a worthy one, as long as we do not succumb to the siren song of phrenology.
What does a “siren song of phrenology” sound like? I imagine it would sound pretty much like this quote by Kuhl:
We are embarking on a grand and golden age of knowledge about child’s brain development. We’re going to be able to see a child’s brain as they experience an emotion, as they learn to speak and read, as they solve a math problem, as they have an idea. And we’re going to be able to invent brain-based interventions for children who have difficulty learning.
I have no doubt that there are some learning difficulties for which a ‘brain-based intervention’ (whatever that is) may be effective. But it’s just a relatively small part of the universe of learning difficulties that it hardly warrants a bombastic claim like the one above. I could find nothing in Kuhl’s narrow research that would support this assertion. Learning and language are complex psycho-social phenomena that are unlikely to have straightforward counterparts in brain activations such as can be seen by even the most advanced modern neuroimaging technology. There may well be some straightforward pathologies that can be identified and have some sort of treatment aimed at them. The problem is that brain pathologies are not necessarily opposites of a typically functioning brain (a fallacy that has long plagued interpretation of the evidence from aphasias) – it is, as brain plasticity would suggest, just as likely that at least some brain pathologies simply create new qualities rather than simply flipping an on/off switch on existing qualities. Plus there is the historical tendency of the self-styled hard sciences to horn in on areas where established disciplines have accumulated lots of knowledge, ignore the knowledge, declare a reductionist victory, fail and not admit failure.
For the foreseeable future, the brain remains a really poor metaphor for language and other social constructs. We are perhaps predestined to finding similarities in anything we look at but researchers ought to have learned by now to be cautious about them. Today’s neuroscientists should be very careful that they don’t look as foolish to future generations as phrenologists and skull measurers look to us now.
In praise of non-reductionist neuroscience
Let me reiterate, I have nothing against brain research. The more of it, the better! But it needs to be much more honest about its achievements and limitations (as much as it can given the politics of research funding). Saying the sort of things Patricia Kuhl does with incredibly flimsy evidence and complete disregard for other disciplines is good for the funding but awful for actually obtaining good results. (Note: The brevity of the TED format is not an excuse in this case.)
A much more promising overview of applied neuroscience is a report by the Royal Society on education and the brain that is much more realistic about the state of neurocognitive research who admit at the outset: “There is enormous variation between individuals, and brain-behaviour relationships are complex.”
The report authors go on to enumerate the things they feel we can claim as knowledge about the brain:
- The brain’s plasticity
- The brain’s response to reward
- The brain’s self-regulatory processes
- Brain-external factors of cognitive development
- Individual differences in learning as connected to the brain and genome
- Neuroscience connection to adaptive learning technology
So this is a fairly modest list made even more modest by the formulations of the actual knowledge. I could only find a handful of statements made to support the general claims that do not contain a hedge: “research suggests”, “may mean”, “appears to be”, “seems to be”, “probably”. This modesty in research interpretation does not always make its way to the report’s policy suggestions (mainly suggestions 1 and 2). Despite this, I think anybody who thinks Patricia Kuhl’s claims are interesting would do well do read this report and pay careful attention to the actual findings described there.
Another possible problem for those making wide reaching conclusions is a relative newness of the research on which these recommendations are based. I had a brief look at the citations in the report and only about half are actually related to primary brain research. Of those exactly half were published in 2009 (8) and 2010 (20) and only two in the 1990s. This is in contrast to language acquisition and multilingualism research which can point to decades of consistently replicable findings and relatively stable and reliable methods. We need to be afraid, very afraid of sexy new findings when they relate to what is perceived as the “nature” of humans. At this point, as a linguist looking at neuroscience (and the history of the promise of neuroscience), my attitude is skeptical. I want to see 10 years of independent replication and stable techniques before I will consider basing my descriptions of language and linguistic behavior on neuroimaging. There’s just too much of ‘now we can see stuff in the brain we couldn’t see before, so this new version of what we think the brain is doing is definitely what it’s doing’. Plus the assumption that exponential growth in precision brain mapping will result in the same growth in brain function identification is far from being a sure thing (cf. genome decoding). Exponential growth in computer speed, only led to incremental increases in computer usability. And the next logical step in the once skyrocketing development of automobiles was not flying cars but pretty much just the same slightly better cars (even though they look completely different under the hood).
The sort of knowledge to learn and do good neuroscience is staggeringly awesome. The scientists who study the brain deserve all the personal accolades they get. But the actual knowledge they generate about issues relating to language and other social constructs is much less overwhelming. Even a tiny clinical advance such as helping a relatively small number of people to communicate who otherwise wouldn’t be able to express themselves makes this worthwhile. But we must not confuse clinical advances with theoretical advances and must be very cautious when applying these to policy fields that are related more by similarity than a direct causal connection.