Friday, July 4, 2008

A (supposed) neural theory of language? Where's the beef??

I think it' s safe to say that I am interested in brain and language, and so when I just came across a book that promises a "neural theory of language" my interest and curiosity were piqued -- and I bought the damn thing ... I have been struggling for years with the issue of what a neural theory of language would even look like (see, for example, Dave Embick's and my paper on this issue, "Defining the relations between linguistics and neuroscience"; a new paper on this topic by us is forthcoming), so I was pretty jazzed to spend some quality time (waiting for a flight at Logan Airport) reading this book.

Jerome A. Feldman's From Molecule to Metaphor. A Neural Theory of Language (MIT Press, 2006), alas, is little more than a mediocre intro to cognitive science with a little neuroscience, uncritically adopted. A real disappointment. You can learn *a little* -- and really just a little -- about certain aspects of computational modeling in language processing that are endorsed by Feldman, mostly in the last third of the book. But the promise of unification, or even coherent discussion, of the challenging links "from molecule to metaphor' are not spelled out.

Greg, this book would make you crazy. I kid you not. The book makes a pitch for embodied cognition, but in a relentlessly naive form. The first seven chapters are a kind of 'light-reading' introduction to some neurobiological concepts (neural connection are tuned; cells are complicated; plasticity matters, mirror neurons are awesome -- that sort of thing). Why any of these specific concepts matter for a neural theory of language is never told to us, even though the subheading for this section of the book is How the Brain Computes.

The critical passages, linking brain and language, contain stuff like this: "There is every reason to believe that ideas, concepts, and the like are represented by neural activity. [Eh .. OK .., yes .. DP] The exact circuitry involved is uncertain, but it suffices for us to assume that some stable connection pattern is associated with each word, concept, schema, and so on. page 91" [Oh well - cop out. DP]"While the details remain unclear, the general idea that mental connections are active neural connections is universally accepted ... (page 94)". No kidding! In the crucial sections on the "Computational Bridge" -- a section I was especially looking forward to, because I think some form of computational theory will be crucial -- the message is that all knowledge is embodied, and that mirror neurons provide the key mechanism. OK, whatever.

The book is (a) superficial [basic cog sci stuff, mediocre intro level] and (b) gullible [apparently Feldman has never come across a mirror neuron claim he didn't believe and adopt]. The basic ingredients for a neural theory of language are: mirror neurons; spreading activation as an account of priming; synaptic plasticity. Sure, all great ideas (even the mirror neurons are a great idea, even if they just don;t work) -- BUT: nothing, and I mean nothing is said at any level of detail about how this is supposed to account for even the most banal aspects of language. Understanding as simulation? Maybe.

There is some rhetoric about language wars -- Chomsky is bad/badder/baddest, Pinker is not very good, Jackendoff also leaves something to be desired -- but there is never anything to sink your teeth into regarding the neurobiological mechanisms that form the basis for any aspect of language. As I said, a disappointment. I'm sure that Feldman has made many important contributions to computational cognitive science and computational lingustics. This book is not one of them.


Anonymous said...

Thanks for engaging, it's better than being ignored.The book is intended as an introduction, but it does have technical content that constrains theories of language by findings in behavioral and neurosciences, brought together by
positive and negative computational results. We are actively engaged in trying to understand the detailed circuitry underlying concepts, but
the results hold more generally.

There is a long history of emotional language wars and your reaction to the book continues that spirit. In fact, Pinker & Jackendoff are treated quite favorably on p.326. Your review does totally omit any mention of Cognitive Linguistics, which is a crucial part of the bridging story.

I just found this blog through Google alerts and will try to see if there is a way to help you understand NTL in your terms - probably not through the blog. Please do send the new paper.

Anonymous said...

I have not read the HBM paper. Generally, I would side with the junior scientists. However, I can see why Peter Fox would want this paper published in HBM, given all the noise he and his collaborators have been making for years about the "default network". Since all their data is based on hemodynamic responses, so they desperately need some electrophysiology. I would not call this a conflict of interest, but almost. On the contrary, I'm not sure why Logothetis would want to veto this paper for personal reasons. What personal reasons? Does he have a theory that goes again spontaneous activity?
It may turn out that Logothetit was entirely right, and the findings were an artifact.
Finally, a minor point, I wonder why "Leopold" was misspelled at the end of the piece (it was "Leopard").

Anonymous said...

Sorry, my previous comment belonged in the Logothetis thread! Argh!