Sunday, October 25, 2015

Brains, grammars and hype (part 1)

This recent neuro-ling paper by Frankland and Greene (F&G) in PNAS has generated a lot of critical comment by linguists, and rightly so.  The paper demonstrates several unfortunate flaws in the thinking (and moral character (I return to this)) of cog-neuro (CN) types. The most startling, intellectually speaking, is CNers deep seated dualism. It appears (if judged by their writing rather than their occasional denunciation of ghosts and souls) that CNers do not believe in the identity thesis (i.e. minds and brains are the same thing). In fact, they seem to believe that behavioral evidence, no matter how subtle or empirically well grounded or replicable or statistically significant or robust or of large effect size or…is inherently incapable of doing much to advance our understanding about how brains are organized. The only real evidence, on this view, comes from fMRI/MEG/ etc. brain studies. Thus, mentalistic investigations of brains (behavioral, cognitive, psychological) cannot possibly inform us about how brains are structured. Talk about dualism! Not even Descartes would have been caught dead saying such things. But for many CNers it’s either show them the meat (literally) or take a hike.

This is quite evident in the F&G paper, which is why it has generated such push back from linguists. Angelika Kratzer’s reaction (here) is right on the mark. Let me quote her:

One quote (attributed to Steven Frankland) in the Harvard Gazette article may point to the source of the communication problem: “This [the systematic representation of agents and themes/patients, A.K.] has been a central theoretical discussion in cognitive science for a long time, and although it has seemed like a pretty good bet that the brain works this way, there’s been little direct empirical evidence for it.” This quote makes it appear as if the idea that the human mind systematically represents agents and themes/patients has had the mere status of a bet before the distinction could be actually localized in the brain. That the distinction is systematically represented in all languages of the world is not given the status of a fact in this quote - it doesn't count as  "empirical evidence". It's like denying your pulse the status of a fact before we can localize the mechanisms that regulate it in the brain.

Yup.[1] As Angelika notes, this quote presupposes that behavioral work, in this case linguistic work, no matter how extensive, does not even rise to the level of “evidence” about brain structure. It seems that minds are one thing and brains another. Dualism anyone?

At any rate, take a look at Angelika’s piece. It makes this point well. Consequently, I thought I would try to do something uncharacteristic in what follows. Instead of zeroing in on the inanities (which, to repeat, are many, and which I will be unable to refrain from mentioning from time to time), I would like to zero in on the substantive contribution F&G makes to our understanding of the brain bases of language competence. But read this for yourself. I am no expert in these matters. However, I am channeling the wisdom of others in what follows. The UMD ling dept congregated recently to discuss the paper and I left convinced that despite the over blown rhetoric (I will return to this) and false advertising (I return to this too) the paper does make a modest contribution. And in the spirit of babies and bathwaters I will try to outline what this might be.  It goes without saying, but I will say it nonetheless, that I am no expert in these matters and I am relying on the knowledge of others here, and I might have screwed things up in translation (and if so, I hope others will chime in), but with all these caveats, here’s why I think that the paper is not just wrongheaded (though it is that too), but makes a possible contribution to our understanding of a very difficult topic. [2]

F&G is a fishing expedition of a kind that we have seen before. For example, Pallier, Devauchelle and Dehaene (discussed here) do something similar in hunting for where the Merge operation lives in the brain. F&G are hunting for brain correlates of “thematic” (yes these are scare quotes, I return to this) information; specifically, “whether and how” (p.11732) the brain codes the “who did what to whom” information that sentence’s express.

The “whether” part of the question, linguists rightly believe, has been already well-established. The most generous reading of F&G is that it agrees but notes that this still leaves open three questions: (1) can we find more neuro based indices of this well established cognitive fact (i.e. fMRI or MEG/EEG or lesion data), (2) can we localize these “thematic” brain effects and (3) what might such localization tell us about how brains code this information. IMO, the most interesting features of F&G regards the first two questions, for which it provides tentative answers. What are these?

F&G conducts two kinds of experiments. The first “identifies a broad region,” the left medial Superior Temporal Cortex (aka: lmSTC) that is able to reliably distinguish sentences that express the same theta information. For example, it can distinguish the sentence pairs John kicked Mary/Mary was kicked by John from Mary kicked John/John was kicked by Mary. By “averaging” over the active and passive pairs, the experiment zeros in on the doers and done-tos and abstracts away from surface syntax.  At the least the experiment shows that this region is not sensitive to just the words involved as these are held constant in the contrasting pairs. What matters is the “thematic” structure.

How well does lmSTC do in distinguishing these contrasts? It succeeds about 57% of the time. By ling standards this is really not enough. After all, humans succeed about 100% of the time. Thus, we need to explain why a region that fails 43% of the time to correctly distinguish what speakers never fail to distinguish (and this is the kind of thing that speakers are virtually perfect at) nonetheless is the brain basis of this overt behavioral capacity.

And there is a ready possible account: the resolution of the fMRI probe is not good enough to eliminate interfering noise and this noise is what reduces discrimination to a mere 57%. However, as the area discriminates above chance then it is a reasonable guess that it is not only sensitive to thematic distinctions, but is where these distinctions get coded and we would see this yet more clearly were we able to get an even finer probe.[3]

So, lmSTC tracks thematic information.  Before going on, I should add that F&G identifies another area that tracks this information at with roughly the same accuracy (the right posterior insula/extreme capsule region (call this region R-2)). However, the paper treats the response of this region as (at best) secondary and most likely not relevant. Why? Because whereas the lmSTC predicts further downstream brain responses, the second area does not.  The neuro crowd at our little UMD discussion got all hot about this wrinkle, so let me tell you a bit about it.

F&G shows that there is a correlation between responses to sentences in lsSTC and the amygdala, where affective responses are apparently evoked. The paper shows that the amygdala gets all hyped up to sentence pairs like The grandfather kicked the baby/the baby was kicked by the grandfather but not to The baby kicked the grandfather/the grandfather was kicked by the baby. Why the differential response? Because the amygdala doesn’t like it when babies are badly treated by wicked granddads but thinks that babies kicking old folks is not really very bad (after all how much harm could a baby kick do?). F&G interprets this, reasonably enough, as showing that the info extracted in lmSTC is used by the amygdala in responding. R-2 shows no such correlation. So whatever is going on there does not correlate with downstream amygdala responses. F&G concludes that R-2 “failed to meet additional minimal functional criteria for encoding sentence meaning” (11733). From what I can tell, the only criteria it failed to code is this downstream impact. Make of this what you will. From where I sit, it does not imply that the same thematic distinctions are not coded in R-2 as in lmSTC, only that the active use of this information pipelines directly from the latter to the amygdala but not the former. But, the CNers really liked this, so I offer it to you for your appreciation.

The second experiment zeros in on the fine structure of lmSTC. In particular, it aims to see whether and where in this already identified region doers and done-tos are coded. F&G does this, again, by seeing how the region’s discrimination powers generalize. How do the subparts of the region react to doers and done-tos as such. Here’s how F&G describes the procedure for isolating doers and done-tos as such in lmSTC :

For our principal searchlight analyses, four-way classifiers were trained to identify the agent or patient using data generated by four out of five verbs. The classifiers were then tested on data from sentences containing the withheld verb. For example, the classifiers were tested using patterns generated by “th dog chased the man,” having never previously encountered patterns generated by sentences involving “chased,” but having been trained to identify “dog” as the agent and “man” as the patient” in other verb contexts. …Thus, this analysis targets regions that instantiate consistent patterns of activity for (for example) “dog as agent” across verb contexts, discriminable from “man as agent” …A region that carries this information therefore encodes “who did it?” across nouns and verb contexts tested. (11734).

It turns out that two proximate yet distinct parts of lmSTC seem to discriminate doers from done-tos. Actually, the results for done-tos is cleaner. The borders of the doer region is muddier (moreover, active and passives of the same roles don’t function quite in parallel).[4] At any rate, F&G conclude that the lmSTC spatially bifurcates the two roles, and, at least to me (and more importantly the neuro-psycho people in the UMD discussion group), this conclusion seems reasonable given the data.

That’s what F&G shows: if correct, it identifies a role sensitive region and areas within that region differentially sensitive to the doer and done-to roles. In other words, if correct, F&G offers a hypothesis about where roles get coded.  But F&G claims to do a whole lot more. Does it? We return to this in the next post.

[1] Lest you think that Frankland indulged in hyperbole for the delectation of the newshounds alone, the same sentiment permeates the PNAS piece.
[2] Thanks particularly to Ellen Lau for organizing this and to Allyson Ettinger for a vigorous defense of the paper. I have shamelessly stolen all that I could from the excellent discussion.
[3] Note that this “guess” is quite a bit more precarious than the “bet” that who did what to whom info is represented in brains. There is no doubt that brains code for such information, and this is much more solid than the proposal that it gets coded in lmSTC. This just reiterates Angelika’s apt remarks above.
[4] F&G discusses why not, but the discussion is pretty inconclusive. If lmSTC exclusively tracks “thematic” information then this results is clearly unexpected.d


  1. Quick clarification on the region that got less love ("R-2" in your discussion). It was not discarded because it played no role in the amygdala response. It was discarded because it did a poor job in a second classification test that should have been easier than the first. The authors described it as a "sanity check". Whereas the main classification test required discrimination between mirror-image sentences, the second classification test required discrimination between unrelated sentences, such as "the truck hit the ball" vs. "the father hit the child" (cf. Supplemental materials, p. 2). The authors' reasoning was that if it fared poorly on something so easy, then the initial success was more likely to be a fluke.

    The reason why we CNers were enthusiastic about the amygdala analyses was that it seems like an ingenious additional check. It could very easily have failed.

  2. If we want to be (very?) generous to the authors of the PNAS paper, we might focus our attention on the localization bit. Here's a quote from the linked post by Angelika Kratzer, with emphasis added:

    "Here, we find evidence supporting a long-standing conjecture of cognitive science: that the human brain encodes the meanings of simple sentences much like a computer, with distinct neural populations representing answers to basic questions of meaning such as “Who did it?” and “To whom was it done?""

    Please correct me if I'm wrong, but while the behavioral evidence (strongly) supports the idea that the brain encodes thematic distinctions, it doesn't imply (or even suggest) that this encoding has to be done by distinct neural populations.

    1. Funny you should bring this up. I try to say something about this in part 2. You are right, this would be an addition (which I think their work suggests). So let me delay further discussion on my part until I post part 2.

  3. My comment is the same as Noah's. Nothing in theoretical linguistics predicts that agents and patients should activate distinct neural populations. If the brain results had NOT shown a distinction, Angelika would not think for a second that she should give up that theoretical distinction. This is simply Marr's distinction between computational theory and procedural level+hardware implementation. Different hardwares can implement the same computational theory--there is no necessity (as Thomas Graf put it to me) that neuronal mechanisms carry their computational properties on their sleeves. (Is that English?)

  4. This comment has been removed by a blog administrator.