Tuesday, February 2, 2016

David Poeppel: The Nijmegen Lectures where "only the best are good enough"

I’m on the train leaving Nijmegen where David Poeppel just finished giving three terrific lectures (slides here) on language and the brain. The lectures were very well attended/received and generated lots of discussion. Lecture 3 was especially animated. It was on the more general topic of how to investigate the cognitive properties of the brain. For those interested in reviewing the slides of the lectures (do it!), I would recommend starting with the slides from the third as it provides a general setting for the other two talks. Though I cannot hope to them justice, let me discuss them a bit, starting here with 3.

Lecture 3 makes three important general points.

First it identifies two different kinds of cog-neuro problems (btw, David (with Dave Embick) has made these important points before (see here and here for links and discussion). The first is the Maps Problem (MP), the second the Mapping Problem (MingP). MP is what most cog-neuro (CN) of language practice today addresses. It asks a “where” question. It takes cognitively identified processes and tries to correlate them with activity in various parts of the brain (e.g. Broca’s does syntax and STS does speech). MingP is more ambitious. It aims to answer the “how” question; how do brains do cognitive computation. It does this by relating the primitives and causal processes identified in a cognitive domain with brain primitives and operations that compute the functions cognition identifies.

A small digression, as this last point seems to generate misunderstanding in neuro types. As it goes without saying, let me say it: this does not mean that neuro is the subservient handmaiden of cognition (CNLers need not bow down to linguists (especially syntacticians) though should you feel so moved, don’t let me stand in your way). Of course there should be plenty of “adjusting” of cog proposals on the basis of neuro insight. The traffic is, and must be, two-way in principle. However, at any given period of research some parts of the whole story might require more adjusting than others as the what we know in some areas might be better grounded than the what that we know in others. And right now, IMO (and I am not here speaking for David) our understanding of how the brain does many things we think important (e.g. how knowledge is represented in brains) is (ahem) somewhat unclear. More honestly, IMO, we really know next to nothing about how cognition gets neutrally realized. David’s excellent slide 3:22 on c-elegans illustrates the general gap between neural structure and understanding of behavior/cognition. In this worm, despite knowing everything there is to know about the neurons, their connectivity, and the genetics of c-elegans we still know next to nothing at all about what it does, except, as David noted, how it poops (a point made by many others, including Cristof Koch (see here)). This is an important result for it belies the claim that we understand the basics of how thought lives in brains but for the details for given all the details we still don’t know squat about how and why the worm does what it does. In short, by common agreement, there’s quite a lot we do not yet understand how to understand.  So, yes, it is a two way street and we should aim to make the traffic patterns richly interactive (hope for accidents?) but as a matter of fact at this moment in time it is unlikely that neuro stuff can really speak with much authority to cog stuff (though see below for some tantalizing possibilities).

Back to the main point. David rightly warned against confusing where with how. He pointedly emphasized that saying where something happens is not at all the same as explaining how it happens, though a plausible and useful first step in addressing the how question is finding where in the brain the how is happening. Who could disagree?

David trotted out Marr (and Aristotle and Tinbergen (slide 3:67)) to make his relevant conceptual distinctions/points richer and gave a great illustration of a successful realization of the whole shebang.[1]  I strongly recommend looking at this example for it is a magnificent accessible illustration of what success looks like in real cog-neuro life (see slides 3:24-32). The case of interest involves localizing the position of something based on auditory information. The particular protagonists are the barn owl and the rodent. Barn owls hunt at night and use sounds to find prey. Rodents are out at night and want to avoid being found. Part of the finding/hiding is to locate where the protagonist is based on the sounds it makes. Question: how do they do this?

Well, consider a Marr like decomposition of the problem each faces. The computational problem is locating position of the “other” based on auditory info. The algorithm exploits the fact that we gather such info through the ears.  Importantly, there are two of them and they sit on opposite sides of the head (they are separated). This means that sounds that are not directly ahead or behind come to each ear at different rates. Calculating the differences locates the source of the sound. There is a nice algorithm for this based on incident detectors and delay lines that an engineer called Jeffres put together in 1948 and precisely the cognitive circuits that support this algorithm were discovered in birds in 1990 by Carr and Konishi. So, we have the problem (place location based on auditory signals differentially hitting the two ears), we have the algorithm due to Jeffres and we have the brain circuits that realize this algorithm due to Carr and Konishi. So, for this problem we have it all. What more could you ask for?

Well, you might want to know if this is the only way to solve the problem and the answer is no. The barn owl implements the algorithm in one kind of circuit and, as it turns out, the rodent uses another. But they both solve the same problem, albeit in slightly different ways. Wow!!! In this case, then, we know everything we could really hope for. What is being done, what computation is being executed and which brain circuits are doing it. Need I say that this is less so in the language case?

Where are we in the latter? Well, IMO, we know quite a bit about the problem being solved. Let me elaborate a bit.

In a speech situation someone utters a sentence. The hearer’s problem is to break down the continuous wave form coming at her and extract an interpretation. The minimum required to do this is to segment the sound, identify the various phonetic parts (e.g. phonemes) use these to access the relevant lexical entries (e.g. morphemes), and assemble these morphemes to extract a meaning. We further know that doing this requires a G with various moving and interacting parts (phonetics, phonology, morphology, syntax, semantics). We know that the G will have certain properties (e.g. generate recursive hierarchies). I also think that we now know that human parsing routines extract these G features online and very quickly. This we know because of the work carried out in the last 60 years. And for particular languages we have pretty good specifications of the actual G rules that best generate the relevant mappings. So, we have a very good description of the computational level problem and a pretty good idea of the representational vocabulary required to “solve” the problem and some of the ways that these representations are deployed in real time algorithms. What we don’t know a lot about are the wetware that realize these representations or the circuits that subserve the computations of these algorithms. Though we do have some idea of where these computations are conducted (or at least places whose activity correlates with this information processing). Not bad, but no barn owl or rodent. What are the current most important obstacles to progress?

In lecture 3, David identifies two bottlenecks. First, we don’t really have plausible “parts-lists” of the relevant primitives and operations in the G or the brain domain for language. Second, we are making our inquiry harder by not pursuing radical decomposition in cog-neuro. A word about each.

David has before discussed the parts list problem under the dual heading of the granularity mismatch problem (GMP) and the ontological incommensurability problem (OIP).  He does so again. In my comments on lecture 3 after David’s lecture (see here), I noted that one of the nice features of Minimalism is that it is trying to address GMP. In particular, if it is possible to actually derive the complexity of Gs as described by GG in, say, its GB form, in terms of much simpler operations like Merge and natural concepts like Extension then we will have identified a plausible set of basic operations that it would make sense to look for neural analogues of (circuits that track merge say (as Pallier et. al. and Freiderici and her group have been trying to do)). So Minimalism is trying to get a neurally useful “parts list” of specifically linguistic primitive operations that it is reasonable to hope that (relatively transparent) parsers use in analyzing sentences in real time.

However, it is (or IMO, should be) trying to do more. The Minimalist conceit is the idea that Gs only use a small number of linguistically special operations, and that most of FL (and the Gs they produce) use cognitively off the shelf elements. What kind? Well operations like feature checking. The features language tracks may be different but the operations for tracking them are cognitively general. IMO, this also holds true of a primitive operation of putting two things together that is an essential part of Merge. At any rate, if this is right, then a nice cog-neuro payoff if Minimalism is on the right track is that it is possible to study some of these primitive operations that apply within FL in other animals. We have seen this being seriously considered for sound where it has been proposed that birds are model species for studying the human sound system (see here for discussion). Well if Merge involves the put-together operation and this operation exists in non-human animals then we can partially study Merge by looking at what they and their brains do. That’s the idea and that’s how contemporary linguistics might be useful to modern cog-neuro.

BTW, there is something amusing about this if it is true. In my experience, cog-neuro of language types hate Minimalism. You know, too abstract, not languagy enough, not contextually situated, etc. But, if what I say above is on the right track, then this is exactly what should make it so appealing, which brings me to David’s second bottleneck.

As David noted (with the hope of being provocative (and it was)), there is lots to be said for radically ignoring lots of what goes in when we actually process speech. There is nothing wrong with ignoring intonation, statistics, speaker intentions, turn-tacking behavior and much much more. In fact, science progresses by ignoring most things and trying to decompose a problem into its interacting sub-parts and then putting them back together again. This last step, even when one has the sub-parts and the operations that manipulate them is almost always extremely complicated. Interaction effects are a huge pain, even in domains where most everything is known (think turbulence). However, this is how progress is made, by ignoring most of what you “see” and exploring causal structures in non-naturalistic settings. David urges us to remember this and implement it within the cog-neuro of language. So screw context and its complexities! Focus in on what we might call the cog-neuro of the ideal speaker hearer. I could not agree more, so I won’t bore you with my enthusiasm.

I’ve focused here on lecture 3. I will post a couple of remarks on the other two soon. But do yourself a favor and take a look. They were great, the greatest thing being the reasonable, hopeful ambition they are brimming with.

[1] Bill Idsardi has long encouraged me to look at this case in detail for it perfectly embodied the Marrian ideal. He was right. It’s a great case.

No comments:

Post a Comment