Of Minds and Language (72 page)

I hope that being here may have been useful to you as students thinking about what you want to do. The lesson is that you better be careful not to be sitting on some silly hill asking a question that nobody wants to know the answer to.

R
IZZI
: It has been a fantastic opportunity for us to be exposed to things that are sufficiently close to being understandable and sufficiently different to offer a slightly different perspective on the things we do. Maybe it's another case of the illusion of difference. Among the various topics that were extremely fascinating and important for the work on language that I and other people do, one thing that really caught my attention is the analogy that can be seen – at least at a certain level of abstraction – between strategies used by different species in the context of species-specific capacities, particularly optimization strategies that have relatively close analogues in language. And Gallistel's talk was illuminating for me and Mark Hauser's talk was also extremely inspiring about the possible existence of optimization strategies including something close to minimal search strategies that are found in natural languages.

The question remains whether we are still at the level of loose analogies or we are at the level of operative principles, which may have a direct causal effect on different cognitive domains, on different capacities across different species. Or maybe there is an intermediate position between the two. Years ago linguists
would have said that we are definitely at the level of vague analogies; but this was partly due to an illusion generated by the technical vocabulary that we used, as well as other aspects that Noam pointed out. If you take the first version of a structured theory of universal grammar, Noam's “Conditions on trans-formations,”
³⁴
the operative principles were called things like “Specified Subject Condition,” “Tenseless Condition,” and so on. They were, in a sense, locality principles. You could have said there was something similar to them in other domains and the activities of other species, but they looked so closely keyed to language that it seemed difficult to come up with some concrete operative generalization. Then things developed significantly in linguistics within the principles and parameters framework and within the minimalist program, and the units of computation are now much more abstract. This makes it much more feasible to look at other cognitive domains and pursue the question of the level of completeness of the analogy.

It seems to me that there are still a number of questions which should be asked. In linguistics we see the relevance of notions like locality and prominence. (Prominence being expressed in terms of notions like c-command, which may be a primitive notion or, as Noam has suggested, may be a derivative notion, but still we have a notion of prominence.) The two notions seem to interact in interesting ways. For instance, locality is computed on tree structures on which prominence is defined, hence an intervener counts only if it is in a certain hierarchical environment with respect to other things. So do we find analogues to these hierarchical properties? To put it very bluntly: given that something like minimal search can be found in other species, do we find anything like c-command in, say, foraging techniques or in the kinds of capacities that have been investigated? These questions remained to be answered.

One second remark on the issue of parameters, which was raised a number of times and very interestingly so by Noam in his comments. In fact, he said something about parameters that I've never thought of before, so I guess that's one of the purposes of these meetings – to discover new things even at the very last moment. There is this basic issue of whether parameters are UG-internal or UG-external and we have heard different varieties of this story (e.g. see pages 211–219). This seems to interact with another issue, which is the locus of the parameters: where are they expressed? Is there a particular locus in our cognitive capacities where they are expressed? On the grounds of restrictiveness, I would still strongly favor the view that parameters are in some sense UG-internal; there's some specific part of UG where parameters are expressed. The main empirical argument has to do with restrictiveness. I mentioned in previous
discussions that there are certain cases that indicate that there are fewer options empirically observable than the options that we could expect if parameters were conceived of as simply a lack of specifications or “holes” within UG. The head– complement parameter is one case: you would expect all sorts of solutions different from the solutions that you actually observe in languages. If UG did not contain any statement about order you would expect that language would resolve the problem of linearization in one sentence through VO order and in the next sentence through OV order, but that's not what we observe. We find fewer options. Questions of simplicity in the sense of absence of structure, and questions of restrictiveness are in tension in some interesting cases. One of our tasks is to resolve this kind of tension. From the viewpoint of these assumptions, I must say that I would be strongly in favor of the optimal scenario for the status of parameters that Noam just mentioned among the various possibilities, which would basically amount to adopting a version of Massimo and Donata's idea that parameters represent an optimal point of equilibrium concerning the amount of specification within UG (see pages 101–102) and what Noam just said about principles being parameters with a fixed value and slightly more complicated than parameters without a fixed value. This seems to be a very interesting and promising way of addressing this question.

U
RIAGEREKA
: A couple of thoughts. For me this has been a great growing experience. I use the word “grow” and not the word “learn” because I think that there's a significant distinction that didn't get enough attention, and since Luigi just made one of the points I was going to make about parameters, I'll say little about parameters. Thinking about the two or even three types of parameters that we talked about, one possibility is that they might all be there. If this entity is complex enough, it may have enough dimensions to it that all forms of variation are there. And that might not be crazy, because languages also change. Putting aside invasions of the usual sort, which is uninteresting, they may change for interesting internal reasons of the sort David Lightfoot (1982, 1999, 2000) talked about a few years ago. So in that case there might be a possibility of drifting elements, but it's not obvious to me that you want to have the drifting part in the core part. The core part may still be really there without this drift, but you want enough messy noise to lead to internal change, though I don't know if the change would be driven by biological considerations. So the suggestion is that maybe we shouldn't eliminate one of the types of parameters in favor of another; we may need to consider all of them.

Another thought that caught my attention is Noam's case of Sylvester, the donkey turned into a rock. I also know about Sylvester because my daughters spend a lot of time talking with me about these things. I just want to mention
two things here. (I tried this with them and they're aged 4 and 6-and-a-half, so they're good subjects for this kind of thing.) Sylvester can turn into a rock and back, but when you make Sylvester turn into two rocks, they get very nervous; unless, of course, you somehow have two aspects of Sylvester, like the tale of a guy who divides into two halves and each half lives an independent life and then finally at the end they get back together. So there are interesting limitations on those transformations that recall a little footnote that Noam had in
Aspects
(1965) referring back to Bertrand Russell's idea of continuity and concepts, which nowadays would be framed in terms of manifolds. There are dimensions to those meanings – of course it has nothing to do with reality but rather with internal topologies that our minds use to prevent us from going from one to two, and so on. I think similar issues also arise going between count and mass: you can get Sylvester to turn into the wind but it's more difficult to get the wind to turn into Sylvester. My only point is that there are interesting dimensions to explore for a very different internalist project, which is where the theory of reference also ought to go.

Finally, I was fascinated by Randy Gallistel's stuff and Chris Cherniak's stuff as well. My biggest challenge for the rest of my career would be to see these two notions get unified: I keep asking myself how these two notions can get together. I'm hoping that I can keep in contact with Chris and Randy to narrow down some of the big problems of unification in terms of something that Randy has said for years: the idea that memory is basically carrying information forward in time. As far as I can see, none of the models out there really help with that. That's what we need but we need to see how models would give us that notion of memory, and my only minor contribution there would be that some of the notions that Noam talked about today (going back to
The Logical Structure of Linguistic Theory
(1955/1975) and mathematical discussions), once you put away all the stuff that is of no interest, which he correctly laid out, there may be a residue for us to think about, and this actually relates to some of the issues that Luigi Rizzi was talking about. After all, when computations get interesting, when you have intervention effects, when you need to know whether there's another one of those or not, you have various notions there that speak to different kinds of memory that also seem to be hierarchical in some sense. So I think it might be time – and maybe Noam can help us with this – to rethink those hierarchies of fifty years ago in I-language terms, in the strong generative capacity terms. It's going to be a difficult task, but if that helps us understand what other memory factors are involved internal to computations, then the task of unifying what we're trying to do with the goal of what Chris Cherniak has already done with his networks might not be fifty years ahead of us but only twenty.

P
IATTELLI-PALMARINI
: Going back to the Royaumont meeting (Piattelli-Palmarini 1980a, 1994), there are some very interesting permanent positive trends that have developed since then. At the time, Piaget's obsession was to build up more power: more and more powerful structures. In contrast, our occupation has been to narrow down things and to constrain the search space. It goes back to the momentous Goodman's paradox: how do you constrain induction?
³⁵
He had a pragmatic solution but the problem is very much with us still. How do you constrain down the search space? Bob Berwick, who never actually published this, made a calculation some years back. He idealized a child who hears a new type of sentence every second and has to guess a grammar with one hundred rules (in the traditional sense – she can get it basically right; or not quite right and then adjust the rule until she gets it right; or keep one rule and change another). He calculated that in order for the child to succeed she needed many thousands of years in an unconstrained search space for grammars. Maybe he got it wrong (though I doubt he did) and it's 1,500 years, or even “only” 150 years, but it's still monstrous, so something else must be happening. The principles and parameters idea really was a wonderful idea and it sounds strange to some who are outside our profession that we insist so much on them, even though we cannot say exactly that we know what all of them are. People like Janet Fodor and Charles Yang do this wonderful work of modeling parametric language acquisition without being sure yet how many there are or where they are, but there doesn't seem to be any other solution. You have to constrain the search space very powerfully. In an aside of our sessions, Noam has expressed the wish that the
minimax
hypothesis that Donata and I have suggested could actually be tested. Not an easy calculation to make, but it will be eminently interesting to try.

And it seems to be the case nowadays also in evolution. The problem is you can only select something that is selectable; you have to have stability, reproducibility, and a narrow space of possibles. What natural selection can select from must be something that can be selected, that has sufficient stability. And this is why the laws of form are coming back; we heard it from Chris Cherniak and there are other examples like optimal foraging, that I mentioned earlier (see page 88). Some species seem to be at optimal foraging and it's a very old problem: what do you need genes for? There are some things you don't need any genes for because it's the physics and chemistry of the situation that dictate the solution.

Going back again to Piaget, he postulated more and more powerful structures with all these complicated things that he invoked: thematization, reflective
abstraction, and so on, which a generation of psychologists have had to study. Another problem was that he was getting to abstractions, the final abstractions, but the issue is to get the initial abstractions. We know there are very basic fundamental abstractions from the very start. This is an important lesson and in linguistics we have known it for many years, and in animal behavior we heard it from Randy Gallistel: the Kantian approach. It has been beautiful to see over the years how Randy and his collaborators have had behaviorism also implode from the inside. Noam made it implode from the outside but they have seen it implode from the inside. You have to invoke very abstract structures to account for what you observe in the different species.

Giacomo Rizzolatti has discovered mirror neurons and this has been fiercely resisted over the years. Nobody could believe in mirror neurons. He couldn't get published because the model was that there were groups of neurons controlling specific muscles, and nobody could believe that one neuron is sensitive to the act of prehension, whether it's done with the right hand, the left hand, or the mouth. He has shown that that was the case, but they didn't believe it so he had to show that it wasn't just from seeing the complete action, because the mirror neuron still fires when a screen comes up and hides the completion of the action from view. But they still resisted and said it was something like grasping rather than the act of prehension. So he designed an experiment with special tweezers which only grasp if you release them (what in French restaurants they give you to eat escargots) and the neuron still fired. So, abstractions from the very, very beginning – this is a very important development and we have to continue along this line.