Thursday, July 27, 2017
I am off for two weeks and will not post anything in that time. I will be sitting by a lake, drinking and eating to excess with friends and family. Hope you can do something similar.
The logic of GG inquiry
In the last post I was quite critical of a piece that I thought mischaracterized the nature of linguistic inquiry of the Chomsky GG variety. I thought that I should do more than hector from the sidelines (though when Hector left the sidelines things did not end well for him). Here is an attempt to outline what is not (or should not be) controversial. It tries to outline the logic of GG investigations, the questions that orient it, and the rational history that follows from pursing these questions systematically. This is not yet a piece for the uninitiated, but fleshed out, I think it could serve as a reasonably good into into what one stripe of linguists do and why. There is need for more filling (illustrations of how linguists get beyond and build on the obvious). But, this is a place to start, IMO, and if one starts here lots of misconceptions will be avoided.
Linguistics (please note the ‘i’ here) revolves around three questions:
(1) What’s a possible linguistic structure in L?
(2) What’s a possible G (for a given PLD)?
(3) What’s a possible FL (for humans)?
These three questions correspond to three facts:
(1’) The fact of linguistic creativity (a native speaker can and does regularly
produce and understand linguistic objects never before encountered by her/him)
(2’) The fact of linguistic promiscuity (any kid can acquire any language in (roughly) the same way as any other kid/language)
(3’) The fact of linguistic idiosyncrasy (humans alone have the linguistic capacities they evidently have (i.e. both (1’) and (2’) are species specific facts)
Three big facts, three big questions concerning those facts. And three conclusions:
(1’’) Part of what makes native speakers proficient in a language is their
cognitive internalization of a recursive G
(2’’) Part of human biology specifies a species wide capacity (UG) to acquire recursive Gs (on the basis of PLD)
(3’’) Humans alone have evolved the Gish capacities and meta-capacities
specified in (2’’) and (3’’) in the sense that our ancestors did not have this meta-capacity (nor do other animals) and we do
IMO, the correctness of these conclusions is morally certain (certain in the sense that though not logically required, are trivially obvious and indubitable once the facts in (1’-3’) are acknowledged. Or, to put this another way, the only way to deny the trivial truths in (1’’-3’’) is to deny the trivial facts in (1’-3’). Note, that this does not mean that these are the only questions one can ask about language, but if the questions in (1-3) are of interest to you (and nobody can force anybody to be interested in any question!), then the consequences that follow from them are sound foundations for further inquiry. When Chomsky claims that many of the controversial positions he has advanced are not really controversial, this is what he means. He means that whatever intellectual contentiousness exists regarding the claims above in no way detracts from their truistic nature. Trivial and true! Hence, intellectually uncontroversial. He is completely right about this.
So, humans have a species specific dedicated capacity to acquire recursive Gs. Is this all that we can trivially deduce from obvious facts? Nope. We can also observe that these recursive Gs have a side that we can informally call a meaning (M), and a side that we can informally can a sound (S) (or, more precisely, an articulation). So, the recursive G pairs meanings with sounds (in Chomsky’s current formulation of the old Aristotelian observation (and yes, it is very old because very trivial). And this unbounded pairing of Ms and Ss is biologically novel in humans. Does this mean that anything we can call language rests on properties unique to humans? Nope. All that follows (but it does follow trivially) is that this unbounded capacity to pair Ms and Ss is biologically species specific. So, even if being able to entertain thoughts is not biologically specific and the capacity to produce sounds (indeed many many) is not biologically unique, the capacity to pair Ms with Ss open-endedly IS. And part of the project of linguistics is to explain (i) the fine structure of the Gs we have that subvene this open-ended pairing, (ii) the UG (i.e. meta-capacity) we have that allows for the emergence of such Gs in humans and (iii) a specification of how whatever is distinctively linguistic about this meta-capacity fits in with all the other non linguistically proprietary and exclusively human cognitive and computational capacities we have to form the complex capacity we group under the encyclopedia entry ‘language.’
The first two parts of the linguistic project have been well explored over the last 60 years. We know something about the kinds of recursive procedures that particular Gs deploy and something about the possible kinds of operations/rules that natural language Gs allow. In other words, we know quite a bit about Gs and UG. Because of this in the last 25 years or so it has become fruitful to ruminate about the third question: how it all came to pass, or, equivalently, why we have the FL we have and not some other? It is a historic achievement of the discipline of linguistics that this question is ripe for investigation. It is only possible because of the success in discovering some fundamental properties of Gs and UG. In other words, the Minimalist Program is a cause for joyous celebration (cue the fireworks here). And not only is the problem ripe, there is a game plan. Chomsky has provided a plausible route towards addressing this very hard problem.
Before outlining the logic (yet again) let’s stop and appreciate what makes the quetion hard. It’s hard because it requires distinguishing two different kinds of universals; those that are cognitively and computationally general from those that are linguistically proprietary, and to do this in a principled way. And that is hard. Very very hard. For it requires thinking of what we formally called UG is an interaction effect, and hence as not a unitary kind of thing. Let me explain.
The big idea behind minimalism is that much of the “mechanics” behind our linguistic facility is not linguistically parochial. Only a small part is. In practical terms, this means that much of what we identified as “linguistic universals” from about the mid 1960s to the mid 1990s are themselves composed of operations only some of which are linguistically proprietary. In other words, just as GB proposed to treat constructions as the interaction of various kinds of more general mechanisms rather than as unitary linguistic “rules” now minimalism is asking that we thing of universals as themselves composed of various kinds of interacting computational and cognitive more primitive operations only some of which are linguistically proprietary.
In fact, the minimalist conceit is that FL is mostly comprised of computational operations that are not specific to language. Note the ‘most.’ However, this means that at least some part of FL is linguistically specific/special (remember 3/3’/3’’ above). The research problem is to separate the domain specific wheat from the domain general chaff. And that requires treating most of the “universals” heretofore discovered as complexes and showing how their properties could arise from the interaction of the general and specific operations that make them up. And that is hard both analytically and empirically.
Analytically it is hard because it requires identifying plausible candidates for the domain general and the linguistically proprietary operations. It is empirically difficult for it requires expanding how we evaluate our empirical results. An analogy with constructions and their “elimination” as grammatical primitives might make this clearer.
The appeal of constructions is that they correspond fairly directly to observable surface features of a language. Topicalizations have topics which sit on the left periphery. Topics have certain semantic properties. Topicalizations allow unbounded dependencies between the topic and a thematic position, though not if the gap is inside an island and the gap is null. Topicalization is similar to, but different from Wh-questions, which are in some ways similar to focus constructions, and in some ways not and all are in some ways similar to relative clause constructions and in some ways not. These constructions have all been described numerous times identifying more and more empirical nuances. Given the tight connection between constructions and their surface indicators, they are very handy ways of descriptively carving up the data because they provide useful visible landmarks of interest. They earn their keep empirically and philologically. Why then dump them? Why eliminate them?
Mid 1980s theory did so because they inadequately answer a fundamental question: why do constructions that are so different in so many ways nonetheless behave the same way as regards, say, movement? Ross established that different constructions behaved similarly wrt island effects, so the question arose as to why this was so. One plausible answer is that despite their surface differences, various constructions are composed from similar building blocks. More concretely, all the identified constructions involve a ‘Move Alpha’ (MA) component and MA is subject to locality conditions of the kind that result in island effects if violated. So, why do they act the same? Because they all use a common component which is grammatically subject to the relevant locality condition.
Question asked. Question answered. But not without failing to cover all the empirical ground constructions did. Thus, what about all the differences? After all, nobody thinks that Topicalization and Relativization are the same thing! Nobody. All that is claimed is that they are formed exploiting a common sub-operation and that is why they all conform to island restrictions. How are the differences handled? Inelegantly. They are “reduced” to “criterial conditions” that a head imposes on its spec or feature requirements that a probe imposes on its goal. In other words, constructions are factored into the UG relevant part (subject to a specific kind of locality) and the G idiosyncratic part (feature/criteria requirements between heads and phrases of a certain sort). In other words, constructions are “eliminated” in the sense of being grammatically basic, not in being objects of the language with the complex properties they have. Constructions, in other words, are the result of the complex interactions of more primitive Gish operations/features/principles. They are interaction effects, with all the complexity this entails.
But this factorization is not enough. One more thing is required to make deconstructing constructions into their more basic constituent parts all theoretically and empirically worthwhile. It is required that we identify some signature properties of the more abstract MA that is a fundamental part of the other constructions, and that’s what all the fuss about successive cyclicity was all about. It was interesting because it provided a signature property of the movement operation: what appears to be unbounded movement is actually composed of small steps, and we were able to track those steps. And that was/is a big deal. It vindicated the idea that we should analyze complex constructions as the interaction of more basic operations.
Let’s now return to the problem of distilling the domain general from the domain specific wrt FL. This will be hard for we must identify plausible operations of each type, show that in combination they yield comparable empirical coverage as earlier UG principles, and identify some signature properties of the domain specific operations/principles. All of this is hard to do, and hence the intellectual interest of the problem.
So what is Chomsky’s proposed route to this end? His proposal is to take recursive hierarchy as the single linguistically specific property of FL. All other features of FL are composite. The operation that embodies this property is, of course, Merge. The conceit is that the simplest (or at least one very simple) operation that embodies this property also has other signature properties we find universally in Gs (e.g. embodies both structure building and displacement, provides G format for interpretation and reconstruction effects, etc.). So identify the right distinctive operation and you get as reward an account for why Gs display some signature properties.
Does this mean that FL only contains Merge? No. If true, it means that Merge is the only linguistically distinctive operation of this cognitive component. FL has other principles and operations as well. So feature checking is a part of FL (Gs do this all the time and is the locus of G differences), though it is unlikely that feature checking is an operation proprietary to FL (even though Gs do it and FL exploits it). Minimality is likely an FL property, but one hopes that it is just a special instance of a more general property that we find in other domains of cognition (e.g. similarity based interference). So too with phases (one hopes), which function to bound the domain of computations, something that well designed systems will naturally do. Again, much of the above are promissory notes, not proposals, but hopefully you get the idea. Merge in combination with these more generic cognitive and computational operations work in concert to deliver an FL.
IMO (not widely shared I suspect), the program is doing quite well in providing a plausible story along these lines. Why do we have the FL we have? Because it is the simplest (or very simple) combination of generic computational and cognitive principles plus one very simple linguistically distinctive operation that yields a most distinctive feature of human linguistic objects, unbounded hierarchy.
Why is simple important? Because it is a crucial ingredient of the phenotypic gambit (see here). We are assuming that simple and evolvable are related. Or, more exactly, we are taking phenotypically simple as proxy for genetically simple as is typical in a lot of work on evolution.
So linguistics starts from three questions rooted in three basic facts and resulting in three kinds of research; into G, into UG and into FL. These questions build on one another (which is what good research questions in healthy sciences do). The questions get progressively harder and more abstract. And, answers to later questions prompt revisions of earlier conclusions. I would like to end this over long disquisition with some scattered remarks about this.
As noted, these projects take in one another’s wash. In particular, the results of earlier lines of inquiry are fodder for later ones. But they also change the issues. MP refines the notion of a universal, distancing it even more than its GB ancestor does from Greenbergian considerations. GB universals are quite removed from the simple observations that motivate a Greenberg style universal recall: they are largely based on negative data). However, MP universals are even some distance from classical GB universals in that MP worries the distinction between those cognitive features that are linguistically proprietary and those that are not in a way that GB seldom (never?) did. Consequently, MP universals (e.g. Merge) are even more “abstract” than their GBish predecessors, which, of course, makes them more remote from the kind of language particular data that linguists are trained to torture for insights.
Or to put this another way: MP is necessarily less philologically focused than even GB was. The focus of inquiry is explicitly the fine structure of FL. This was also true of earlier GBish theories, but, as I’ve noted before, this focus could be obscured. The philologically inclined could have their own very good reasons for “going GB,” even absent mentalist pretentions. MP’s focus on the structure of FL makes it harder (IMO, impossible) to evade a mentalist focus.
A particularly clear expression of the above is the MP view of parameters. In GBish accounts parameters are internal properties of FL that delimit the class of possible Gs. Indeed, Chomsky made a big deal of the fact that in P&P theories there were a finite number of Gs (though perhaps a large finite number) dependent on the finite number of choices for values FL allowed. This view of parameters fit well with the philologists interest in variation, for it proposed that variation was severely confined, limited to a finite number of possible differences. On this view, the study of variation feeds into a study of FL/UG by way of a study of the structure of the finite parameter space. So, investigating different languages and how they vary is, on this view, the obvious way of studying the parametric properties of FL.
But, from an MP point of view, parameters are suspect. Recall, the conceit is that the less linguistic idiosyncrasy built into FL, the better. Parameters are very very idiosyncratic (is TP or CP a bounding node? Are null subjects allowed?). So the idea of FL internal parameters is MP unwelcome. Does this deny that there is variation? No. It denies that variation is parametrically constrained. Languages differ, there is just no finite limit to how they might.
Note that this does not imply that anything goes. It is possible that no Gs allow some feature without it being the case that there is a bound on what features a G will allow. So invariances (aka: principles) are fine. It’s parameters that are suspect. Note, that on this view, the value of work on variation needs rethinking. It may tell you little about the internal structure of FL (though it might tell you a lot about the limits of the invariances).
Note further that this further drives a wedge between standard linguistic research (so much is dedicated to variation and typology) and the central focus of MP research, the structure of FL. In contrast to P&P theories where typology and variation are obviously relevant for the study of FL, this is less obvious (I would go further) in an MP setting. I tend to think that this fact influences how people understand the virtues and achievements of MP, but as I’ve made this point before, I will leave it be here.
Last, I think that the MP problematic encourages a healthy disdain for surface appearances, even more so than prior GBish work. Here’s what I mean: if your interest is in simplifying FL and relating the distinctive features of language to Merge then you will be happy downplaying surface morphological differences. So, for example, if MP leads you to consider a Merge based account of binding, then reflexive forms (e.g. ‘himself’) are just the morphological residues of I-merge. Do they have interesting syntactic properties? Quite possibly not. They are just surface detritus. Needless to say, this way of describing things can be seen, from another perspective, as anti-empirical (believe me, I know whereof I write). But if we really think that all that is G distinctive leads back to Merge then if you think that c-command is a distinctive product of Merge and you find this in binding then you will want to unify I-merge and binding theory so as to account for the fact that binding requires c-command. But this will then mean ignoring many differences between movement and binding, and one way to do this is to attribute the differences to idiosyncratic “morphology” (as we did in eliminating constructions). In other words, from an MP perspective there are reasons to ignore some of the data that linguists hold so dear.
There is a line (even Chomsky has pushed it) that MP offers nothing new. It is just the continuation of what we have always done in GG. There is one sense in which I think that this is right. The questions asked linguistics have investigated follow a natural progression if one’s interest is in the structure of FL. MP focuses on the next natural question to ask given the prior successes of GG. However, the question itself is novel, or at least it is approachable now in ways that it wasn’t before. This has consequences. I believe that one of the reasons behind a palpable hostility to MP (even among syntacticians) is the appreciation that it does change the shape of the board. Much of what we have taken for granted is rightly under discussion. It is like the shift away from constructions, but in an even more fundamental way.
 I discuss this again in a forthcoming post. I know you cannot wait.
 In other words, this argument form is not particularly novel when applied to language. As such one should beware to avoid methodological dualism and not subject the linguistic application of this gambit to higher standards than generally apply.
 A personal judgment: I don’t believe that cross-linguistic study has generally changed our views about the principles. But this is very much a personal view, I suspect.
Thursday, July 20, 2017
Is linguistics a science?
I have a confession to make: I read (and even monetarily support) Aeon. I know that they publish junk (e.g. Evans has dumped junk on its pages twice), but I think the idea of trying to popularize the recondite for the neophyte is a worthwhile endeavor, even if it occasionally goes awry. I mention this because Aeon has done it again. The editors clearly understand the value (measured in eyeballs) of a discussion of Chomsky. And I was expecting the worst, another Evans like or Everett like or Wolfe like effort. In other words I was looking forward to extreme irritation. To my delight, I was disappointed. The piece (by Arika Okrent here) got many things right. That said, it is not a good discussion and will leave many more confused and misinformed than they should be. In what follows I will try to outline my personal listing of pros and cons. I hope to be brief, but I might fail.
The title of Okrent’s piece is the title of this post. The question at issue is whether Chomskyan linguistics is scientific. Other brands get mentioned in passing, but the piece Is linguistics a science? (ILAS), is clearly about the Chomsky view of GG (CGG). The subtitle sets (part of) the tone:
Much of linguistic theory is so abstract and dependent on theoretical apparatus that it might be impossible to explain
ILAS goes into how CGG is “so abstract” and raises the possibility that this level of abstraction “might” (hmm, weasel word warning!) make it incomprehensible to the non-initiated, but it sadly fails to explain how this distinguishes CGG from virtually any other inquiry of substance. And by this I mean not merely other “sciences” but even biblical criticism, anthropology, cliometrics, economics etc. Any domain that is intensively studied will create technical, theoretical and verbal barriers to entry by the unprepared. One of the jobs of popularization is to allow non-experts to see through this surface dazzle to the core ideas and results. Much as I admire the progress that CGG has made over the last 60 years, I really doubt that its abstractions are that hard to understand if patiently explained. I speak from experience here. I do this regularly, and it’s really not that hard. So, contrary to ILAS, I am quite sure that CGG can be explained to the interested layperson and the vapor of obscurity that this whiff of ineffability spritzes into the discussion is a major disservice. (Preview of things to come: in my next post I will try (again) to lay out the basic logic of the CGG program in a way accessible (I hope) to a Sci Am reader).
Actually, many parts of ILAS are much worse than this and will not help in the important task of educating the non-professional. Here are some not so random examples of what I mean: ILAS claims that CGG is a “challenge to the scientific method itself” (2), suggests that it is “unfalsifiable” Popper-wise (2), that it eschews “predictions” (3), that it exploits a kind of data that is “unusual for a science” (5), suggests that it is fundamentally unempirical in that “Universal grammar is not a hypothesis to be tested, but a foundational assumption” (6), bemoans that many CGG claims are “maddeningly circular or at the very least extremely confusing” (6), complains that CGG “grew ever more technically complex,” with ever more “levels and stipulations,” and ever more “theoretical machinery” (7), asserts that MP, CGG’s latest theoretical turn confuses “even linguists” (including Okrent!) (7), may be more philosophy than science (7), moots the possibility that “a major part of it is unfalsifiable” and “elusive” and “so abstract and dependent on theoretical apparatus that it might be impossible to explain” (7), moots that possibility that CGG is post truth in that there is nothing (not much?) “at stake in determining which way of looking at things is the right one” (8), and ends with a parallel between Christian faith and CGG which are described as “not designed for falsification” (9). These claims, spread as they are throughout ILAS, leave the impression that CGG is some kind of weird semi mystical view (part philosophy, part religion, part science), which is justifiably confusing to the amateur and professional alike. Don’t get me wrong: ILAS can appreciate why some might find this obscure hunt for the unempirical abstract worth pursuing, but the “impulse” is clearly more Aquarian (as in age of) than scientific. Here’s ILAS (8):
I must admit, there have been times when, upon going through some highly technical, abstract analysis of why some surface phenomena in two very different languages can be captured by a single structural principle, I get a fuzzy, shimmering glimpse in my peripheral vision of a deeper truth about language. Really, it’s not even a glimpse, but a ghost of a leading edge of something that might come into view but could just as easily not be there at all. I feel it, but I feel no impulse to pursue it. I can understand, though, why there are people who do feel that impulse.
Did I say “semi mystical,” change that to pure Saint Teresa of Avila. So there is a lot to dislike here.
That said, ILAS also makes some decent points and in this it rises way above the shoddiness of Evans, Everett and Wolfe. It correctly notes that science is “a messy business” and relies on abstraction to civilize its inquiries (1), it notes that “the human capacity for language,” not “the nature of language,” is the focus of CGG inquiry (5), it notes the CGG focus on linguistic creativity and the G knowledge it implicates (4), it observes the importance of negative data (“intentional violations and bad examples”) to plumbing the structure of the human capacity (5), it endorses a ling vs lang distinction within linguistics (“There are many linguists who look at language use in the real world … without making any commitment to whether or not the descriptions are part of an innate universal grammar”) (6), it distinguishes Chomsky’s conception of UG from a Greenberg version (sans naming the distinction in this way) and notes that the term ‘universal grammar’ can be confusing to many (6):
The phrase ‘universal grammar’ gives the impression that it’s going to be a list of features common to all languages, statements such as ‘all languages have nouns’ or ‘all languages mark verbs for tense’. But there are very few features shared by all known languages, possibly none. The word ‘universal’ is misleading here too. It seems like it should mean ‘found in all languages’ but in this case it means something like ‘found in all humans’ (because otherwise they would not be able to learn language as they do.)
And it also notes the virtues of abstraction (7).
Despite these virtues (and I really like that above explanation of ‘universal grammar’), ILAS largely obfuscates the issues at hand and gravely misrepresents CGG. There are several problems.
First, as noted, a central trope of ILAS is that CGG represents a “challenge to the scientific method itself” (2). In fact one problem ILAS sees with discussions of the Everett/Chomsky “debate” (yes, scare quotes) is that it obscures this more fundamental fact. How is it a challenge? Well, it is un-Popperian in that it insulates its core tenets (universal grammar) from falsifiability (3).
There are two big problems with this description. First, so far as I can see, there is nothing that ILAS says about CGG that could not be said about the uncontroversial sciences (e.g. physics). They too are not Popper falsifiable, as has been noted in the philo of science literature for well over 50 years now. Nobody who has looked at the Scientific Method thinks that falsifiability accurately describes scientific practice. In fact, few think that either Falsificationism or the idea that science has a method are coherent positions. Lakatos has made this point endlessly, Feyerabend more amusingly. And so has virtually every other philosopher of science (Laudan, Cartwright, Hacking to name three more). Adopting the Chomsky maxim that if a methodological dictum fails to apply to physics then it is not reasonable to hold linguistics to its standard, we can conclude that ILAS’s observation that certain CGG tenets are falsifiable (even if this is so) is not a problem peculiar to CGG. ILAS’s suggestion that it is is thus unfortunate.
Second, as Lakatos in particular has noted (but Quine also made his reputation on this, stealing the Duhem thesis), central cores of scientific programs are never easily directly empirically testable. Many linking hypotheses are required which can usually be adjusted to fend off recalcitrant data. This is no less true in physics than in linguistics. So, having cores that are very hard to test directly is not unique to CGG.
Lastly, being hard to test and being unempirical are not quite the same thing. Here’s what I mean. Take the claim that humans have a species specific dedicated capacity to acquire natural languages. This claim rests on trivial observations (e.g. we humans learn French, dogs (smart as they are) don’t!). That this involves Gs in some way is trivially attested by the fact of linguistic creativity (the capacity to use and understand novel sentences). That it is a species capacity is obvious to any parent of any child. These are empirical truisms and so well grounded in fact that disputing their accuracy is silly. The question is not (and never has been) whether humans have these capacities, but what the fine structure of these capacities is. In this sense, CGG is not a theory, anymore than MP is. It is a project resting on trivially true facts. Of course, any specification of the capacity commits empirical and theoretical hostages and linguists have developed methods and arguments and data to test them. But we don’t “test” whether FL/UG exists because it is trivially obvious that it does. Of course, humans are built for language like ants are built to dead reckon or birds are built to fly or fish to swim. So the problem is not that this assumption is insulated from test and thus holding it is unempirical and unscientific. Rather this assumption is not tested for the same reason that we don’t test the proposition that the Atlantic Ocean exists. You’d be foolish to waste your time. So, CGG is a project, as Chomsky is noted as saying, and the project has been successful as it has delivered various theories concerning how the truism could be true, and these are tested every day, in exactly the kinds of ways that other sciences test their claims. So, contrary to ILAS, there is nothing novel in linguistic methodology. Period. The questions being asked are (somewhat) novel, but the methods of investigation are pure white bread. That ILAS suggests otherwise is both incorrect and a deep disservice.
Another central feature of ILAS is the idea that CGG has been getting progressively more abstract, removed from facts, technical, and stipulative. This is a version of the common theme that CGG is always changing and getting more abstruse. Is ILAS pining for the simple days of LSLT and Syntactic Structures? Has Okrent read these (I actually doubt it given that nobody under a certain age looks at these anymore). At any rate, again, in this regard CGG is not different from any other program of inquiry. Yes, complexity flourishes for the simple reason that more complex issues are addressed. That’s what happens when there is progress. However, ILAS suggests that contemporary complexity contrasts with the simplicity of an earlier golden age, and this is incorrect. Again, let me explain.
One of the hallmarks of successful inquiry is that it builds on insights that came before. This is especially true in the sciences where later work (e.g. Einstein) builds on early work (e.g. Newton). A mark of this is that newer theories are expected to cover (more or less) the same territory as previous ones. One way of doing this for newbies to have the oldsters as limit cases (e.g. you get Newton from Einstein when speed of light is on the low side). This is what makes scientific inquiry progressive (shoulders and giants and all that). Well linguistics has this too (see here for first of several posts illustrating this with a Whig History). Once one removes the technicalia (important stuff btw), common themes emerge that have been conserved through virtually every version of CGG accounts (constituency, hierarchy, locality, non-local dependency, displacement) in virtually the same way. So, contrary to the impression ILAS provides, CGG is not an ever more complex blooming buzzing mass of obscurities. Or at least not more so than any other progressive inquiry. There are technical changes galore as bounds of empirical inquiry expand and earlier results are preserved largely intact in subsequent theory. The suggestion that there is something particularly odd of the way that this happens in CGG is just incorrect. And again, suggesting as much is a real disservice and an obfuscation.
Let me end with one more point, one where I kinda like what ILAS says, but not quite. It is hard to tell whether ILAS likes abstraction or doesn’t. Does it obscure or clarify? Does it make empirical contact harder or easier? I am not sure what ILAS concludes, but the problem of abstraction seems contentious in the piece. It should not be. Let me end on that theme.
First, abstraction is required to get any inquiry off the ground. Data is never unvarnished. But more importantly, only by abstracting away from irrelevancies can phenomena be identified at all. ILAS notes this in discussing friction and gravitational attraction. It’s true in linguistics too. Everyone recognizes performance errors, most recognize that it is legit to abstract away from memory limitations in studying the G aspects of linguistic creativity. At any rate, we all do it, and not just in linguistics. What is less appreciated I believe is that abstraction allows one to hone one’s questions and make it possible to make contact with empirics. It was when we moved away from sentences uttered to judgments about well formedness investigated via differential acceptability that we were able to start finding interesting Gish properties of native speakers. Looking at utterances in all their gory detail, obscures what is going on. Just as with friction and gravity. Abstraction does not make it harder to find out what is going on, but easier.
A more contemporary example of this in linguistics is the focus on Merge. This abstracts away from a whole lot of stuff. But, it also by ignoring many other features of G rules (besides the capacity to endlessly embed) allows for inquiry to focus on key features of G operations: they spawn endlessly many hierarchically organized structures that allow for displacement, reconstruction, etc. It also allows one to raise in simplified form new possibilities (do Gs allow for SW movement? Is inverse control/binding possible?). Abstraction need not make things more obscure. Abstracting away from irrelevancies is required to gain insight. It should be prized. ILAS fails to appreciate how CGG has progressed, in part, by honing sharper questions by abstracting away from side issues. One would hope a popularization might do this. ILAS did not. It made appreciating abstractions virtues harder to discern.
One more point: it has been suggested to me that many of the flaws I noted in ILAS were part of what made the piece publishable. In other words, it’s the price of getting accepted. This might be so. I really don’t know. But, it is also irrelevant. If this is the price, then there are worse things than not getting published. This is especially so for popular science pieces. The goal should be to faithfully reflect the main insights of what one is writing about. The art is figuring out how to simplify without undue distortion. ILAS does not meet this standard, I believe.
 The CGG as mysticism meme goes back a long way. I believe that Hockett’s review of Chomsky’s earliest work made similar suggestions.
 In fact, few nowadays are able to identify a scientific method. Yes, there are rules of thumb like think clearly, try hard, use data etc. But the days of thinking that there is a method, even in the developed sciences, is gone.
Monday, July 17, 2017
The Gallsitel-King conjecture; another brick in the wall
Several people sent me this piece discussing some recent work showing how to store and retrieve information in "live" (vs synthetic) DNA. It's pretty cool. Recall the Gallistel-King conjecture (GKC) is that a locus of cognitive computing will be intra-cellular and that large molecules like DNA will be the repository of memories. The advantage is that we know how to"write to" and "read from" such chemical computers and that this is what we need if we are to biologically model the kinds of computations that behavioral studies have shown to be what is going on in animal cognition. The proof of concept that this is realistic invites being able to do this in "live" systems. This report shows that it has been done.
Now we need to find more plausible mechanisms by which this kind of process might take place. But, this is a cool first step and makes the GKC a little less conjectural.
The images and videos the researchers pasted inside E. Coli are composed of black-and-white pixels. First, the scientists encoded the pixels into DNA. Then, they put their DNA into the E. coli cells using electricity. Running an electrical current across cells opens small channels in the cell wall, and then the DNA can flow inside. From here, the E. Coli’s CRISPR system grabbed the DNA and incorporated it into its own genome. “We found that if we made the sequences we supplied look like what the system usually grabs from viruses, it would take what we give,” Shipman says.
Once the information was inside, the next step was to retrieve it. So, the team sequenced the E. coli DNA and ran the sequence through a computer program, which successfully reproduced the original images. So the running horse you see at the top of the page is really just the computer's representation of the sequenced DNA, since we can’t see DNA with the naked eye.
Now we need to find more plausible mechanisms by which this kind of process might take place. But, this is a cool first step and makes the GKC a little less conjectural.
Thursday, July 13, 2017
Some recent thoughts on AI
Kleanthes sent me this link to a recent lecture by Gary Marcus (GM) on the status of current AI research. It is a somewhat jaundiced review concluding that, once again, the results have been strongly oversold. This should not be surprising. The rewards to those that deliver strong AI (“the kind of AI that would be as smart as, say a Star Trek computer” (3)) will be without limit, both tangibly (lots and lots of money) and spiritually (lots and lots of fame, immortal kinda fame). And given hyperbole never cripples its purveyors (“AI boys will be AI boys” (and yes, they are all boys)), it is no surprise that, as GM notes, we have been 20 years out from solving strong AI for the last 65 years or so. This is a bit like the many economists who predicted 15 of the last 6 recessions but worse. Why worse? Because there have been 6 recessions but there has been pitifully small progress on strong AI, at least if GM is to be believed (and I think he is).
Why despite the hype (necessary to drain dollars from “smart” VC money) has this problem been so tough to crack? GM mentions a few reasons.
First, we really have no idea how open ended competence works. Let me put this backwards. As GM notes, AI has been successful precisely in “predefined domains” (6). In other words, where we can limit the set of objects being considered for identification or the topics up for discussion or the hypotheses to be tested we can get things to run relatively smoothly. This has been true since Winograd and his block worlds. Constrain the domain and all goes okishly. Open the domain up so that intelligence can wander across topics freely and all hell breaks loose. The problem of AI has always been scaling up, and it is still a problem. Why? Because we have no idea how intelligence manages to (i) identify relevant information for any given domain and (ii) use that information in relevant ways for that domain. In other words, how we in general figure out what counts and how we figure out how much it counts once we have figured it out is a complete and utter mystery. And I mean ‘mystery’ in the sense that Chomsky has identified (i.e. as opposed to ‘problem’).
Nor is this a problem limited to AI. As FoL has discussed before, linguistic creativity has two sides. The part that has to do with specifying the kind of unbounded hierarchical recursion we find in human Gs has been shown to be tractable. Linguists have been able to say interesting things about the kinds of Gs we find in human natural languages and the kinds of UG principles that FL plausibly contains. One of the glories (IMO, the glory) of modern GG lies in its having turned once mysterious questions into scientific problems. We may not have solved all the problems of linguistic structure but we have managed to render them scientifically tractable.
This is in stark contrast to the other side linguistic creativity: the fact that humans are able to use their linguistic competence in so many different ways for thought and self-expression. This is what the Cartesians found so remarkable (see here for some discussion) and that we have not made an iota of progress understanding. As Chomsky put it in Language & Mind (and is still a fair summary of where we stand today):
Honesty forces us to admit that we are as far today as Descartes was three centuries ago from understanding just what enables a human to speak in a way that is innovative, free from stimulus control, and also appropriate and coherent. (12-13)
All-things-considered judgments, those that we deploy effortlessly in every day conversation, elude insight. That we do this is apparent. But how we do this remains mysterious. This is the nut that strong AI needs to crack given its ambitions. To date, the record of failure speaks for itself and there is no reason to think that more modern methods will help out much.
It is precisely this roadblock that limiting the domain of interest removes. Bound the domain and the problem of open-endedness disappears.
This should sound familiar. It is the message in Fodor’s Modularity of Mind. Fodor observes that modularity makes for tractability. When we move away from modular systems, we flat on our faces precisely because we have no idea how minds identify what is relevant in any given situation and how it weights what is relevant in a given situation and how it then deploys this information appropriately. We do it all right. We just don’t know how.
The modern hype supposes that we can get around this problem with big data. GM has a few choice remarks about this. Here’s how he sees things (my emphasis):
I opened this talk with a prediction from Andrew Ng: “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future.” So, here’s my version of it, which I think is more honest and definitely less pithy: If a typical person can do a mental task with less than one second of thought and we can gather an enormous amount of directly relevant data, we have a fighting chance, so long as the test data aren’t too terribly different from the training data and the domain doesn’t change too much over time. Unfortunately, for real-world problems, that’s rarely the case. (8)
So, if we massage the data so that we get that which is “directly relevant” and we test our inductive learner on data that is not “too terribly different” and we make sure that the “domain doesn’t change much” then big data will deliver “statistical approximations” (5). However, “statistics is not the same thing as knowledge” (9). Big data can give us better and better “correlations” if fed with “large amounts of [relevant!, NH] statistical data”. However, even when these correlational models work, “we don’t necessarily understand what’s underlying them” (9).
And one more thing: when things work it’s because the domain is well behaved. Here’s GM on AlphaGo (my emphasis):
Lately, AlphaGo is probably the most impressive demonstration of AI. It’s the AI program that plays the board game Go, and extremely well, but it works because the rules never change, you can gather an infinite amount of data, and you just play it over and over again. It’s not open-ended. You don’t have to worry about the world changing. But when you move things into the real world, say driving a vehicle where there’s always a new situation, these techniques just don’t work as well. (7)
So, if the rules don’t change, you have unbounded data and time to massage it and the relevant world doesn’t change, then we can get something that approximately fits what we observe. But fitting is not explaining and the world required for even this much “success” is not the world we live in, the world in which our cognitive powers are exercised. So what does AI’s being able to do this in artificial worlds tell us about what we do in ours? Absolutely nothing.
Moreover, as GM notes, the problems of interest to human cognition have exactly the opposite profile. In Big Data scenarios we have boundless data, endless trials with huge numbers of failures (corrections). The problems we are interested in are characterized by having a small amount of data and a very small amount of error. What will Big Data techniques tell us about problems with the latter profile? The obvious answer is “not very much” and the obvious answer, to date, has proven to be quite adequate.
Again, this should sound familiar. We do not know how to model the everyday creativity that goes into common judgments that humans routinely make and that directly affects how we navigate our open-ended world. Where we cannot successfully idealize to a modular system (one that is relatively informationally encapsulated) we are at sea. And no amount of big data or stats will help.
What GM says has been said repeatedly over the last 65 years. AI hype will always be with us. The problem is that it must crack a long lived mystery to get anywhere. It must crack the problem of judgment and try to “mechanize” it. Descartes doubted that we would be able to do this (indeed this was his main argument for a second substance). The problem with so much work in AI is not that it has failed to crack this problem, but that it fails to see that it is a problem at all. What GM observes is that, in this regard, nothing has really changed and I predict that we will be in more or less the same place in 20 years.
Since penning(?) the above I ran across a review of a book on machine intelligence by Gary Kasparov (here). The review is interesting (I have not read the book) and is a nice companion to the Marcus remarks. I particularly liked the history on Shannon’s early thoughts on chess playing computers and his distinction on how the problem could be solved:
At the dawn of the computer age, in 1950, the influential Bell Labs engineer Claude Shannon published a paper in Philosophical Magazine called “Programming a Computer for Playing Chess.” The creation of a “tolerably good” computerized chess player, he argued, was not only possible but would also have metaphysical consequences. It would force the human race “either to admit the possibility of a mechanized thinking or to further restrict [its] concept of ‘thinking.’” He went on to offer an insight that would prove essential both to the development of chess software and to the pursuit of artificial intelligence in general. A chess program, he wrote, would need to incorporate a search function able to identify possible moves and rank them according to how they influenced the course of the game. He laid out two very different approaches to programming the function. “Type A” would rely on brute force, calculating the relative value of all possible moves as far ahead in the game as the speed of the computer allowed. “Type B” would use intelligence rather than raw power, imbuing the computer with an understanding of the game that would allow it to focus on a small number of attractive moves while ignoring the rest. In essence, a Type B computer would demonstrate the intuition of an experienced human player.
As the review goes on to note, Shannon’s mistake was to think that Type A computers were not going to materialize. They did, with the result that the promise of AI (that it would tell us something about intelligence) fizzled as the “artificial” way that machines became “intelligent” simply abstracted away from intelligence. Or, to put it as Kasparov is quoted as putting it: “Deep Blue [the machine that beat Kasparov, NH] was intelligent the way your programmable alarm clock is intelligent.”
So, the hope that AI would illuminate human cognition rested on the belief that technology and brute calculation would not be able to substitute for “intelligence.” This proved wrong, with machine learning being the latest twist in the same saga, per the review and Kasparov.
All this fits with GM’s remarks above. What both do not emphasize enough, IMO, is something that many did not anticipate; namely that we would revamp our views of intelligence rather than question whether our programs had it. Part of the resurgence of Empiricism is tied to the rise of the technologically successful machine. The hope was that trying to get limited machines to act like we do might tell us something about how we do things. The limitations of the machine would require intelligent design to get it to work thereby possibly illuminating our kind of intelligence. What happened is that getting computationally miraculous machines to do things in ways that we had earlier recognized as dumb and brute force (and so telling us nothing at all) has transformed into the hypothesis that there is no such things as real intelligence at all and everything is “really” just brute force. Thus, the brain is just a data cruncher, just like Deep Blue is. And this shift in attitude is supported by an Empiricist conception of mind and explanation. There is no structure to the mind beyond the capacity to mine the inputs for surfacy generalizations. There is no structure to the world beyond statistical regularities. On this Eish viw, AI has not failed, rather the right conclusion is that there is less to thinking than we thought. This invigorated Empiricism is quite wrong. But it will have staying power. Nobody should underestimate the power that a successful (money making) tech device can have on the intellectual spirit of the age.
Thursday, July 6, 2017
The logic of adaptation
I recently ran across a nice paper on the logic of adaptive stories (here), along with a nice short discussion of its main points (here) (by Massimo Pigliucci (P)). The Olson and Arroyo-Santos paper (OAS) argues that circularity (or “loopiness”) is characteristic of all adaptive explanations (indeed, of all non-deductive accounts) but that some forms of loopiness are virtuous while others are vicious. The goal, then, is to identify the good circular arguments from the bad ones, and this amounts to distinguishing small uninteresting circles from big fat wide ones. Good adaptive explanations distinguish themselves from just-so stories in having independent data afforded from the three principle kinds of arguments evolutionary biologists deploy. OAS adumbrates the forms of these arguments and uses this inventory to contrast lousy adaptive accounts from compelling ones. Of particular interest to me (and I hope FoLers) is the OAS claim that looking at things in terms of how fat a circular/loopy account is will make it easy to see why some kinds of adaptive stories are particularly susceptible to just-soism. What kinds? Well ones like those applied to the evolutions of language, as it turns out. Put another way, OAS leads to Lewontin like conclusions (see here) from a slightly different starting point.
An example of a just-so story helps to illustrate the logic of adaptation that OAS highlights. Why do giraffes have long necks? So as to be able to eat leaves from tall trees. Note, that giraffes eat from tall trees confirms that having long necks is handy for this activity, and the utility of being able to eat from tall trees would make having a long neck advantageous. This is the loopiness/circularity that OAS insists is part of any adaptational account. OAS further insists that this circularity is not in itself a problem. The problem is that in the just-so case the circle is very small, so small as to almost shrink to a point. Why? Because the evidence for the adaptation and the fact that the adaptation explains is the same: tall necks are what we want to explain and also constitute the evidence for the explanation. As OAS puts it:
…the presence of a given trait in current organisms is used as the sole evidence to infer heritable variation in the trait in an ancestral population and a selective regime that favored some variants over others. This unobserved selective scenario explains the presence of the observed trait, and the only evidence for the selective scenario is trait presence (168).
In other words, though ‘p implies p’ is unimpeachably true, it is not interestingly so. To get some explanation out of an account that uses these observations we need a broader circle. We need a big fat circle/loop, not an anorexic one.
OAS’s main take home message is that fattening circles/loops is both eminently doable (in some cases at least) and is regularly done. OAS lists three main kinds of arguments that biologists use to fatten up an adaptation account: comparative arguments, population arguments, and optimality arguments. Each brings something useful to the table. Each has some shortcomings. Here’s how OAS describes the comparative method (169):
The comparative method detects adaptation through convergence (Losos 2011). A basic version of comparative studies, perhaps the one underpinning most state- ments about adaptation, is the qualitative observation of similar organismal features in similar selective contexts.
A second kind of argument focuses on variations in a single population and sees how this affects “heritability and fitness between potentially competing individuals” (171). These kinds of studies involve looking at extant populations and seeing how their variations tie up with heritability. Again OAS provides an extensive example involving “the curvature of floral nectar spurs” in some flowers (171) and shows how variation and fitness can be precisely measured in such circumstances (i.e. where it is possible to do studies of “very geographically and restricted sets of organisms under often unusual circumstances” (172)).
This method, too, has a problem. The biggest drawback is that the population method “examines relatively minor characters that have not gone to fixation” and “extrapolation of results to multiple species and large time scales” is debatable (171, table 1). In other words, it is not that clear whether the situation in which population arguments can be fully deployed reveal the mechanisms that are at play “in generating the patterns of trait distribution observed over geological time and clades” because it is unclear whether the “very local population phenomena are…isomporphic with the factors shaping life on earth at large” (172).
The third type of argument involves optimality thinking. This aims to provide an outline of the causal mechanisms “behind a given variant being favored” and rests on a specification of the relevant laws driving the observed effect (e.g. principles of hydronamics for body contour/sleekness in aquatic animals). The downside to this mode of reasoning is that it is not always clear what variables are relevant for optimization.
OAS notes that adaptive explanations are best when one can provide all three kinds of reasons (as one can in the case, for example, of aquatic contour and sleekness (see figure 4 and the discussion in P). Accounts achieve just-so status when none of the three methods can apply and none have been used to generate relevant data. The OAS discussion of these points is very accessible and valuable and I urge you take a look.
The OAS framing also carries an important moral, one that both OAS and P note: if going from just-so to serious requires fattening with comparative, population and optimization arguments then some fashionable domains of evolutionary speculation relying on adaptive consideration are likely to be very just-soish. Under what circumstances will getting beyond hand waving prove challenging? Here’s OAS (184, my emphasis):
Maximally supported adaptationist explanations require evidence from comparative, populational, and optimality approaches. This requirement highlights from the outset which adaptationist studies are likely to have fewer layers of direct evidence available. Studies of single species or unique structures are important examples. Such traits cannot be studied using comparative approaches, because the putatively adaptive states are unique (cf. Maddison and FitzJohn 2015). When the traits are fixed within populations, the typical tools of populational studies are unavailable. In humans, experimental methods such as surgical intervention or selective breeding are unethical (Ruse 1979). As a result, many aspects of humans continue to be debated, such as the female orgasm, human language, or rape (Travis 2003; Lloyd 2005; Nielsen 2009; MacColl 2011). To the extent that less information is available, in many cases it will continue to be hard to distinguish between different alternative explanations to decide which is the likeliest (Forber 2009).
Let’s apply these OAS observations to a favorite of FoLers, the capacity for human language. First, human language capacity is, so far as we can tell, unique to humans. And it involves at least one feature (e.g. hierarchical recursion) that, so far as we can tell, emerges nowhere else in biological cognition. Hence, this capacity cannot be studied using comparative methods. Second, it cannot be studied using population methods, as, modulo pathology, the trait appears (at least at the gross level) fixed and uniform in the human species (any kid can learn any language in more or less the same way). Experimental methods, which could in principle be used (for there probably is some variation across individuals in phenomena that might bear on the structure of the fixed capacity (e.g. differences in language proficiency and acquisition across individuals) will, if pursued, rightly land you in jail or at the World Court in the Hague. Last, optimization methods also appear useless for it is not clear what function language is optimized for and so the dimensions along which it might be optimized are very obscure. The obvious ones relating to efficient information transmission are too fluffy to be serious.
P makes effectively the same point, but for evo-psych in general, not just evo-lang. In this he reiterates Lewontin’s earlier conclusions. Here is P:
If you ponder the above for a minute you will realize why this shift from vicious circularity to virtuous loopiness is particularly hard to come by in the case of our species, and therefore why evolutionary psychology is, in my book, a quasi-science. Most human behaviors of interest to evolutionary psychologists do not leave fossil records (i); we can estimate their heritability (ii) in only what is called the “broad” sense, but the “narrow” one would be better (see here); while it is possible to link human behaviors with fitness in a modern environment (iii), the point is often made that our ancestral environment, both physical and especially social, was radically different from the current one (which is not the case for giraffes and lots of other organisms); therefore to make inferences about adaptation (iv) is to, say the least, problematic. Evopsych has a tendency to get stuck near the vicious circularity end of Olson and Arroyo-Santos’ continuum.
There is more, much more, in the OAS paper and P's remarks are also very helpful. So those interested in evolang should take a look. The conclusion both pieces draw regarding the likely triviality/just-soness of such speculations is a timely re-re-re-reminder of Lewontin and the French academy’s earlier prescient warnings. Some questions, no matter how interesting, are likely to be beyond our power to interestingly investigate given the tools at hand.
One last point, added to annoy many of you. Chomsky’s speculations, IMO, have been suitably modest in this regard. He is not giving an evolang account so much as noting that if there is to be one then some features will not be adaptively explicable. The one that Chomsky points to is hierarchical recursion. Given the OAS discussion it should be clear that Chomsky is right in thinking that this will not be a feature liable to an adaptive explanation. What would “variation” wrt Merge be? Somewhat recursive/hierarchical? What would this be and how would the existence of 1-merge and 2-merge systems get you to unbounded Merge? It won’t, which is Chomsky’s (and Dawkins’) point (see here for discussion and references). So, there will be no variation and no other animals have it and it doesn’t optimize anything. So there will be no available adaptive account. And that is Chomsky’s point! The emergence of FL whenever it occurred was not selected for. Its emergence must be traced to other non adaptive factors. This conclusion, so far as I can tell, fits perfectly with OAS’s excellent discussion. What Chomsky delivers is all the non-trivial evolang we are likely to get our hands on given current methods, and this is just what OAS, P and Lewontin should lead us to expect.
 Note that Chomsky’s conception of optimal and the one discussed by OAS are unrelated. For Chomsky, FL is not optimized for any phenotypic function. There is nothing that FL is for such that we can say that it does whatever better than something else might. For example structure dependence has no function so that Gs that didn’t have it would be worse in some way than ones (like ours) that do.
Subscribe to: Posts (Atom)