Comments

Thursday, May 25, 2017

Naturalized philosophy

I went to graduate school in philosophy a long time ago. At that time, there was a premium put on “naturalized” research, the idea being that good philosophy needed grounding in a “real” (non-philosophical) discipline. It was a time when Newton and Einstein and Boyle and Godel and Poincare joined the usual dead white European males that we all know and love in the pantheon of philosophical greats. In this setting, it is no surprise that Chomsky and his work made frequent appearances in the pages of the most prestigious philo journals and was a must read for a philosopher of language. It actually took some effort for the discipline to relegate Chomsky to the domain of “philosophical naïf” (I think this was Putnam’s phrase) and it coincided with endless debates about the implications of the referentialist worldview for narrow content and semantic meaning. IMO, this work did not deliver much in the way of insight, though it did manage to make quite a few careers. At any rate, Chomsky’s exit from the main stage coincided with a waning of the naturalizing project and a return to the metaphysical (and metalinguistic) abstruseness that philosophy is, it appears, endemically attracted to. If nothing else, de-naturalizing philosophy establishes academic protective boundaries providing philosophy with a proprietary subject matter that can protect deep thinkers from the naturalizers and their empirical pretentions.[1] Why do I mention this? Because I am a big fan of the kind of naturalized philosophy that the above mentioned luminaries practiced and so I am usually on the lookout for great examples thereof.

What are the distinctive marks of this kind of work? It generally rests on a few pretty “obvious” empirical premises and demonstrates their fertile implications. Chomsky’s work offers an excellent illustration.

What is Chomsky’s most significant contribution to philosophy (and indeed linguistics)? He identified three problems in need of solution: what does a native speaker know when s/he knows her/his native language? What meta-capacity underlies a native speaker’s capacity to acquire her/his native language? And how did this meta-capacity arise in the species?  These are the big three questions he put on the table. And the subsequent questions they naturally lead to: How do native speakers use their knowledge to produce and understand language, how do LADs use their meta-capacity to acquire their native capacity? How are one’s knowledge of language embodied in wetware?  The last three rely on glimmers of answers to the first three. Chomsky has taught us how to understand the first three. 

Here’s the argument. It is based on few really really really obvious facts. First, that nothing does language like humans do. Birds fly, fish swim, humans do language. It is a species specific capacity unlike anything we find anywhere else. Second, a native speaker displays linguistic creativity. This means that a native speaker can use and understand an unbounded number of linguistic objects never before encountered and does this relatively effortlessly. Third, any kid can reflexively acquire any language when placed in the right linguistic environment (linguistic promiscuity), an environment which, when one looks even moderately closely, vastly underdetermines the knowledge attained (poverty of the linguistic stimulus). These three facts make it morally certain that part of linguistic competence implies internalization of a G, that the human meta-capacity of interest involves a higher order capacity to acquire certain kinds of Gs and not others and that this meta-capacity rests on some distinctive species specific capacities of humans. These three conclusions rest solidly on these obvious facts and together they bring forth a research program: what properties to human Gs have and what is the fine structure of the meta-capacity. That Gs exist and that FL/UG exists is trivially true. What their properties are is anything but.[2]

As FoLers know, Chomsky has recently added a third question to the agenda: how FL/UG could possibly have arisen. He argues that the relative rapidity of the emergence of FL and its subsequent stability argues for an intriguing conclusion: that the change that took place was necessarily pretty small and that whatever is proprietary to language must be quite minor. I tend to think that Chomsky is right about this, and that it motivates a research program that (i) aims to limit what is linguistically special while (ii) demonstrating how this special secret sauce allows for an FL like ours in the context of other more cognitively and computationally general mental capacities it is reasonable to believe that our pre-linguistic ancestors enjoyed. Imo, this line of thinking is less solidly based on “obvious” facts, but the line of inquiry is sufficiently provocative to be very inviting. Again, the details are up for grabs, as they should be.

So what are the marks of naturalized philosophy? Identifying questions motivated by (relatively) straightforward facts that support a framework for asking more detailed questions using conventional modes of empirical inquiry. Chomsky is a master of this kind of thinking. But he is not alone. All of the above is actually in service of advertising another such effort by Randy Gallistel. The paper of interest, which is a marvelous piece of naturalized philosophy appeared here in TiCS. I want to say a word or two about it.

Gallistel’s paper is on the coding question. The claim is that this question has been effectively ignored in the cog-neuro world with baleful effects. The aim is to put it front and center on the research agenda and figure out what kind of neural system is compatible with a reasonable answer to that question. The argument in the paper is roughly as follows.

First, there is overwhelming behavioral evidence that animals (including humans) keeps track of numerical quantities (see box 1, (3)). Hence the brain must have a way to code for number. It must be able to store these numbers in some way and must be able to transmit this stored information in signals in some way. So it must be able to write this information to memory and read this information from memory.

Second, if the brain does code for number it must do so in some code. There are various kinds, but the two the paper discusses are hash/rate/tally codes vs combinatorial codes (4-5). The former are “unary” codes. What this means is that “to convey a particular number one must use as many code elements are the numerosity to which the number refers.” Thus, if the number is 20 then there are 20 hash marks/strokes/dotes whatever representing the number.

The paper distinguishes such codes from “combinatorial” codes. These are the ones we are familiar with. So for example, ‘20’ conveys the number 20 and does so by using 10 digits in order sensitive configurations (i.e. 21 differs from 12). Note, combinatorial code patterns are not isomorphic to the things they represent.[3] 

The paper explores the virtues of combinatorial codes as against hash/rate/tally codes. The latter are “vastly more efficient” by orders of magnitude. Rate codes “convey 1 bit per spike” (5) while it is known that spike trains convey between 3-7 bits per spike. Rate codes are very energy expensive, combinatorial codes can be “exponentially smaller” (6). Last of all, there is evidence that spike trains use combinatorial codes because “reordering the intervals changes the message” (recall ‘21’ vs ‘12’), as expected if they spike trains are expressing a combinatorial code. 

The conclusion: the brain uses a combinatorial code, and this is interesting because this seems to require that the code be “symbolic” in the sense that its abstract (syntactic) structure matters for the information being conveyed.  And this strongly suggests that this info is not stored in synapses as supposed in a neural net system.

This last conclusion should not be controversial. When first put on the market of ideas, neural nets were confidently sold as being non-representational. Rumelhart and McClelland focused on this as one of their more salient properties and Fodor and Pylyshyn criticized such models for precisely this reason. The Gallistel paper is making the additional point that being asymbolic is, in addition to being cognitively problematic, is also neurophysiologically a problem as the kind of codes we are pretty sure we need are the kinds that neural nets are designed not to support. And this means that these are the wrong neuro models for the brain: “In neural net models, plastic synapses are molded by experience” and were intended to model “associative bonds” which “were never conceived of as symbols, and neither are their neurobiological proxies” (8).

Note, we can conclude that neural nets are the wrong model even if we have no idea what the correct model is. We can know what kind of code it is and what this means for the right neurophysiology without knowing what the right neurophysiology is. And if the codes are combinatorial/symbolic then there is no way that the right physiology for memory can be neural nets. This takes the Fodor-Pylyshyn critique on major step further.

So, if not in nets, what kind of architecture. Well, you all know by now. The paper notes that we can get everything we want from a chemical computer. We can physically model classical von Neuman/Turing machines in chemistry, with addresses, reading from, writing to, etc. Moreover, chemical computation has some nice biological features. Complex chemicals can be very stable for long periods of time (what we want from a long term memory store), and writing to such molecules is very energy efficient (8). In addition, chemical computing can be fast (some “can be altered on a nanosecond time scale”) and we know of instances of this kind of chemical computing that are behaviorally relevant. Last chemical computations are very energy efficient. Both storing and computing can be done cheaply if done chemically.

All of this leads to the conclusion that the locus of neurobiological computing is chemical and where are the relevant chemicals? Inside the cell. So, in place of neural nets we have the “cell intrinsic memory hypothesis” (1). Happily, there is now evidence that some computing gets done intra-celluarly (1-2). But if some gets done there…

This paper is great naturalized philosophy: we argue from pretty simple behavioral evidence that a certain kind of coding format is required and then that these kinds of formats prefer certain kinds of physical systems to support such codes and end with conclusions about the locus of the relevant computations. Thus we move from numbers are required, to combinatorial codes are the right kind, to neural nets won’t cut it to chemical computing within the cell. The big open meaty empirical question is what particular combinatorial code is exploited. It’s the cog-neuro analogue of how DNA stores genetic information and uses it. Right now, we do not know. At all.

This last analogy to DNA is important, and, IMO, is the strongest reason for thinking that this line of thinking is correct. Conventional computers provide an excellent model of how computation can be physically implemented. We know how to chemically “build” a conventional computer. We know that biology already uses chemistry to store and use information in hereditary and development. Is it really plausible that this in place machinery is not used for cognitive computation? Or as the paper puts it in the last sentence: “Why should the conveyance of acquired information proceed by principles fundamentally different from those that govern the conveyance of heritable information?” Why indeed! Isn’t the contrary assumption (the machinery is there for the using but it is never used) biologically scandalous? Wouldn’t Darwin be turning in his grave if he considered this? Isn’t assuming it to be false a kind of cognitive creationism? Yup, connetionists and neural net types are the Jerry Falwells of biology! Who would have thunk it: the road from Associationsim to Creationism is paved with Empiricist intentions. Only Rationalism and Naturalized Philosophy can save you.



[1] Let me quickly add that I consider philosophy training a very useful aid to right thinking. Nothing allows you to acquire the feel for good argumentation than a stressful philosophical workout. And by “good” I mean understanding how premises related to conclusions, how challenging premises can allow one to understand how to evaluate conclusions, understanding that it is always reasonable to ask what would happen to a conclusion should such and such a premise be removed etc.  In other words, philosophy prizes deductive structure and this is a useful talent to nurture regardless of what conclusions you are interested in netting and premises you are interested in frying.
[2] Chomsky, as you all know, not only posed the questions but showed how to go about empirically investigating them. This is what puts him in with the Gods: he discovered interesting questions and figured out technology relevant to answering them.
[3] Thus, the numeral’s patterning represents the number in the former but not the latter. The difference between the two kinds of codes is similar to the one made (here) between patterns that track the patterning and those that do not.

1 comment:

  1. Well said! Chomsky is a great philosopher in his own right. If he had published nothing other than Aspects chp1, Cartesian Linguistics, and the first two chps. of K of L, he would still be head and shoulders above most of the last 50 years. As it is...

    ReplyDelete