What's innate?

Johan Bolhuis sent me a copy of a recent comment in TiCS(Priors in animal and artificial intelligence (henceforth Priors))on the utility of rich innate priors in cognition, both in actual animals and artificially in machines. Following Pinker, Priorsframes the issue in terms of the blank slate hypothesis (BSH) (tabula rasafor you Latin lovers). It puts the issue as follows (963):

Empiricists and nativists have clashed for centuries in understanding the architecture of the mind: the former as a tabula rasa, and the latter as a system designed prior to experience…The question, summarized in the debate between the nativist Gary Marcus and the pioneer of machine learning, Yann LeCun, is the following: shall we search for a unitary general learning principle able to flexibly adapt to all conditions, including novel ones, or structure artificial minds with driving assumptions, or priors, that orient learning and improve acquisition speed by imposing limiting biases?

Marcus’ paper (here) (whose philosophical framework Priorsuses as backdrop for its more particular discussion) relates BSH to the old innateness question, which it contends revolves around “trying to reduce the amount of innate machinery in a given system” (1). I want to discuss this way of putting things, and I will be a bit critical. But before diving in, I want to say that I really enjoyed both papers and I believe that they are very useful additions to the current discussion. They both make excellent points and I agree with almost all their content. However, I think that the way they framed the relevant issue, in terms of innateness and blank slates, is misleading and concedes too much to the Empiricist (E) side of the debate. 

My point will be a simple one: the relevant question is not how much innate machinery, but what kindof innate machinery. As Chomsky andQuine observed a long time ago, everyonewho discusses learning and cognition is waste deep in a lot of innate machinery. The reason is that learning without a learning mechanismis impossible. And if one has a learning mechanism in terms of which learning occurs, then that learning mechanism is not itself learned. And if it is note learned then it is innate. Or, to put this more simply, the mechanism that allows for learning is a precondition for learning and preconditions are fixed prior to that which they precondition. Hence all features of the learning mechanism are innate in the simple sense of not themselves being learned. This is a simple logical point, and all who discuss these issues are aware of this point. So the question is not, never has been, and never could not have been is there innate structure?Rather the question is, always has been and always will be what structure is innate?

Why is putting things in this way important? Because arguing about the amountof innate structure gives Eists the argumentative edge. Ockham like considerations will always favor using less machinery rather than more all things being equal. So putting things as the Marcus paper and Priorsdoes is to say that the Eist position is methodologically preferable to the Rationalist (R) one. Putting things in terms of what kinds of innate machinery is required (to solve a given learning problem), rather than how much considerably levels the methodological playing field. If both E and R conceptions require boatloads of innate machinery to get anywhere, then the question moves from whether innate structure is needed (as the BSH slyly implicates) to what sort is needed (which is the serious empirical question).

This said, let’s zero in on some specifics. What makes an approach Eist? There are two basic ingredients. The first important ingredient is associationism (Aism). This is the glue that holds “ideas” together. However, this is not all. There is a second important ingredient: perceptualism (Pism). Pism is the idea that all mental contents are effectively reducible to perceptual contents, which are themselves effectively reducible to sensory concepts (sensationalism (Sism)). 

This pair of claims lies at the center of Eist theories of mind. And the notion of the blank slate emphasizes the second. We find this reflected in a famous Eish slogan: “there is nothing in the mind that is not first in the senses.” The Eish conception unites P/Sism with Aism to get to the conclusion that all mental contents are either primitive sensory/perceptual “ideas” or constructed out of sensory/perceptual input via association. The problems with Eism arise from both sources and revolve around two claims: the denial that mental concepts interrelate other than by association (they have no further interesting logical structure) and that all ideas are congeries of sensory perceptions. These two assumptions combine to provide a strong environmentalist approach to cognition wherein the structure of the environment largely shapes the contents of the mind through the probabilistic distributions of sensory/perceptual inputs. Rism denies bothclaims. It argues that association is not the fundamental conceptual glue that relates mental contents anddenies that all complex mental contents are combinations of sensory/perceptual inputs. To wax metaphorical, for Eists, only sensation can write on our mental blank slates and the greater the sensations the more vivid the images that appear. Rists think this is empirical bunk.

Note that this combination of cognitive assumptions has a third property. Given Eist assumptions, cognition is general purpose. If cognition is nothing but tracking the frequencies of sensory inputs then all cognition is of a piece, the only difference being the sensations/perceptions being tracked. There is no modularity or domain specificity beyond that afforded by the different sensory mechanisms, nor rules of “combination” beyond those tracking the differential exposure to some sensations over others. Thus for Eists, the domain generality of cognition is not an additional assumption. It is the consequence of Eisms two foundational premises.

Now, we actually know today that Eism will not work (actually, we knew this way back way back when). In particular, Pism/Sism was very thoroughly explored at the turn of the 20^thcentury and shown to be hopeless. There were vigorous attempts to reduce our conceptual contents to sense data. And these efforts completely failed! Pism/Sism, in other words, is a hopeless position. So hopeless, in fact, that the only place it still survives is in AI and certain parts of psychology. Deep Learning (DL), it seems, is the latest incarnation of P/Sism+Aism right now. BothPriorsand Marcus elegantly debunk DLs inflated pretentions by showing both that the assumptions are biologically untenable and that they are adhered to more in the PR discussions than in the practice of the parade cases meant to illustrate successful AI learners.[1] I refer you to their useful discussions. See especially their excellent points concerning how much actual learning in humans and animals is based on very little input (i.e. from a very limited number of examples). DL requires Big Data (BD) to be even remotely plausible. And this data must be quite carefully curated (i.e. supervised) to be of use. Both papers make the obvious point that much biological learning is done from very few example cases (sparse data) and is unsupervised (hence notcurated). This makes most of what DLers have “discovered” largely irrelevant as models for biologically plausible theories of cognition. Sadly, the two papers do notcome right out and say this, though they hint at it furiously. It seems that the political power of DL is such that frankly saying that this emperor is hardly clothed will not be well rewarded.[2]Hence, though the papers make this point, it is largely done in a way that bends over backwards to emphasize the virtues of DL and not appear to be critically shrill. IMO, there is a cost to this politeness.

One last point and I stop. Priorsmakes a cute observation, at least one that I never considered. Eists of the DL and connectionist variety loveplasticity. They want flexible minds/brains because these are what the combination of Aism and P/Sism entails.Priorsmakes the nice observation that if flexibility is understood as plasticity then plasticity is something that biology only values in smalldoses. Brains cease being plastic after a shortish critical period. This Priorsnotes implies that there is a biological cost of being relentlessly open minded. You can see why I might positively reverberate to this observation.

Ok, nuff said. The two papers are very good and are shortish as well. Priorsis perfect for anyone wanting to have a non human case to illustrate Rish themes in a class on language and mind. The Marcus piece is part of a series of excellent papers he has been putting out reviewing the hype behind DL and taking it down several pegs (though, again, I wish he were less charitable). From these papers and the references they cite, it strikes me that the hype that has surrounded DL is starting to wear thin. Call me a hopeless romantic, but maybe when the overheated PR dies down and it becomes clear that the problems the latest round of Eish accounts solved were not the central problems in cognition, we can return to some serious science.  

---

[1]An aside: there is more than a passing similarity between the old attempts to reduce mental contents to sense data and the current fad in DL of trying to understand everything in terms of pixel distributional properties. History seems to constantly repeat; the first time as insight, the second time as a long con. Not surprisingly, the attempt to extract the notion “object” or “cat” from pixel distributions is no more successful today than were prior attempts to squeeze such notions from sense data. Ditto with algebraic structure from associations. It is really useful to appreciate how long we have known that Eism cannot be a serious basis for cognition. The failures Priorsand Marcus observe are not new ones, just the same old failures gussied up in technically spiffier garb.

[2]Some influential voices are becoming far more critical. Shalizi (here) notes that much of DL is simply a repackaging of perceptrons (“extracting features from the environment which work in that environment to make a behaviorally-relevant classificationor prediction or immediate action”) and will have roughly the same limitations that perceptrons had (viz. “This sort of perception is fast, automatic, and tuned to very, very particular features of the environment… They generalize to more data from their training environment, but not to new environments…”).  Shalizi, like Marcus andPriors, locates the problems with these systems in their lack of “abstract, compositional, combinatorial understanding we (and other animals) show in manipulating our environment, in planning, in social interaction, and in the structure of language.” 

            In other words, DL is basically the same old stuff repackaged for the credulous “smart” technopilic shopper. You cannot keep selling perceptrons, so repackage and sell it as DeepLearning (the ‘deep’ here is, no doubt, the contribution of the marketing department). The fact is that the same stuff that was problematic before is problematic still. There is no way to “abstract” out compositional and combinatorial principles and structures from devices aimed to track “particular features of the environment.” 

Two pictures of the mind?(brain)

Empiricists (E) and Rationalists (R) have two divergent “pictures” of how the mind/brain functions (henceforth, I use ‘mind’ unless brains are the main focus).[1]

For Es, the mind/brain is largely a passive instrument that, when running well, faithfully records the passing environmental scene. Things go awry when the wrong kinds of beliefs intrude between the sensory input and receptive mind to muddy the reception. The best mind is a perfectly receptive mind. Passive is good. Active leads to distortion.[2]

For Rs there is no such thing as a passive mind. What you perceive is actively constructed along dimensions that the mind makes available. Perception is constructed. There is no unvarnished input, as transduction takes place along routes the mind lays out and regulates. More to the point, sensing is an activity guided by mental structure.

All of this is pretty old hat. However, that does not mean that it has been well assimilated into the background wisdom of cog-neuro.  Indeed, from what I can tell, there are large parts of this world (and the closely related Big Data/Deep Mind world) that take the R picture to be contentious and the E picture to be obvious (though as we shall see, this seems to be changing).  I recently ran across several nice pieces that discuss these issues in interesting ways that I would like to bring to your attention. Let me briefly discuss each of them in turn.

The first appeared here (let’s call the post TF (Teppo Felin being the author)) and it amusingly starts by discussing that famous “gorilla” experiment. In case you do not know it, it goes as follows (TF obligingly provides links to Youtube videos that will allow you to be a subject and “see” the gorilla (or not) for yourself). Here is TF’s description (2):

 In the experiment, subjects were asked to watch a short video and to count the basketball passes. The task seemed simple enough. But it was made more difficult by the fact that subjects had to count basketball passes by the team wearing white shirts, while a team wearing black shirts also passed a ball. This created a real distraction.

The experiment came with a twist. While subjects try to count basketball passes, a person dressed in a gorilla suit walks slowly across the screen. The surprising fact is that some 70 per cent of subjects never see the gorilla. When they watch the clip a second time, they are dumbfounded by the fact that they missed something so obvious. The video of the surprising gorilla has been viewed millions of times on YouTube – remarkable for a scientific experiment. Different versions of the gorilla experiment, such as the ‘moonwalking bear,’ have also received significant attention.

Now, it’s hard to argue with the findings of the gorilla experiment itself. It’s a fact that most people who watch the clip miss the gorilla.

The conclusion that is generally drawn (including by heavyweights like Kahneman) is that humans are “ ‘blind to the obvious, and blind to our blindness.’” The important point that TF makes is that thisdescription of the result presupposes that there is available a well defined mind independent notion of “prominence or obviousness.” Or, in my (tendentious) terms, it presupposes an Eish conception of perception and a passive conception of the mind.  The problem is that this conception of obviousness is false. As TF correctly notes, “all kinds of things are readily evident in the clip.” In fact, I would say that there are likely to be an infinite number of possible things that could be evident in the clip in the right circumstances. As Lila Gleitman once wisely observed, a picture is worth a thousand words and that is precisely the problem. There is no way to specify what is “obvious” in the perception of the clip independent of the mind doing the perceiving. As TF puts it, obviousness only makes sense relativized to perceivers’ mental capacities and goals. 

Now, ‘obviousness’ is not a technical cog-neuro term. The scientific term of art is ‘salience.’ TF’s point is that it is quite standardly assumed that salience is an objective property of a stimulus, rather than a mind mediated relation. Here is TF on Kahneman again (3).

Kahneman’s focus on obviousness comes directly from his background and scientific training in an area called psychophysics. Psychophysics focuses largely on how environmental stimuli map on to the mind, specifically based on the actual characteristics of stimuli, rather than the characteristics or nature of the mind. From the perspective of psychophysics, obviousness – or as it is called in the literature, ‘salience’ – derives from the inherent nature or characteristics of the environmental stimuli themselves: such as their size, contrast, movement, colour or surprisingness. In his Nobel Prize lecture in 2002, Kahneman calls these ‘natural assessments’. And from this perspective, yes, the gorilla indeed should be obvious to anyone watching the clip. 

TF gets one thing askew in this description IMO: the conception of salience it criticizes is Eish, not psychophysical.[3]True, psychophysics aims to understand how sensation leads to perception and sensations are tied to the distal stimuli that generate them. But this does not imply that salience is an inherent property of the distal stimulus. The idea that it is, is pure Eism. On this view, minds that “miss” the salient features of a stimulus are minds that are misfiring. But if minds makestimuli salient (rather than simply tracking what is salient), then a mind that misses a gorilla in a video clip when asked to focus on the number of passes being executed by members of a team may be functioning perfectly well (indeed, optimally). For this purpose the gorilla is a distraction and an efficient mind with the specific count-the-passes mandate in hand might be better placed to accomplish its goal were it to “ignore” the gorilla in the visual scene.[4]

Let me put this another way: if minds are active in perception (i.e. if minds are as Rs have taken them to be) then salience is not a matter of what you are looking atbut what you are looking for (this is TF’s felicitous distinction). And if this is so, every time you hear some cog-psych person talking about “salience” and attributing to it causal/explanatory powers, you should appreciate that what you are on the receiving end of is Eish propaganda. It’s just like when Es press “analogy” into service to explain how minds generalize/induce. There is no scientifically usefully available notions of either except as relativized to the specific properties of the minds involved. Again as TF puts it (4):

Rather than passively accounting for or recording everything directly in front of us, humans – and other organisms for that matter – instead actively look for things. The implication (contrary to psychophysics[5]) is that mind-to-world processes drive perception rather than world-to-mind processes.

Yup, sensation and perception are largely mind mediated activities. Once again, Rism is right and Eism is wrong (surprise!).

Now, all of this is probably obvious to you(at least once it is pointed out). But it seems that these points are still considered radical by some. For example, TF rightly observes that this view permeates the Big Data/Deep Learning (BD/DL) hoopla. If perception is simply picking out the objectively salient features of the environment unmediated by distorting preconceptions, then there is every reason to think that being able to quickly assimilate large amounts of input and statistically massage them quickly is the road to cognitive excellence. Deep Minds are built to do just that, and that is the problem (see herefor discussion of this issue by “friendly” critics of BD/DL). 

But, if Rism is right, then minds are not passive pattern matchers or neutral data absorbers but are active probers of the passing scene looking for information to justify inferences the mind is built to make. And if this is right, and some objective notion of salience cannot be uncritically taken to undergird the notion of relevance, then purely passive minds (i.e. current Deep Minds) won’t be able to separate what is critical from what is not. 

Indeed, this is what lies behind the failure of current AI to get anywhere on unsupervised learning. Learning needs a point of view. Supervised learning provides the necessary perspective in curating the data (i.e. by separating out the relevan-to-the-task (e.g. find the bunny)) data from the non-relevant-to-the-task data). But absent a curator (that which is necessarily missing from unsupervised learning), the point of view (what is obvious/salient/relevant) must come from the learner (i.e. in this case, the Deep Mind program). So if the goal is to get theories of unsupervised learning, the hard problem is to figure out what minds consider relevant/salient/obvious and to put this into the machine’s mind. But, and here is the problem, this is precisely the problem that Eism brackets by taking salience to be an objective feature of the stimulus. Thus, to the degree that BD/DL embrace Eism (IMO, the standard working assumption), to that degree it will fail to address the problem of unsupervised learning (which, I am told, is theproblem that everyone (e.g. Hinton) thinks needs solving).[6]

TF makes a few other interesting observations, especially as relates to the political consequences of invidiously comparing human and machine capacities to the detriment of the former. But for present purposes, TF’s utility lies in identifying anotherway that Eism goes wrong (in addition, for example, to abstracting away from exactly how minds generalize (remember, saying that the mind generalizes via “analogy” is to say nothing at all!)) and makes it harder to think clearly about the relevant issues in cog-neuro.

Sam Epstein develops this same theme in a linguistic context (here (SE)). SE starts with correctly observing that the process of acquiring a particular G relies on two factors, (i) an innate capacity that humans bring to the process and (ii) environmental input (i.e. the PLD). SE further notes that this two factor model is generally glossed as reflecting the contributions of “nature” (the innate capacity) and “nurture” (the PLD). And herein we find the seeds of a deep Eish misunderstanding of the process, quite analogous to the one the TF identified.  Let me quote SE (197-198):

[I]t is important to remember—as has been noted before, but

perhaps it remains underappreciated—that it is precisely the organism’s biology

(nature) that determines what experience, in any domain, can consist of …

To clarify, a bee, for example, can perform its waggle dance for me a million times, but that ‘experience’, given my biological endowment, does not allow me to transduce the visual images of such waggling into a mental representation (knowledge) of the distance and direction to a food source. This is precisely what it does mean to a bee witnessing the exact same environmental event/waggle dance. Ultrasonic acoustic disturbances might be experience for my dog, but not for me. Thus, the ‘environment’ in this sense is not in fact the second factor, but rather, nurture is constituted of those aspects of the ill-defined ‘environment’ (which of course irrelevantly includes a K-mart store down the street from my house) that can in principle influence the developmental trajectory of one or more organs of a member of a particular species, given its innate endowment.

In the biolinguistic domain, the logic is no different. The apparent fact that

exposure to some finite threshold amount of ‘Tagalog’ acoustic disturbances in

contexts (originating from outside the organism, in the ‘environment’) can cause

any normal human infant to develop knowledge of ‘Tagalog’ is a property of

human infants…. Thus the standard statement that on the one hand, innate properties of the organism and, on the other, the environment, determine organismic development, is profoundly misleading. It suggests that those environmental factors that can influence the development of particular types of organisms are definable, non-biologically—as the behaviorists sought, but of course failed, to define ‘stimulus’ as an organism-external construct. We can’t know what the relevant developmental stimuli are or aren’t, without knowing the properties of the organism.

This is, of course, correct. What counts as input to the language acquisition device (LAD) must be innately specified. Inputs do not come marked as linguistically vs non-linguistically relevant. Further what the LAD does in acquiring a G is the poster child example of unsupervised learning. And as we noted above, without a supervisor/curator selecting the relevant inputs for the child and organizing them into the appropriate boxes it’s the structure of the LAD that mustbe doing the relevant curating for itself. There really is no other alternative. 

SE points out an important consequence of this observation for nature vs nurture arguments within linguistics, including Poverty of Stimulus debates.  As SE notes (198): 

… organism external ‘stimuli’ cannot possibly suffice to explain any aspects of the developed adult state of any organism. 

Why? For the simple reason that the relevant PLD “experience” that the LAD exploits is itself a construction of the LAD. The relevant stimulus is the proximal one, and in the linguistic domain (indeed in most cognitively non-trivial domains) the proximal stimulus is only distantly related to the distal one that triggers the relevant transduction. Here is SE once more (199):

…experience is constructed by the organism’s innate properties, and is very different from ‘the environment’ or the behaviorist notion of ‘stimulus’.

As SE notes, all of this was well understood over 300 years ago (SE contains a nice little quote from Descartes). Actually, there was a lively discussion at the start of the “first cognitive revolution” (I think this is Chomsky’s term) that went under the name of the “primary/secondary quality distinction” that tried to categorize those features of proximate stimuli that reflected objective features of their distal causes and those that did not. Here appears to be another place where we have lost clear sight of conceptual ground that our precursors cleared.

SE contains a lot more provocative (IMO, correct) discussion of the implications of the observation that experience is a nature-infested notion. Take a look.

Let me mention one last paper that can be read along side TF and SE. It is on predictive coding, a current fad, apparently, within the cog-neuro world (here). The basic idea is that the brain makes top down predictions based on its internal mental/brain models about what it should experience, perception amounting to checking these predictions against the “input” and adjusting the mental models to fit these. In other words, perception is cognitively saturated. 

This idea seems to be getting a lot of traction of late (a piece in Quantais often a good indicator that an idea is “hot”). For our purposes, the piece usefully identifies how the new view differs from the one that was previously dominant (7-8):

The view of neuroscience that dominated the 20th century characterized the brain’s function as that of a feature detector: It registers the presence of a stimulus, processes it, and then sends signals to produce a behavioral response. Activity in specific cells reflects the presence or absence of stimuli in the physical world. Some neurons in the visual cortex, for instance, respond to the edges of objects in view; others fire to indicate the objects’ orientation, coloring or shading…

Rather than waiting for sensory information to drive cognition, the brain is always actively constructing hypotheses about how the world works and using them to explain experiences and fill in missing data. That’s why, according to some experts, we might think of perception as “controlled hallucination.”

Note the contrast: perception consists in detecting objective features of the stimulus vs constructing hypotheses about how the world works verified against bottom up “experience.” In other words, a passive feature detector vs an active mind constructing hypothesis tester.  Or, to be tendentious one more time, an Eish vs an Rish conception of the mental. 

One point worth noting. When I was a youngster oh so many decades ago, there was a big fight about whether brain mechanisms are largely bottom up or top down computational systems. The answer, of course, is that it uses both kinds of mechanisms. However the prevalent sentiment in the neuro world was that brains were largely bottom up systems, with higher levels generalizing over features provided by lower ones. Chomsky’s critique of discovery procedures (see herefor discussion) hit at exactly this point, noting that in the linguistic case it was not possible to treat higher levels as simple summaries of the statistical properties of lower ones. Indeed, the flow of information likely went from higher to lower as well. This has a natural interpretation in terms of brains mechanisms involving feed forward as well as feed back loops. Interestingly, this is what has also driven the trend towards predictive coding in the neuro world. It was discovered that the brain has many “top down feedback connections” (7)[7]and this sits oddly with the idea that brains basically sit passively waiting to absorb perceptual inputs. At any rate, there is an affinity between thinking brains indulge in lots of feed forward processing and taking brains to be active interpreters of the passing perceptual scene.

That’s it. To repeat the main message, the E vs R conceptions of the mind/brain and how it functions are very different, and importantly so. As the above papers note, it is all too easy to get confused about important matters if the differences between these two views of the mental world are not kept in mind. Or, again to be tendentious: Eism is bad for you! Only a healthy dose of Rism can protect you from walking its fruitless paths. So arm yourself and have a blessed Rish day.

---

[1]They also have two divergent pictures of how data and theory relate in inquiry, but that is not the topic of today’s sermonette.

[2]I have argued elsewhere (here) that this passivity is what allows Es to have a causal semantic theory. 

[3]Nor from what I can gather from Kahneman’s Noble lecture is he committed to the view that salience is a property of objects. Rather it is a property of situations a sentient agent finds herself in. The important point for Kahneman is that they are more or less automatic, fast, and unconscious. This is consistent with it being cognitively guided rather than a transparent reflection of the properties of the object. So, though TF’s point is useful, I suspect that he did not get Kahneman quite right. Happily none of that matters here.

[4]A perhaps pointless quibble: the fact that people cannot reportseeing a gorilla does not mean that they did not perceive one. The perceptual (and even cognitive) apparatus might indeed have registered a gorilla without it being the case that that viewers can access this information consciously. Think of being asked about the syntax of a sentence after hearing it and decoding its message. This is very hard to retrieve (it is below consciousness most of the time) but that does not mean that the syntax is not being computed. At any rate, none of this bears on the central issues, but it was a quibble that I wanted to register.

[5]NH: again, I would replace ‘psychophysics’ with ‘Eism.’

[6]As TF notes, this is actually a very old problem within AI. It is the “frame problem.” It was understood to be very knotty and nobody had any idea how to solve it in the general case. But, as TF noted, it has been forgotten “amid the present euphoria with large-scale information- and data-processing” (6).

            Moreover, it is a very hard problem. It is relatively easy to identify salient features givena context. Getting a theory of salience, in contrast, (i.e. a specification of the determinants of salience acrosscontexts) is very hard. As Kahneman notes in his Nobel Lecture (456), it is unlikely that we will have one of these anytime soon. Interestingly, early on Descartes identified the capacity for humans to appropriatelyrespond to what’s around them as an example of stimulus free (i.e. free and creative) behavior. We do not know more about this now than Descartes did in the 17^thcentury, a correct point that Chomsky likes to make.

[7]If recollection serves (but remember I am old and on the verge of dementia) the connections from higher to lower brain levels is upwards of five times those from lower to upper. It seems that the brain is really eager to involve higher level “expectations” in the process of analyzing incoming sensations/perceptions.

Newton as empiricist

One thing Empiricism (E) got right is that there are no foundational assumptions of general scientific utility that are empirically inviolable. Rationalists (R) once thought otherwise, thinking that the basics of the mechanical philosophy (matter is geometrical and all forces are contact forces) as well as an innate appreciation of some of God’s central features could undergird permanent foundations for investigation of the physical world. One of Newton’s big debates with the Cartesians revolved around this point and he argued (as convincingly examined hereby Andrew Janiak) that we have no privileged access to such theoretical starting points. Rather, even our most basic assumptions are subject to empiricalevaluation and can be overturned.

This is now the common view. The Rs were wrong, the Es right. Or at least someof them were. Descartes and his crew were metaphysical foundationalists in that they took the contours of reality to be delivered via clear and distinct innateideas. As these ideas reflected what had to be the casethey served to found physical inquiry in a mechanical view of the world. So for Rs innate ideas were critical in the sciences in that they were metaphysically foundational. Newton denied this. More exactly, he denied that there were empirically unassailable foundations for natural philosophy. And he did this by arguing that even foundational assumptions (e.g. the mechanical philosophy) were subject to experimental evaluation. This he proceeded to do in the Principia Andrew Janiakwhere he argued that the mechanical philosophers had it all wrong, and that mass and gravity were inconsistent with mechanism and so false. Great argument.

As Janiak tells the story, three features of Newton’s story really made his conclusions convincing. First, he provided a mathematical formulation of the force laws and the law and gravitation. Second, he showed how these unify terrestrial and celestial mechanics. That they wereunified was a staple of Cartesian mechanical thinking. However, they could not show that this conviction was scientifically justifiable. Newton unified the two domains via gravity (plus he threw in the tides as a bonus) thus achieving what the mechanical philosophers wanted by using a force they despised. Third, Newton provided a principled account for Galileo’s observation that acceleration was independent of the shape/size/density of the accelerating objects. Why this should be so was a real puzzle, and an acknowledged one. The link between gravitation and mass that Newton forged had this independence fall out as a trivial consequence. The second and third achievements were substantive within the framework of Cartesian physics (they solved problems that Cartesians recognized as worthy of explanation) but they were inconsistent with Cartesian mechanical philosophy (because they were based on Newton’s conceptions of mass and gravity), as was widely understood. This is why Newton’s physical interpretation of his formal work was resisted, though everyone agreed that the math was wondrous. The problem was not the math, but what Newton took the math to mean.

The Janiak book goes into lots of detail regarding Newton’s argumentative approach to arguing against Cartesian orthodoxy. It is a really great read. The principle form of the argument goes as follows: it is possible to know thatsomething is true without knowing exactly howit can be.[1]This involved unpacking Newton’s famous dictum that the does not feign hypotheses. This Janiak notes does not mean that Newton did not advance theories. Clearly he did. Rather Newton meant that he believed that he could both say that he knew something to be the case and that he did not know exactly how what he knew to the case could be the case. Sound familiar? This is a critical distinction and Newton’s point is that unless one carefully distinguishes the questions one is asking it is very hard to evaluate the quality of the answers. Some data for some questions can be completely compelling. The same data for other questions hardly begin to scratch the surface. 

Newton showed that it is both possible to know thatGravity is real and not know exactly whyit has the properties we know it to have or even to exactly know whatit is (a relation or a quality). What Newton did know was that one could believe thatgravity was real without thinking that it was an inherent property of bodies that acted at a distance. He insisted that local action was consistent with gravity (there was no contradiction between the two) even though he did not know how it could be (he had no constructive theory of local action that included gravity). 

I would have put things somewhat differently: Newton knew that gravity existed and what its signature features were and some central cases of its operations. He had an effective theory. But he did not believe that he had a fundamental account of gravity, though he knew that there could not be a mechanical explanation of these gravitational effects. Moreover, he knew that nothing known at the time implied that it was inconsistent with local action of some other yet to be determinedkind. So, he knew a lot but not everything, and that was good enough for him.

Note what convinced contemporaries: unification, and explanation of outstanding generalizations. So too today: we want to unify distinctive domains in syntax (MP requires this) and want to explain why the generalizations we appear to have discovered hold. This is the key to moving forward, and imitating Newton is not the worst path forward.

So, Eism did come out right on this point, but most Es did not. Newton was somewhat of an outlier on these foundational matters. He rightly understood that Cartesian metaphysics and its innate ideas window into reality would not wash. But he did not appear to embrace an Eist epistemology (or at least theory of mind). Unlike many, he did not endorse the view that there are no ideas in the intellect that are not first in the senses. Thus, his beef with Rists was not their nativism but their supposition that being innate conferred some kind of privileged metaphysical status. The way Newton short circuited the metaphysical conclusions was by showing that they did not hold up to experimental and theoretical scrutiny. He did not argue that the ideas could not be useful because they could not exist (i.e. there are no innate ideas) or that the only decent ideas were those based in perception. This was fortunate for curiously the Eists did not really grasp N’s basic take home message: ideaswherever they come fromrequire scientific validation through experiment and theory. And every part of every theory is in principle up for grabs, whatever its psychological source. Eists have tended not to understand that this was Newton’s message and have concluded that because Newton showed that Rist foundationalism failed that it failed because it countenanced innate ideas. 

But this was not the problem. The problem came with the added assumption that innate ideas in virtue of being innate are empirically unassailable. Curiously, Eists came up with their own form of foundationalism based on their view that the only good idea are the ones based in the senses. This idea didn’t turn out very well either and gave rise to its own forms of foundationalism, also false by Newton’s standards.

So, Rs were wrong about the relation of ideas to truth even though they were largely right about the psychology. Newton was right that there are no useful a priori foundations for the sciences (i.e. foundations that are not ultimately empirically justified). Es were right in that they believed that Newton did show that Rism was wrong to the degree that it was foundationalist. Where Es got lost is in rejecting the Rish psychology becauseit failed to provide metaphysical foundations. Eism ended up looking for epistemological foundations that would demarcate legit (i.e. scientific thinking) from non legit thinking (everything else). Eists pursued the idea that if one grounded ones ideas in the senses will guarantee good foundations. Newton would not have approved. The right conclusion is that that there are no epistemological shortcuts to good science. 

---

[1]These are two questions that, though related, should not be confused. I have argued that they often have been within linguistics too. Thathumans have an FL that undleries their unique capacity to acquire language is almost a tautology, IMO. How FL allows humans to do this is a substantive problem that we have just started to crack. 

Mendivil-Giro cleans the Augean stables

I am delighted to be writing this very short post advertising a very nice paper. It has appeared in the Journal of Linguisticsbut is available on lingbuzz (here). The paper is aptly entitled Is Universal Grammar ready for retirement? A short review of a longstanding misinterpretation. The author is Jose-Luis Mendivil-Giro (MG) (put in appropriate diacritics on the vowels). Here is the abstract:

In this paper I consider recent studies that deny the existence of Universal Grammar (UG), and I show how the concept of UG that is attacked in these works is quite different from Chomsky’s, and thus that such criticisms are not valid. My principal focus is on the notions of “linguistic specificity” and of “innateness”, and I conclude that, since the controversy about UG is based on misinterpretations, it is rendered sterile and thus does unnecessary harm to linguistic science. I also address the underlying reasons for these misunderstandings and suggest that, once they have been clarified, there is much scope for complementary approaches that embrace different research traditions within current theoretical linguistics.

The paper reads quickly and is surprisingly judicious and generous without being conciliatory.  Readers will note that I have made similar points far less charitably in FoL. MG surmises that the reason for the multiple confusions he identifies lies in the perfectly reasonable fact that different people are (or can be) interested in different issues relating to the wide ranging concept of ‘language.’ Perhaps. There are indeed different people interested in different things and given the complexity of the phenomena we categorize under the term ‘language.’ Further, MG is right to think that these different approaches are complementary rather than incompatible. FoL has made exactly this point several times. However, I believe that MG is being far too generous with GGs critics. I doubt that MG has correctly identified the source of the confused discussion in the literature. And one reason I believe this is that MG’s point has been made repeatedly over the last 60 years to absolutely no avail. GGers have generally bent over backwards conceding that there is room for non-GG style work in investigating the myriad properties that language knowledge and use have. What GG has insisted upon is that it’s own style of work addresses real questions and provides legitimate answers to these questions. Critics have repeatedly rejected this, as MG’s own excellent review of the literature amply demonstrates. So, if there is a confusion (or “misinterpretations”), it is rabid, and not traceable to mere differences to tastes in scientific questions. It has deeper roots. 

Ok, let me say it: the difference really lies in two incompatible conceptions of what science consists in, especially as regards the mental/behavioral sciences. The Empiricism/Rationalism (E/R) divide is the one that I have in mind, but as I have discussed it endlessly on FoL I will not go over it again here. Suffice it to say, that ifone is an Eist then GG is basically muddleheaded confusion. It cannotbe right and so its results need notbe considered. Consequently, if GG’s critics were largely Eish, it would explain the depth of their misunderstanding and their congenital inability to resist confusion/misinterpretation. 

Here’s what I mean. The tenor of many (most?) of the critiques as MG notes hardly ever go into any detail concerning specific GG proposals. As MG notes this results in critiques that are overwhelmingly dumb. The sheer ignorance of the critical discussion is wondrous to behold. The critics that MG cites and discusses really appear to know nothing at all and many (most?) completely ignore everything that GG has discovered over 60 years of research. MG notes this, and seems a bit disoriented by the fact that the main culprits seem so blithely uninformed. And it is not just one or two. They are alllike this, from Chater and Christiensen to Tomasello, Everettt, Levinson etc. etc. etc. Their critiques are really useless (and many times based on simple equivocation (I am talking to you Everett!), even if they contain a grain of truth or two (though color me very skeptical, I have been told that Tomasello’s stuff has someinteresting points) that are worth preserving given a reasonable conception of the enterprise. These kinds of “misunderstanding” are best explained methodologically. The critics don’t go into the details because they don’t believe the problem is one of detail. It is one of principle. The GG enterprise is faulty because its Rish presuppositions are untenable. If you believe this (and these people do, really!), then it is no wonder that they don’t do a deep dive into the details and confront what GGers take to be their most significant contributions.

 In other words, for the critics, the problem is the GG belief that a reasonable view of language would root the research program in an Rish vision of science in general and the mental/behavioral sciences in particular. The critics, being Eish, reject this, and as the divide between E and R conceptions is wide, we can identify its basic unbridgeability as the underlying source of the shockingly shoddy criticisms that MG so ably surveys. Given this, I am far less hopeful than MG is that “there is a glimmer of hope” (p. 23) that these disagreements will be resolved in a rational manner.[1]They cannot be for the very idea of what is the right form of “rational” inquiry is what is being debated.

I have other quibbles with the paper. For example, I found the discussion of reduction and emergence in section 3 somewhat confusing in that it mixes up two different questions: how do linguistic claims get cashed out in wetware? and are linguistic primitives reducible to those of other cognitive domains? These are different questions (as I am sure MG knows) but the paper seems to run them together. The question of FL’s linguistic “specificity” relates more to the second than the first. Of course, if we assume that cognition supervenes on brains and brains are made up of regular biological material then linguistic objects, dependencies and principles even if very linguistically sui generiswill live in biological tissue of these brains. Where else?[2]

However, that is not, nor has it ever been the relevant issue. The question has always been whether the FoL is cognitively independent. To put this crudely in “program” talk: is the FoL program just cobbled together from routines extant in other domains of animal cognition or does it require its own specific features (primitives, subroutines, addressing mechanisms etc.). One might imagine that FL is a kind of Rube Goldberg device assembled from bits and pieces of other available cognitive faculties. This is a possibility. However, I personally doubt it, and the Merge Hypothesis (i.e. that Merge is the linguistically specific sauce that one needs to add to general cognitive and computational powers to yield FL) does as well, though it limits the specificity to this one small operation. 

Honesty compels us (me!) to admit that, to date, no minimalist account has managed to eliminate all operations rather than Merge in accounting for well established features of FL. So, to date, there is reason to think that there is more to the UG parts of FL than just Merge.[3]So whereas the Minimalist Program’s ambitions are alive and well, to date, there is still quite a bit of air between the hopes and the results. And to date, there is good reason to think that FL has quite a bit more UG in it than the standard advertising supposes. This is not a serious problem for the program, but it is worth keeping in mind when we advertise the ambitions given that the program is not exactly in its infancy anymore (it’s a robust 25 years old).

I have other quibbles as well, but enough really. MG has written a terrific paper which makes some very useful points (e.g. I love the discussion in section 4 a lot and his discussion of Tomasello, Everett and Chater and Christiansen are excellent). The paper should be widely read and I hope that it helps change the discussion to a more reasonable one. It shoulddo this. But even if it fails to blunt the overwhelming stupidity of the common critiques, it is a very good paper for insidersto read. I suspect that nowadays many GGers do not really care for the larger cognitive biological issues that once animated the field. This makes it hard to properly rebut the many claims that GG is dead that abound in the popular press. MG’s paper is a good starting point for those interested in reclaiming the cognitive/biological roots of the GG enterprise.

That said I am going to end on a pessimistic note. Despite MG’s excellent discussion, I doubt it will much change the discussion for the reasons outlined above. We are entering a new age of Eism (Deep Learning and Big Data being conspicuous signs of this), and not just in otherareas of cognition.  Its allure is alive in linguistics as well. The idea that FL exists and has special features and that it is a proper object of linguistic study is, IMO, actually taken to be rather quaint within linguisticcircles. GGers with a cognitive bent should not only worry about the barbarians at the gate, the horse has been dragged within the city limits. Let’s hope that MG’s reasonable discussion can redirect this tide, but I am not counting on it.

Last point, I was delighted to see that a major journal published MG’s paper. I could not imagine this appearing in today’s Cognitionor LIor NLLT. Kudos to the Journal of Linguistics.  

---

[1]Of course, that said, one should always be ready to integrate useful findings from those one disagrees with, even deeply. Those grains of truth are (perhaps) worthwhile.

[2]Though who knows, maybe there really is mind stuff. The belief that there is isn’t is largely a matter of faith.

[3]Indeed, an interesting paradox, IMO, of much contemporary Minimalist work is that it is not Merge that does most of the Grammatical heavy lifting. Rather the prime grammatical operation is AGREE and the long distance feature checking that accompanies it. I-Merge is a very secondary feature of most contemporary accounts and nobody had bothered to consider how linguistically specific the properties of AGREE are. To the degree that they are not, this is a problem for the idea that onlyMerge is linguistically proprietarty. Ditto with the features of the basic lexical atoms. Their idiosyncrasies have been well discussed by Chomsky. To the degree that they remain, there is more to UG than Merge.

Gallistel rules

There is still quite a bit of skepticism in the cog-neuro community about linguistic representations and their implications for linguistically dedicated grammar specific nativist components. This skepticism is largely fuelled, IMO, by associationist-connectionist (AC) prejudices steeped in a nihilistic Empiricist brew.  Chomsky and Fodor and Gallistel have decisively debunked the relevance of AC models of cognition, but these ideas are very very very (very…) hard to dispel. It often seems as if Lila Gleitman was correct when she mooted the possibility that Empiricism is hard wired in and deeply encapsulated, thus impervious to empirical refutation. Even as we speak the default view in cog-neuro is ACish and that there is a general consensus in the cog-neuro community that the kind of representations that linguists claim to have discovered just cannot be right for the simple reason that the brain simply cannot embody them.

Gallistel and Matzel (see here) have deftly explored this unholy alliance between associationist psych and connectionist neuro that anchors the conventional wisdom. Interestingly, this anti representationalist skepticism is not restricted to the cog-neuro of language. Indeed, the Empiricist AC view of minds and brains has over the years permeated work on perception and it has generated skepticism concerning mental (visual) maps and their cog-neuro legitimacy.  This is currently quite funny for over the last several years Nobel committees have been falling all over themselves in a rush to award prizes to scientists for the discovery of neural mental maps. These awards are well deserved, no doubt, but what is curious is how long it’s taken the cog-neuro community to admit mental maps as legit hypotheses worthy of recognition.  For a long time, there was quite a bit of excellent behavioral evidence for their existence, but the combo of associationist dogma linked to Hebbian neuro made the cog-neuro community skeptical that anything like this could be so. Boy were they wrong and, in retrospect, boy was this dumb, big time dumb!

Here is a short popular paper (By Kate Jeffery) that goes over some of the relevant history. It traces the resistance to the very idea of mental maps stemming from AC preconceptions. Interestingly, the resistance was both to the behavioral evidence in favor of these (the author discusses Tolman’s work in the late 40s. Here’s a quote (5):

Tolman, however, discovered that rats were able to do things in mazes that they shouldn’t be able to do according to Behaviourism. They could figure out shortcuts and detours, for example, even if they hadn’t learned about these. How could they possibly do this? Tolman was convinced animals must have something like a map in their brains, which he called a ‘cognitive map’, otherwise their ability to discover shortcuts would make no sense. Behaviourists were skeptical. Some years later, when O’Keefe and Nadel laid out in detail why they thought the hippocampus might be Tolman’s cognitive map, scientists were still skeptical.

Why the resistance? Well ACism prevented conceiving of the possibility.  Here’s how Jeffery put it (5-6).

One of the difficulties was that nobody could imagine what a map in the brain would be like. Representing associations between simple things, such as bells and food, is one thing; but how to represent places? This seemed to require the mystical unseen internal ‘black box’ processes (thought and imagination) that Behaviourists had worked so hard to eradicate from their theories. Opponents of the cognitive map theory suggested that what place cells reveal about the brain is not a map, so much as a remarkable capacity to associate together complex sensations such as images, smells and textures, which all happen to come together at a place but aren’t in themselves spatial.

Note that the problem was not the absence of evidence for the position. Tolman presented lots of good evidence. And O’Keefe/Nadel presented more (in fact enough more to get the Nobel prize for the work). Rather the problem was that none of this made sense in an AC framework so the Tolman-O’Keefe/Nadel theory just could not be right, evidence be damned.[1]

What’s the evidence that such maps exist? It involves finding mental circuits that represent spatial metrics, allowing for the calculation of metric inferences (where something is and how it is from where you are). The two kinds of work that have been awarded Nobels involve place cells and grid cells. The former involve the coding of direction, the latter coding distance. The article does a nice job of describing what this involves, so I won’t go into it here.  Suffice it to say, that it appears that Kant (a big deal Rationalist in case you were wondering) was right on target and we now have good evidence for the existence of neural circuits that would serve as brain mechanisms for embodying Kant’s idea that space is a hard wired part of our mental/neural life. 

Ok, I cannot resist. Jeffery nicely outlines he challenge that these discoveries pose for ACism. Here’s another quote concerning grid cells (the most recent mental map Nobel here) and how badly it fits with AC dogma (8):[2]

The importance of grid cells lies in the apparently minor detail that the patches of firing (called ‘firing fields’) produced by the cells are evenly spaced. That this makes a pretty pattern is nice, but not so important in itself – what is startling is that the cell somehow ‘knows’ how far (say) 30 cm is – it must do, or it wouldn’t be able to fire in correctly spaced places. This even spacing of firing fields is something that couldn’t possibly have arisen from building up a web of stimulus associations over the life of the animal, because 30 cm (or whatever) isn’t an intrinsic property of most environments, and therefore can’t come through the senses – it must come from inside the rat, through some distance-measuring capability such as counting footsteps, or measuring the speed with which the world flows past the senses. In other words, metric information is inherent in the brain, wired into the grid cells as it were, regardless of its prior experience. This was a surprising and dramatic discovery. Studies of other animals, including humans, have revealed place, head direction and grid cells in these species too, so this seems to be a general (and thus important) phenomenon and not just a strange quirk of the lab rat.

As readers of FL know, this is a point that Gallistel and colleagues have been making for quite a while now and every day the evidence for neural mechanisms that code for spatial information per se grows stronger. Here is another very recent addition to the list, one that directly relates to the idea that dead-reckoning involves path integration. A recent Science paper (here) reports the discovery of neurons tuned to vector properties. Here’s how the abstract reports the findings:

To navigate, animals need to represent not only their own position and orientation, but also the location of their goal. Neural representations of an animal’s own position and orientation have been extensively studied. However, it is unknown how navigational goals are encoded in the brain. We recorded from hippocampal CA1 neurons of bats flying in complex trajectories toward a spatial goal. We discovered a subpopulation of neurons with angular tuning to the goal direction. Many of these neurons were tuned to an occluded goal, suggesting that goal-direction representation is memory-based. We also found cells that encoded the distance to the goal, often in conjunction with goal direction. The goal- direction and goal-distance signals make up a vectorial representation of spatial goals, suggesting a previously unrecognized neuronal mechanism for goal-directed navigation.

So, like place and distance, some brains have the wherewithal to subserve vector representations (goal direction and distance). Moreover, this information is coded by single neurons (not nets) and is available in memory representations, not merely for coding sensory input. As the paper notes, this is just the kind of circuitry relevant to “the vector-based navigation strategies described for many species, from insects to humans (14–19)— suggesting a previously unrecognized mechanism for goal-directed navigation across species” (5).

So, a whole series of neurons tuned to abstracta like place, distance, goal, angle of rotation, and magnitude that plausibly subserve the behavior that has long been noted implicates just such neural circuits. Once again, the neuroscience is finally catching up with the cognitive science. As with parents, the more neuro science matures the smarter classical cognitive science becomes.
Let me emphasize this point, one that Gallistel has forcefully made but is worth repeating at every opportunity until we can cleanly chop off the Empiricist zombie’s head. Cognitive data gets too little respect in the cog-neuro world. But in those areas where real progress has been made, we repeatedly find that the cog theories remain intact even as the neural ones change dramatically. And not only cog-neuro theories. The same holds for the relation of chemistry to physics (as Chomsky noted) and genetics to biochemistry (as Gallistel has observed). It seems that more often than not what needs changing is the substrate theory not the reduced theory. The same scenario is being repeated again in the cog-neuro world. We actually know very little about brain hardware circuitry and we should stop assuming that ACish ideas should be given default status when we consider ways of unifying cognition with neuroscience.

Consider one more interesting paper that hits a Gallistel theme, but from a slightly different angle. I noted that the Science paper found single neurons coding for abstract spatial (vectorial) information. There is another recent bit of work (here) that ran across my desk[3] that is also has a high Gallistel-Intriguing (GI) index.

It appears that slime molds can both acquire info about their environment and can pass this info on to other slime molds. What’s interesting is that these slime molds are unicellular, thus the idea that learning in slime molds amounts to fine tuning a neural net cannot be correct. Thus whatever learning is in this case must be intra, not inter-neural.  And this supports the idea that one has intra cellular cognitive computations. Furthermore, when slime molds “fuse” (which they apparently can do, and do do) the information that an informed slime mold has can transfer to its fused partner. This supports the idea that learning can be a function of the changed internal state of a uni-cellular organism.
This is clearly grist for the Gallistel-King conjecture (see here for some discussion) that (some) learning is neuron, not net, based. The arguments that Gallistel has given over the years for this view have been both subtle, abstract and quite arm-chair (and I mean this as a compliment). It seems that as time goes by, more and more data that fits this conception comes in. As Gallistel (and Fodor and Pylyshyn as well) noted, representational accounts prefer certain kinds of computer architectures over others (Turing-von Neumann architectures). These classical computer architectures, we have been told, cannot be what brains exploit. No, brains, we are told repeatedly, use nets and computation is just the Hebb rule with information stored in the strength of the inter-neuronal connections. Moreover, this information is very ACish with abstracta at best emergent, rather than endogenous features of our neural make-up. Well, this seems to be wrong. Dead wrong. And the lesson I draw form all of this is that it will prove wrong for language as well. The sooner we dispense with ACism, the sooner we will start making some serious progress. It’s nothing but a giant impediment, and has proven to be so again and again.

---

[1] This is a good place to remind you of the difference between Empiricist and empirical. The latter is responsiveness to evidence. The former is a theory (which, IMO, given its lack of empirical standing has become little more than a dogma).

[2] It strikes me as interesting that this sequence of events reprises what took place in studies of the immune system. Early theories of antibody formation were instructionist because how could the body natively code for so many antibodies? As work progressed, Nobel prizes streamed to those that challenged this view and proposed selectionist theories wherein the environment selected from a pre-specified innately generated list of options (see here). It seems that the less we know, the greater the appeal of environmental conceptions of the origin of structure (Empiricism being the poster child for this kind of thinking). As we come to know more, we come to understand how rich is the contribution of the internal structure of the animal to the problem at hand. Selectionism and Rationalism go hand in hand. And this appears to be true for both investigations of the body and the mind.

[3] Actually, Bill Idsardi feeds me lots of this, so thx Bill.