Revisiting Gallistel's conjecture

I recently received two papers that explore Gallistel’s conjecture (see here for one discussion) concerning the locus of neuronal computation. The first (here) is a short paper that summarizes Randy’s arguments and suggests a novel view of synaptic plasticity. The second (here: http://www.nature.com/nature/journal/v538/n7626/full/nature20101.html)[1] accept Randy’s primary criticism of neural nets and couples a neural net architecture with a pretty standard external memory system. Let me say a word about each.

The first paper is by Patrick Trettenbrein (PT) and it appears in Frontiers in Systems Neuroscience. It does three things.

First, it reviews the evidence against the idea that brains store information in their “connectivity profiles” (2). This is the classical assumption that inter-neural connection strengths are the locus of information storage. The neurophysiological mechanisms for this are long term potentiation (LTP) and long term depression (LTD). LTP/D are the technical terms for whatever strengthens or weakens interneuron connections/linkages. I’ve discussed Gallistel and Matzel’s (G&M) critique of the LTP/D mechanisms before (see here). PT reviews these again and emphasizes G&M’s point that there is an intimate connection between this Hebbian “fire together wire together” LTP/D based conception of memory and associationist psychology. As PT puts it: “Crucially, it is only against this background of association learning that LTP and LTD seem to provide a neurobiologically as well as psychologically plausible mechanism for learning and memory” (88). This is why if you reject associationsim and endorse “classical cognitive science” and its “information processing approach to the study of the mind/brain” you will be inclined to find contemporary connectionist conceptions of the brain wanting (3).

Second, there is recent evidence that connection strength cannot be the whole story. PT reviews the main evidence. It revolves around retaining memory traces despite very significant alterations in connectivity profiles. So, for example, “memories appear to persist in cell bodies and can be restored after synapses have been eliminated” (3), which would be odd if memories lived in the synaptic connections. Similarly it has recently been shown that “changes in synaptic strength are not directly related to storage of new information in memory” (3). Finally, and I like this one the best (PT describes it as “the most challenging to the idea that the synapse is the locus of memory in the brain”), PT quotes a 2015 paper by Bizzi and Ajemian which makes the following point:

If we believe that memories are made of patterns of synaptic connections sculpted by experience, and if we know, behaviorally, that motor memories last a lifetime, then how can we explain the fact that individual synaptic spines are constantly turning over and that aggregate synaptic strengths are constantly fluctuating?

Third, PT offers a reconceptualization of the role these neural connections. Here’s an extended quote (5):

…it occurs to me that we should seriously consider the possibility that the observable changes in synaptic weights and connectivity might not so much constitute the very basis of learning as they are the result of learning.

This is to say that once we accept the conjecture of Gallistel and collaborators that the study of learning can and should be separated from the study of memory to a certain extent, we can reinterpret synaptic plasticity as the brain's way of ensuring a connectivity and activity pattern that is efficient and appropriate to environmental and internal requirements within physical and developmental constraints. Consequently, synaptic plasticity might be understood as a means of regulating behavior (i.e., activity and connectivity patterns) only after learning has already occurred. In other words, synaptic weights and connections are altered after relevant information has already been extracted from the environment and stored in memory.

This leaves a place for connectivity, but not as the mechanism of memory but as what allows memories to be efficiently exploited.[2] Memories live within the cell but putting these to good use requires connections to other parts of the brain where other cells store other memories. That’s the basic idea. Or as PT puts it (6):

The role of synaptic plasticity thus changes from providing the fundamental memory mechanism to providing the brain’s way of ensuring that its wiring diagram enables it to operate efficiently…

As PT notes, the Gallistel conjecture and his tentative proposal are speculative as theories of the relevant cell internal mechanisms don’t currently exist. That said, neuroiphsyiological (and computational, see below) evidence against the classical Hebbian view are mounting and the serious problems for storing memories in usable form in connections strengths (the bases of Gallistel’s critique) are becoming more and more well recognized.

This brings us to the second Nature paper noted above. It endorses the Gallistel critique of neural nets and recognizes that neural net architectures are poor ways of encoding memories. It adds a conventional RAM to a neural net and this combination allows the machine to “represent and manipulate complex data structures.”

Artificial neural networks are remarkably adept at sensory processing, sequence learning and reinforcement learning, but are limited in their ability to represent variables and data structures and to store data over long timescales, owing to the lack of an external memory. Here we introduce a machine learning model called a differentiable neural computer (DNC), which consists of a neural network that can read from and write to an external memory matrix, analogous to the random-access memory in a conventional computer. Like a conventional computer, it can use its memory to represent and manipulate complex data structures, but, like a neural network, it can learn to do so from data.

Note that the system is still “associationist” in that learning is largely data driven (and as such will necessarily run into PoS problems when applied to any interesting cognitive domain like language) but it at least recognizes that neural nets are not good for storing information. This latter is Randy’s point. The paper is significant for it comes from Google’s Deep Mind Project and this means that Randy’s general observations are making intellectual inroads with important groups. Good.

However, this said, these models are not cognitively realistic for they still don’t make room for the domain specific knowledge that we know characterizes (and structures) different domains. The main problem remains the associationism that the Google model puts at the center of the system. As we know that associationism is wrong and that real brains characterize knowledge independently of the “input,” we can be sure that this hybrid model will need serious revision if intended as a good cog-neuro model.

Let me put this another way. Classical cog sci rests on the assumption that representations are central to understanding cognition. Fodor and Pylyshyn and Marcus long ago agued convincingly that connectionism did not successfully accommodate representations (and, recall, that connectionist agreed that their theories dumped representations) and that this was a serious problem for connectionist/neural net architectures. Gallistel further argued that neural nets were poor models of the brain (i.e. and not only of the mind) because they embody a wrong concpetion of memory; one that that makes it hard to read/write/retrieve complex information (data structures) in usable form. This, Gallistel noted, starkly contrasts with more classical architectures. The combined Fodor-Pylyshyn-Marcus-Gallistel critique then is that connectionist/neural net theories were a wrong turn because they effectively eschewed representations and that this is a problem both from the cognitive and the neuro perspective. The Google Nature paper effectively concedes this point, recognizes that representations (i.e. “complex data structures) are critical  and resolves the problem by adding a classical RAM to a connectionist front end.

However, there is a second feature of most connectionist approaches that is also wrong. Most such architectures are associationist. They embody the idea that brains are entirely structured by the properties of the inputs to the system. As PT puts it (2):

Associationism has come in different flavors since the days of Skinner, but they all share the fundamental aversion toward internally adding structure to contingencies in the world (Gallistel and Matzel 2013).

Yes! Connectionists are weirdly attracted to associationism as well as rejecting representations. This is probably not that surprising. Once on thinks of representations then it quickly becomes clear that many of their properties are not reducible to statistical properties of the inputs. Representations have formal properties above and beyond what one finds in the input, which, once you look, are found to be causally efficacious. However, strictly speaking associationsim and anti-representationalism are independent dimensions. What makes Behaviorists distinctive among Empiricists is their rejection of representations. What unifies all Empiricists is their endorsement of associationism. Seen form this perspective, Gallistel and Fodor and Pylyshyn and Marcus have been arguing that representations are critical. The Google paper agrees. This still leaves associationism however, and position the Googlers embrace.[3]

So is this a step forward? Yes. It would be a big step forward if the information processing/representational model of the mind/brain became the accepted view of things, especially in the brain sciences. We could then concentrate (yet again) all of our fire on pernicious Empiricism so many Cog-neuro types embrace.[4] But, little steps my friends, little steps. This is a victory of sorts. Better to be arguing against Locke and Hume than Skinner![5]

That’s it. Take a look.

---

[1] Thx to Chris Dyer for bringing the paper to my attention. I put in the URL up rather than link to the paper directly as the linking did not seem to work. Sorry.

[2] Redolent of a competence/performance distinction, isn’t it?  The physiological bases of memory should not be confused with the physical bases for the deployment of memory.

[3] I should add that it is not clear that the Googlers care much about the cog-neuro issues. Their concerns are largely technological, it seems to me. They live in a Big Data world, not one where PoS problems (are thought to) abound. IMO, even in a uuuuuuge data environment, PoS issues will arise, though finding them will take more cleverness. At any rate, my remarks apply to the Google model as if intended as a cog-neuro one.

[4] And remember, as Gallistel notes (and PT emphasizes) much of the connectionism one sees in the brain sciences rests on thinking that the physiology has a natural associationist interpretation psychologically. So, if we knock out one strut, the other may be easier to dislodge as well (I know that this is wishful thinking btw).

[5] As usual, my thinking on these issues was provoked by some comments by Bob Berwick. Thx.

The concise Gallistel on how brains compute

Jeff Lidz sent me this great little piece by Randy Gallistel on his favorite theme: how most neuroscientists have misunderstood how brains compute. I’ve discussed Randy’s stuff in various FoL posts (here, here, and here). Here in just four lucid pages, Randy makes his main point again. If he is right (and the form of his argument seems impeccable to me), then much of what goes on in neuroscience is just plain wrong. Indeed, if Randy is right, then current neo-connectionist/neural net assumptions about the brain are about as accurate as 1950s-60s behaviorist conceptions were about the mind. In other words, at best of tertiary interest and, more likely, deserving to be completely forgotten.[1] At any rate, Randy here makes four main points.

First, that there is recent evidence (discussed here) strongly pointing to the conclusion that information can be stored inside a single neuron (rather than in connections of many neurons).

Second, that there is scads of behavioral evidence showing that brains store number values and that there is no way of storing numbers this in connection weights, thus implying that any theory of the brain that limits itself to this kind of hardware must be at best incomplete and at worst wrong.

Third, that there is a close connection between neural net “plasticity” conceptions of the brain and traditional empiricist conceptions of the mind (especially learning). In fact, Randy argues that these are largely flip sides of the same coin.

Fourth, that brains already contain all the hardware that is required to function like classical computers, the latter being the perfect complements for the computational cognitive theories that replaced behaviorism.

And all in four pages.

There is one argument that Randy hints at but doesn’t stress that I would like to add to his four. It is a conceptual argument. Here it is.

Whatever one thinks of cognition, it is clear that animals use large molecules like DNA and RNA for information processing. Indeed, this is now standard biological dogma. As Gallistel and King (here) illustrates, this system has all the capacities of a classical computer (addresses, read-write memory, variables, binding etc.). So here’s the conceptual argument: imagine that you had an animal with the wherewithal to classically compute hereditary information but instead of repurposing (exapting) this system for cognitive ends it developed an entirely different additional system for this purpose. In other words, it had all it needed sitting there but ignored these resources and embodied cognition in a completely different way. Does this seem plausible? Is this the way evolution typically works? Isn’t opportunism the main mover in the evolution game? And if it is, doesn’t this suggest that Randy’s conjecture must be right? In fact, wouldn’t it be weird if large chunks of cognition did not exploit that computational machinery already sitting there in DNA/RNA and other large molecules? In fact, wouldn’t the contrary assumption bear a huge burden of proof? Well, you know what I think!

Why is this not the common perception? Why is Randy’s position considered exotic? Here’s the one word answer: Empiricism! In the cog-neuro world this is the default view. There is little to empirically support this conception (see here for a review of the pas de deux between unsupported empiricism in psychology and tendentious reasoning in neural net neuroscience). Indeed, it largely flourishes when we know next to nothing about some domain of inquiry. However, it is the default conception of the mind. What Randy is pointing out (and has repeatedly pointed out and is right to point out) is that it is fatally flawed, not only as a theory of mind but also as a theory of the brain. And its flaws are conceptual as well as empirical. I can’t wait for the day that this becomes the conventional wisdom, though given the methodological dualism characteristic of the cog-neuro-sciences, I suspect that this day is not just around the corner. Too bad.

---

[1] Note that I say “deserving” of amnesia. This concedes the sad fact that neo-behaviorism is making a vigorous comeback within cognition. Yet another indication of the collapse of civilization.

Brains, grammars and hype (part 2)

In the previous post (here), I showed how Frankland and Greene identifies a role sensitive region of cortex and sub-areas within that region that are differentially sensitive to the doer and done-to roles. In other words, if correct, F&G offers a hypothesis about where roles like doer and done-to get coded. Finding a region sensitive to thematic parameters would be a useful contribution given our vast ignorance concerning the brain bases of anything (see here discussed here). Let me repeat this loudly lest I not be heard: FINDING A REGION SENSITIVE TO THEMATIC PARAMETERS WOULD BE A USEFUL CONTRIBUTION GIVEN OUR VAST IGNORANCE CONCERNING THE BRAIN BASES OF ANYTHING. However, F&G claims to do a whole lot more than this. Here I want to consider if it does do more. So the question for what follows: does F&G explain how the brain codes thematic information as it appears to claim to do?

No. Not really. The paper may have identified a region that correlates to role information but F&H’s claim that it explains how brains code such information seems to me quite overblown.[1] Here’s what I mean.

What would it mean to show how brains code such information? F&G tells us. In the abstract, it takes its discovered empirical results to support the following claim:

At a high level, these regions may function like topographically defined data registers, encoding the fluctuating values of abstract semantic variables. This functional architecture, which in key respects resembles that of a classical computer, may play a critical role in enabling humans to flexibly generate complex thoughts.

What’s this mean? Those familiar with earlier critiques of connectionism should recognize the allusions. People like Fodor and Pylyshyn, Marcus, and Gallistel argued that brains had a Turing rather than a connectionist architecture. They provided various arguments for this, including observations about the systematicity of cognition (in particular in language), which makes perfect sense if one assumes that brains embodied read/write memories with variables and valuation of variables, being key elements.  Most of the arguments provided were behavioral (though see Gallistel for more direct arguments that brains cannot be connectionist either). F&G is clearly pointing to these claims in the abstract above (indeed, Fodor and Pylyshyn, Marcus and Pinker are noted in the bibliography in relation to this). So, F&G clearly intends its results to be an argument in favor of Turing architectures and a challenge for connectionist architectures. However, if this is the intent, I don’t see that F&G’s argument adds anything to the earlier behavioral arguments. Why not?

F&G notes that its results are consistent with Turing architectures, but then so are most connectionist models so far as I can tell. There is nothing in these models that prevents the hidden layers (appropriately tuned) from isolating doer and done-to roles. Indeed, this is regularly done in such models for other abstract categories. So, if F&G intends to use its results to argue for classical architectures, then it is unclear to me what it has actually added to the arguments advanced by Fodor & Pylyshyn, Marcus or Gallistel. Note, I have nothing against the conclusion that connectionist architectures are bad neural models (less coyly: I am pretty confident that connectionist architectures suck). What I don’t see is that F&G adds anything to the previous arguments. I would go further (as you probably knew I would). The concluding discussion section of F&G notes that there is a “class of models that use matrix operations to combine spatially distributed representations into conjunctive representations…that could potentially be augmented…[to] encode conjunctive representations for distinct semantic roles” (11737). For the uninitiated, this is connectionist speak. In other words, as F&G notes, its results do not argue against a connectionist conception in favor of a more classical Turing view. Or more correctly, the F&G results do not add anything to the earlier (completely compelling arguments) arguments. So, if F&G intends its “how” contribution to consist in an argument for a classical architecture and against a connectionist one, then, by its own admission, it fails.[2]

What else could the “how” mean? Another possible contrast is between the kinds of codes the brain uses to track information; in particular does the brain use a place code or a rate code to track doers and done-tos. Let me expand a bit.

One line of thinking (that F&G says its results endorse) exploits geography to code information: “functional segregation corresponding to spatial segregation” and binding of variables to values executed by bringing the two into spatial proximity. This contrasts with another view wherein binding is signaled through temporal proximity (synchronization) rather than spatial. F&G claims that its results (my emphasis)

…suggest that such temporal correlations may be unnecessary in this case because the bindings may instead be encoded through the instantiation of distributed patterns of activity in spatially dissociable patches of cortex devoted to representing distinct semantic variables” (11736). 

However as the paper notes, and the highlighted mealy-mouthed modals indicate, this conclusion is not particularly well supported by their experiments. Or, more correctly, F&G’s tools preclude a strong choice between the two.  As F&G notes, the hunt was conducted using fMRI and because these have limited temporal resolution (on the order of 1000 ms) fMRI probes cannot generally “see” rate codes. The best that F&G can conclude is that because it was able to localize roles in geographically proximate yet distinct locals this suggests that a place coding of roles might be right, though not to the exclusion of rate codes. The logic is that place codes require segregated (proximate?) geography and this was found. Hence the finding supports the claim that for role information the brain uses a place code. But this conclusion does not follow. To establish it firmly one needs the inverse: if segregated regions then place code. But this is not obviously true. Moreover, and here I am asking, do neuro people believe that anytime they can localize functions in different (nearby) places that this is evidence for place codes? Sounds wrong to me, but, hey, I don’t do this.[3]

I should add that the second experiment is the crucial one for this conclusion, and it is less robust than the first as F&G notes. The bifurcation of lmSTC into doer and done-to areas is quite subtle empirically and some of the participants in the UMD discussion thought that the data here was quite brittle. Again, this is beyond my pay grade.

F&G, then, really says very little (if anything) about the how question. In fact, it never really addresses it except tangentially. “Where?,” not “how?”, is what F&G addresses.  Let me squawk about this for a moment.

IMO, neuro types often confuse how does X work with where is X located. Why they think answering one answers the other I do not know. I don’t object to the claim that knowing where things are in the brain might be/is likely to be a good first step in figuring out how the brain does what it does. But reading F&G (and this paper is hardly unique) leads me to think that CNers can’t tell the difference between where and how.  And this is a problem.

One consequence of the confusion is that it denigrates the cognitive work that it presupposes. F&G relies on an unanalyzed conception of thematic roles. In fact, it relies on a truism: that sentences like John saw Mary do not mean the same as Mary saw John and that the difference has something to do with the fact that what sentences say about John/Mary in the first sentence is effectively reverses what the second sentence says about them.  This is a truism, or as close to one as might be imagined.  However, as any linguist knows, there are many different theories to explain how this truism is true. Some exploit theta roles, some grammatical roles, some the internal/external distinction, some first vs second merge, some predicate argument structure with 1^st and 2^nd argument positions of a predicate, some Deep Structures, some kernel sentences, etc. When a linguist asks how is thematic information represented, s/he means how can we distinguish between these apparently different conceptions all of which code/represent the observed doer/done-to difference. F&G cannot tell us which of these is right, nor does it intend to. This “how?” question is beyond the technical reach of current neuro apparatus.  That’s not a criticism. Here is the criticism: by confusing where with how, F&G continues the tradition of treating distinctions beyond the range of its probes as non-questions, rather than as questions beyond the resolution of its methods. The fact is that cognitive probes into the structure of brains is right now far more powerful than the currently most fashionable technology in neuro-science. fMRI might generate pretty pictures, but it's a pretty coarse technology. Right now, behavioral methods generally allow us to probe brain structure in a far more refined way than neuro methods do. That CN technology cannot usefully probe well motivated behaviorally based claims is what we should expect, and is what we find.

A second feature of the where/how confusion is that it leads one to abstract away from the most serious question in the neuro-sciences. Call it Gallistel’s question: how do brains embody mental constructs?  For example, how does wetware code for a variable or a value thereof? How do brains read and write to memory, bind a variable, distinguish between types and tokens?  Nobody knows. In fact, as Gallistel has observed, most CNers don’t even understand that this is the “how?” question that needs addressing (see here for discussion). The cognitive literature, including that in linguistics, has shown that we need these notions. Much of current neuroscience assumes that brain architectures that cannot do any of this (indeed that apparently deny, if Gallistel is right, that brains ever do this) are serviceable. This is partly abetted by the fact that current thinking fails to distinguish where from how. F&G is another example of this wider confusion.

I could go on, but I won’t. F&G makes a contribution: it identifies one possible place for where role information in some sense (however it is represented and whether it is specifically linguistic or not) might live. Given the current state of neuroscience, this is not nothing. However, the paper’s rhetoric (BS really) is way over the top. The introduction and conclusion motivate the investigation by pointing to really big issues (in particular recursion and Turing architecture). It purports to address these issues but in truth it can’t. The results are neutral wrt them. In the process, F&G sows lots of confusion and makes lots of simple errors thereby makind it hard to find the useful kernel in the morass. This leads me to one final observation.

I have heard it argued that without the overstatement and the BS the paper could never have been published. This is sometimes said in apparent justification of the BS and hype. If so neuroscience is in really bad shape. Moreover, I am skeptical that the hype is necessary, though I am sure that even if it is, it is odious to sling it nonetheless. Let me vent.

First, I doubt that a more measured presentation would have prevented publication. The result is not trivial and could have been presented as relevant to finding where linguistically/conceptually important concepts live in brain tissue.

Second, wanting to get published is no excuse for BS. This is not show business. BS goes against the fundamental values of the scientific enterprise and should not be tolerated, even if it might be useful career-wise.[4] The big problem is that such BS is fast becoming part of standard practice.  And like all S it greases a slippery slope: BS facilitates publication, we become more indulgent towards it and this will serve to further BSify research and publication. There is no excuse for this, or at least not one that should pass the smell test (and BS does smell). Whatever, F&G has told us about brains, it is mired in overstatement and self promotion. That’s the main reason many have reacted so strongly, and rightly so.[5] And that’s too bad because F&G does have something to tell us of interest.

---

[1] Steve Pinker’s tweet highlights these F&G ambitions as well. It reads: “The most important paper in cognitive neuroscience in many years: How does the brain represent who did what to whom.” Note the “how.” I wonder if the tweet would have had the same impact if we replaced ‘how’ with ‘where.’ I can’t tell, though I think that the howish version sounds far more interesting. And this is exactly the problem.

[2] In the discussion section, F&G observes relations between its results and some previous findings in the literature. An interesting one relates to deficit studies that identify insult to the lmSTC results in “who did what to whom” problems for stimuli presented aurally and visually. This suggests the possibility, as F&G note, that this area is not linguistically dedicated. In other words, this area might be part of an “amodal language of thought.” If this is so, it might be interesting to see if analogous areas in non-linguistically endowed animals can similarly discriminate doers from done-tos.  This might even have some interesting linguistic significance concerning the theoretical utility of theta roles as discussed here. F&G leaves the linguistic status of lmSTC for future research. Hope it gets done.

[3] Also, how important is the proximity? Say that doers were found in one area and done-tos were found several sulci away. Would this be a problem for place codes? I don’t know. At any rate, the relation between being localizable and being place coded strikes me as looser than F&G suggests. In fact, I could imagine that even were rate codes employed to code some functional feature the sources generating the relevant rates might nonetheless localize somewhat.  I don’t know that this is so, but nothing F&G says leads me to think that this is impossible or even false. So a question to cognoscenti: is this inference from localizable to place code legit?

[4] IMO, BS is the most corrosive feature of much current research. As Frankfurt has argued, it might be even worse than lying for unlike the latter it has no regard for truth whatsoever. Stan Dehaene was the editor for the paper and he should really have removed this BS from the paper. He knows better.

[5] BTW, F&G does not get its BS right either. See the box marked “Significance” on the first page of the paper. It suggests that the problem of theta roles is the same as the problem of recursion. This is false. The roles that F&G addresses have nothing to do with Humboldt’s making infinite use of finite means. Here we have a finite set of possible sentences templatically specifiable wrt roles of two arguments. Recursion gives you sentences with many doers and many done tos, in fact unboundedly many. F&G has nothing to say about where the brain codes this.