Cognitive models of reasoning

Approaches that endorse the causal nature of representation have difficulties in explaining the selectivity that any living organism shows in its behaviour.  Not every represented feature in the environment is relevant. The organism’s (construed) goal seems to affect the ways by it handles incoming information. Worse still, such information seems to be interpreted in relation to the goal and the organism’s experiences of the conditions for attaining it. This is a long known problem. In cognitive science philosophy it is discussed as the relevance problem.  Richard Samuels has presented a recent summary of the issue and the various attempts at solving it.

Reasoning seems to be the pivotal cognitive process that transforms incoming information into action plans. In order to provide a naturalistic account of cognition, this mental activity should be modelled in terms of some algorithmic procedures. As Fodor suspected in “The mind doesn’t work that way”, and Samuels concludes in 2010, neither the classical computational theory of cognition nor the connectionist approaches have succeeded. He enlists various reasons, but to me the most interesting concerns the ways representations are understood.

Fodor opted for symbolic representations (as the vehicles for the language of thought) very early on, because the resemblance theory of representations  was completely inadequate to account for “propositional attitudes”, like beliefs or wishes about some state of affair. This conception generated the “functional role semantics” with all sorts of paradoxes and Fodor is both an honest man and enough sharp-eyed to admit the difficulties, which many of his followers (Like Susan Schneider and Richard Samuels) have not wanted to do. 

There is a marvellous debate between Steven Pinker and Jerry Fodor in 2005 about how to understand the workings of the mind. Fodor’s “The mind doesn’t work that way” was a critique of the classical computational approach, stimulated by Pinker’s book “How the mind works” (1997). It took Pinker five years to defend his evolutionary account of cognitive processes in the paper published in Mind and Language. You can retrieve it at Pinker’s home page. Fodor’s response in the same issue is, again, most exhilarating but it is sadly locked up in Wiley’s archives from which you cannot get it for free unless you have an institutional access. 

First, Fodor shows how the syntactic and “local” meaning of symbolic representations inevitably cause problems. He also shows how cognitive science tries to account for these by the massive modularity models and by calling evolutionary accounts to help (as Pinker does). Fodor's conclusion is very straightforward: “So how does the mind work? I don’t know. You don’t know. Pinker doesn’t know. And, I rather suspect, such is the current state of the art, that if God were to tell us, we wouldn’t understand him.”

Samuels admits five years later that Fodor’s view of the state of art in cognitive science is accurate. However, he does not want to explain it as the inherent (logical) flaws in the paradigm as Fodor did. He regards it mainly as an empirically solvable problem once the methodological difficulties in studying reasoning have been cleared away. Why is Samuels reluctant to admit the logical problem? He provides the answer, partly in a footnote. I am astonished how often important insights are buried in footnotes. Perhaps they are too dangerous to be inserted in the main line of argument. Here it is: “...if one seeks to provide a substantive general explanation of failures in cognitive science [to explain the problem of relevance], then the most obvious option is to reject the mechanistic assumption. [The footnote:  this is the position that Descartes (1637) famously adopts in the Discourse on method; and for surprisingly similar reasons. Roughly, he thinks that it is impossible to provide a mechanistic account of reason]. 

If our only alternative to computational,  connectionist, dynamic systems etc. models of cognitive processes is to revert back to Cartesian dualism, it is obvious that we must go on trying to fit these models to the mind and hope that improved ways of doing so will eventually get us out of the danger. I do not think we have to renounce a monistic account of the mind. We only have to renounce the representationist accounts of the “tokens in the brain”.

Farewell to Fodor?

I think I have now reached the point at which I can leave behind Jerry Fodor’s long-standing attempt at understanding the causal chain from representations to their symbolic expression. When I sensed the Aristotelian smell in the way Fodor wrote about perceptual modules, I was not yet aware that it was just an offspring of his grand project of overcoming the Aristotelian rift between incoming information and its transformation into language. It is really impressive regarding its scope, its time span, and volume. And yet, in the end, it seems to fail. The source of the failure lurks in the very beginning, in the conception of “Language of thought” , which Fodor launched in a book (1975) with this title. 

If we start by assuming that mental representations are language-like, we are free to look at the complex permutations, combinations and transformations that thinking seems to accomplish with such “symbolic representations”. But how does this coincide with the causal theory of mental representations?  

The most succinct formulation I have found by Fodor on this issue is from an exhilarating paper, jointly written with LePore in 1993, that shows the inconsistencies of inferential role semantics (endorsed by post-structuralists and constructionists of the time).

 “..what makes 'dog' mean dog is some sort of symbol-world connection; perhaps some sort of causal or informational or nomological connection between tokens of the expression and tokens of the animal.”

There are two important things to notice here. First, two sets of “tokens” [presumably in the brain] are postulated. One for symbols and the other for the object to which the symbol refers. The second thing  concerns the nature of the relationship between the two sets. I have got the impression that Fodor emphasised causal nature of the object and the token [object representation?] earlier in his writings. This is not addressed here. Instead, the relationship between the two token sets are characterised by three possible alternatives. So, symbol tokens inherit their meaning content from the [causally] tokened object representations. Such mental symbols can combined, permuted, etc., in the language of thought and, eventually, also be expressed in natural languages.

But how is this “causal or informational or nomological” relationship implemented (in the brain, presumably)? How is the first set of tokens transformed into the second set? Do we not have here exactly the problem of how meaning is attached to representations, that Daniel Stern’s homemade attempt in Motherhood Constellation addressed in a small scale? Fodor’s edifice is obviously much more extensive but nevertheless it runs into the same problems.

It interesting to notice that Fodor’s work is refuted from two sides. Both the representatives of semantic theories, like Jylkkä (2011), and strict cognitive science philosophers, as Cummins, show that his functionalist account entails forbidden amalgamations. On one hand, you cannot do away the inferential nature of symbols, concepts, and their verbal expressions. On the other hand, by assuming “symbolic representations” you destroy the idea of representation as it is used in cognitive science by overextending its meaning. It seems that both camps are justified in their arguments.

In order to accomplish a unitary chain from perception to expression, mediated by a central process, Fodor had to invent a number of bridging concepts. I felt awkward when encountering the terms “symbolic representation”, “language of thought” or LOT, and “tokens”. Reading commentaries and reviews did not help much. Consider the following summary of LOT:

“…cognitive processes are the causal sequencing of tokenings of symbols in the brain… LOT also claims that mental representations have a combinatorial syntax. A representational system has a combinatorial syntax just in case it employs a finite store of atomic representations that may be combined to form compound representations… Relatedly, LOT holds that symbols have a compositional semantics—the meaning of compound representations is a function of the meaning of the atomic symbols, together with the grammar. ” (Schneider & Katz, 2011).

Will I ever get out of such a conceptual mess? Or should I just leave it behind to rest in the peace of countless journal articles, hidden in the archives of publishing houses that charge 30 $ for an access.

It seems that different strands of cognitive science philosophy work with quite different conceptions of representation. The computational approaches to sense perception (illustrated by David Marr’s early research) understand representations as an end product of some basic algorithmic procedures on the sense data, like retinal images.  Symbolic representations, in turn, are products of algorithmic procedures with “atomic symbols”  - if I can make any sense of the above quotation.  

 
But the same problem that can be recognised in Fodor & LePore’s formulation is present. How do we get from the representations that are constructed (computed) out of sense data to the symbolic representations (“atomic symbols”)? We cannot, as Cummins and Roth argue in a recent article.  There is a rift between the causal sequences that operate on “incoming forms"  and the “symbol manipulating operations” performed by the “central processor”.  Perceptual representations are not symbolic.  In my next post, I will summarise this line of arguments. Then I shall explore a bit closer what Schneider and Katz mean by “atomic symbols”.

Resurrection of resemblance theory

Has the Aristotelian “resemblance theory” of representation been abandoned, when we embark on a computational analysis of cognitive processes?  This question came to my mind when reading Geir Kirkebøen’s excellent article on “Descartes’ psychology of vision and cognitive science” (1998). I stumbled upon it after Fodor’s modularity book had sent me to reading texts on computational neuroscience and I came across David Marr’s influential book on vision (1982). It appeared posthumously after his death of leukaemia only at the age of 35. A reprint of the book has been issued quite recently. There are several good reviews of Marr’s conceptualisations.  Kirkebøen traces the historical antecedents of Marr’s view and demonstrates its similarity with Descartes’ theory of vision. This article, once again, demonstrates the immeasurable value of a historical approach to the theoretical issues of psychology. If you can get hold of this article, do read it.

Kirkebøen argues that Marr’s computational understanding is a restatement of Descartes’ solution to the problem, how  res extensa can become res cogitas. Descartes showed that extension (a “picture” on the retina) is transformed into mechanical movements in the optical nerve by a computational process that models two-dimensional objects like his analytical geometry modelled Euclidean forms.  This seems indeed to be the end of the “resemblance theory” of representations, as Fodor pointed out.

However, the idea that our visual perception uses algorithmic procedures to generate three-dimensional perceptibles from two-dimensional sense data does not abolish the problem of the relationship between the source and its retinal imprint. Oron Shagrir’s recent (2010) commentary on Marr’s theory is helpful here. The following quotations are terminologically difficult for us who are not trained in the language of computational neuroscience, but I believe you can recognise the point:

“In the case of edge detection, Marr (1982, 68ff.) refers to the constraint of spatial localization, which means (in this context) that the things in the world that give rise to intensity changes are spatially localized… Another pertinent physical fact is that intensity changes in the [retinal] image result from “surface discontinuities or from reflectance or illumination boundaries.” (Shagrir, 2010, p. 488)

“Marr’s explanation appeals to similarity between the internal mapping relations and external relations between the features that are being represented. The similarity is not at  the level of physical properties. After all, the physiological properties of the brain are quite different from the physical and optical properties that make up our visual field. The similarity is at a more abstract level of mathematical properties.” (p. 489)

“Thus, Marr not only demonstrates that the internal mathematical function correlates with the contingent world that we live in. He also underscores the basis of this correlation, which is a similarity of mathematical structures. This mathematically based similarity, I maintain, is the key in addressing the appropriateness problem…” (p. 489)

Shagrir does not refer to Kirkebøen’s paper and is perhaps not familiar with it. The second quotation, however, reproduces almost literally the Cartesian solution to the difference between the optic imprint and the physiological properties  of the brain. The similarity is established by mathematical modelling. To me this is a restatement of the resemblance theory. It is still about “form without matter” even if the form is not any more understood in pictorial terms. Shagrir also restates Aristotle's idea that vision is about distinguishing light from darkness.

Fodor, Aristotle, and the modular mind

I mentioned some time ago, when discussing Daniel Stern’s attempt at tracing meaningful interpersonal interactions, that there is a convenient way to bypass the knotty problem of how representations - or what ”comes in” - obtain meanings that can be expressed by referential, or symbolic, means. You just ignore the divide and concentrate on either side. Because Stern did not want to do this, he crashed into the problem. His effort to solve it demonstrate that a number representational modes must be assumed and, specifically, one that allows for the transition. Stern called it “proto-narrative envelope”. To me, this construction is not entirely convincing.

Disregarding the number and nature of representational modes that make up our experience of the world, the other problem is the number of mental faculties that must be postulated  to account for the incoming information. Stimulated by Ibn Sina’s Medieval account of the five faculties inthe perceptual soul, I left the issue of meaning for a while and began reading Jerry Fodor’s and Peter Carruthers’ views of mind modules. What an interesting world of thought it is. As a psychotherapist being inspired by Lev Vygotsky, Valentin Voloshinov, Mikhail Bakhtin, and Donald Winnicott, I left behind the cognitive approaches to the mind in the early 1970s. Hence, the more philosophical discourse on  the computational models of the mind is a foreign territory to me and I have to proceed cautiously.

Both Fodor and Carruthers subscribe to the assumption on multiple functions, or modules, that mediate between the flow of sensory input and the mind. Their main difference seems to be in the level and extent of such modules. Fodor claims that the central (computational) processes are non-modular, while Carruthers supports a “massive modularity” theory of the mind. Both share the view that mental modules are innate, hard-wired systems that operate according to their inner rules.

Interestingly, when Fodor published his seminal book “The modularity of mind” (1983), he added “An essay on faculty psychology” as the subheading. Indeed, within the Aristotelian framework of how the mind relates to the world there is logically no alternative.  Although written almost thirty years ago, Fodor’s book is a relevant work that introduces the modern version of the ancient model of understanding the soul’s operations. Here is an illustration how Fodor approaches the issue of incoming forms:

Any mechanism whose states covary with environmental ones can be thought of as registering information about the world; and, given the satisfaction of certain further conditions, the output of such systems can reasonably be thought of as representations of the environmental states with which they covary … But if cognitive processors are computational systems, they have access to such information solely in virtue of the form of the representations in which it is couched. Computational processes are, by definition, syntactic; a device which makes information available to such processes is therefore responsible for it format as well as its quality… if we think of the perceptual mechanisms as analogous to such devices, then we are saying that what perception must do is to so represent the world as to make it accessible to thought. ” (Fodor: The modularity of mind, pp. 39-40, original italics).

It is not easy to unpack this statement. In fact, the whole book is an elaboration of this basic formulation. For the time being, I only pay attention to the first and the last sentences in the passage. Aristotle’s view of the co-varying mechanism covered the sense organs and the internal common sense, whose joint operation produced phantasmata as the output, in a mode that was accessible to thought. This analogue does not do justice to Fodor because, elsewhere, he explicitly rejects the “resemblance theory” of mental representations. 

True, since Descartes representations could not anymore be conceived of as “forms without matter”. The division between “res extensa” and “res cogitans” implied a radical transformation from the material to the ideal mode of existence. The concept of representation has indeed changed a lot. However, the notion of sensory processes that  covary with external states, register information about the world, and present this information to thought, faithfully repeats the ancient model of the mind.