![Sample our Behavioral Sciences journals, sign in here to start your access, latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/110801/TOC-Subject-Sample-2021-Behavioral-Sciences.jpg"})
Abstract
Despite its prominent role in cognitive psychology, its relevance for the research of consciousness, and some helpful clarification (e.g., Revonsuo 1999), the binding problem is still surrounded by considerable confusion. In this paper, I first give an informal but systematic overview on the diversity of forms the binding problem can assume, and then attempt to extract, on the basis of “working definitions” of various much-discussed types of binding, a common denominator. I propose that at the heart of the binding problem lies the notion of representing an entity as having a certain property, and discuss several objections that could be raised against the proposed analysis, as well its usefulness and implications.
Notes
Notes
[1] For convenience, I will continue to use the familiar singular form (‘the binding problem’), even when talking of more than one. And even where the plurality matters, I will often prefer to speak of different ‘versions’ of the problem, rather than of different ‘binding problems’.
[2] I take the term ‘physico-functional’ from Loar (Citation1990) (where it occurs as ‘physical-functional’), intending to use it as follows: A physico-functional description is a description that is either couched entirely in physical terms or “abstracts from physical detail” in a transparent way. ‘In a transparent way’ means here, roughly, that the description may replace some or all of the physical detail of a less abstract description by specifications of properties or relations that are at least in principle derivable from the original description, in conjunction with known physical facts (in particular, facts about physical laws). By extension, the term ‘physico-functional state’ (or ‘physico-functional property’, etc.) is here used to refer to states (or properties) that are defined by physico-functional descriptions, just like a phenomenal state is a state that is defined by a phenomenological description.
[3] That phenomenal states are (or may very well be) physico-functional states is a thesis that is typically endorsed by so-called new wave materialists, such as Loar (Citation1990, Citation1997) and Hill (Citation1997). Their opponents, on the other hand—antiphysicalists such as Chalmers (Citation2007) and Horgan and Tienson (Citation2001)—regard phenomenal and physico-functional states as distinct from each other. Hence, a formulation of a binding problem that is couched in phenomenological terms will on the latter's view be of a different problem when compared to a physico-functional formulation. For example, Bayne and Chalmers (Citation2003) distinguish rather sharply between access unity and phenomenal unity of consciousness, following Block's (Citation1995) distinction between access consciousness and phenomenal consciousness. (In addition to this distinction, and orthogonally to it, Bayne and Chalmers recognize different forms of unity of consciousness, some of which—like objectual unity—can be seen as results of successful binding. With respect to objectual unity, then, they identify as two different problems that of establishing objectual access unity and that of establishing objectual phenomenal unity.) In a seemingly more physicalistic vein, by contrast, Revonsuo (Citation1999, Citation2006) speaks of neurological, cognitive and phenomenal levels of description on which he takes different formulations of the binding problem to be situated. On the other hand, Revonsuo cites Smythies (Citation1994a; Citation1994b), whose distinction between (what might be called) a representational and a phenomenological binding problem is drawn in a decidedly anti-physicalistic spirit. A similar distinction can be found in Garson (Citation2001), on which more below.
[4] This distinction between “perceptual” and “cognitive” binding problems can certainly be seen as somewhat artificial, as Roskies herself admits (p. 7). Consider for example the problem of visual or auditory sentence comprehension: in some way or other, the meanings of individual words have to be combined to yield an interpretation of the whole sentence. If this constitutes a binding problem (and prima facie at least, it might), is it still perceptual, or already cognitive?
[5] It is common to distinguish between two ways of using the verb ‘to represent’. The first usage is that in which a representation represents its object, whereas the second usage is at play when a cognitive system is said to represent one or more entities, which it does (at least on the present assumption) by way of employing representations of those entities in its information processing. Consequently, there is also a second way of using the term ‘representation’, namely, to stand for a system's act of representing (in the second sense of ‘to represent’). Not to cause unnecessary confusion, I will try to avoid this latter usage of ‘representation’, although the second way of using the verb ‘to represent’ is often quite useful, and will in most cases be disambiguated by the context.
[6] Another possibility, especially in the light of such remarks as Treisman's above, is to regard the goal in binding as lying in some kind of phenomenal experience. Given the unclarities surrounding the topic of phenomenology, however, a primarily phenomenological conception of binding is, I think, unlikely to have very much theoretical value. Further reservations against even a partly phenomenological conception of binding are expressed in the next section.
[7] I should emphasize that what I here call ‘representational content’ is not supposed to be the same thing as what, in the literature on mental content, is discussed under the label ‘narrow content’. In particular, I do not claim that representational content is completely “in the head”, or that the question of what a given representation's object is should be seen as irrelevant for psychological explanation (cf. Stalnaker Citation1999).
[8] For a classic example, von der Malsburg cites Rosenblatt (Citation1961).
[9] Suppose that object 1 is represented by representations of the features A and B, and object 2 is represented only by a representation of the feature C, and that the only realistically possible combinations of these features are (A and B) and (B and C). Now, even if we use this “realism constraint” as a guiding rule in determining what we should regard as the content of the system's representations, there will be considerable ambiguity if representations of all three features are activated at the same time. In particular, it would be unclear whether the system is representing one object (perhaps object 1) as characterized by the features A and B and another one (object 2?) as characterized by the feature C, or whether one is represented as characterized by A and the other as characterized by B and C. There is even the possibility that the system is representing three objects, each one of which being represented as characterized by a different feature.
[10] Cf. Treisman (Citation1996, p. 171), who uses the term ‘property binding’ for what is here (and elsewhere) called ‘feature-binding’.
[11] I am here not differentiating between those parts of an object that could be seen as objects in their own right (like the nose and mouth of a face) and those that are typically not seen as objects (like the upper and lower half of a forearm). This does not mean, however, that “perceptual grouping” according to basic gestalt principles, such as good continuation, common fate etc., is also a form of part-binding.
[12] The reason why I say that relation-binding “may” involve representations of spatial relations, such as LOCATED AT 3° OF VISUAL ANGLE TO THE LEFT OF … or TILTED 85° RELATIVE TO THE MAIN AXIS OF … , lies in the possibility that some of these representations may be made superfluous by the use of spatial maps. If two objects are, by way of location-binding (on which more below), bound to certain parts of such a map, and thus (presumably) to representations of locations in visual space, the information about their relative distance will thereby be implicitly encoded and conceivably require no further representation. On the ways in which the spatial arrangement of objects is represented in cortex, see e.g. Robertson (Citation2004).
[13] Cf., e.g.: Doumas and Hummel (Citation2005); Fodor and Pylyshyn (Citation1988); Sharkey (Citation1991).
[14] Though the conception of location-binding in accordance with (B2) seems straightforward enough, it could be objected that location-binding does in fact also fit (B1): For, arguably, the relationship between an object-representation and a place-representation could itself be seen as a representation, namely, of the state of affairs that the object is located at the place in question. Up to this point, I do not disagree with the objection. Yet the question is whether the supposed representation also represents that state of affairs as involving the object and place in question, for only then would the fit with (B1) be complete. Unfortunately, the answer to this question would require an investigation into what it means for a representation to represent its object as having a certain property, which would go well beyond the scope of this paper. Another possibility for location-binding to conform with (B1) rather than (B2) would obtain if there were in fact no object-representations except for sets (or mereological sums) of somehow combined feature-representations. Location would then simply be another class of features whose representations have to be combined with the former to yield a new representation (i.e., a larger set or mereological sum of feature-representations) that then represents its object as being at a certain place. Evidently, the process of this combination would constitute an example of (B1)-like binding.
[15] My use here of “the representation R” is only meant to make clear that I am using the same variable that has already been introduced in (I). It should not be taken to suggest that a representation R already exists (since it clearly need not exist if only the first disjunct of (I) is true).
[16] The verb ‘to modify’ is here used in a broad sense, referring both to intrinsic and relational changes.
[17] Not all sceptics with regard to the binding problem are philosophers; some are cognitive scientists. To my knowledge, however, the latter do not tend to diagnose any conceptual confusions, but instead only argue that the binding problem is much easier to solve than is commonly thought (Riesenhuber and Poggio Citation1999), or that it may not actually pose itself in existing biological systems (Rensink Citation2000, p. 1483).
[18] Thus, he anticipates (to a certain degree) two important points of the previous section: He explicitly acknowledges that the “true” binding problem is a matter of representation, and moreover, his conception of binding as having to do with “the assignment of instances […] to types” anticipates (B2). Unfortunately, it is unclear to what extent his conception is based on a consideration of other types of binding than feature-binding.
[19] Here is a brief but central passage from Hardcastle's discussion of how binding might be achieved without a binding process: If we consider entire cortical regions as a single network system acting as one resonating unit, then a vehicle for uniting disparate feature-bits would be unnecessary. Instead, the higher level statistical aggregate of firing neurons would unite sets of feature primitives. (p. 262) It must be this “uniting” of “sets of feature primitives” that she takes binding to consist in. But except for the vague idea of a connection between feature primitives that is somehow based on resonance, there is not a hint of a reason why we should consider it a case of binding.
[20] Cf. the open peer commentary to their article in the same issue of BBS (in particular the contributions by Broackes, by J. Cohen, by O’Brien and Opie, by Pani, by Van Gulick, and still several others), as well as the authors’ response, in which they declare that, if ‘representation’ is understood in a sufficiently broad and behaviorally relevant sense, they “do not deny that there are representations” (p. 1017). Of the behavioral relevance of representation, there will also be something to say in the following discussion.
[21] Similar talk can, however, also be found in Revonsuo (Citation1999). Already Dennett (Citation1991, p. 257f.) takes “current neuroscientific research” to task for offering “incautious formulations of ‘the binding problem’”.
[22] Incidentally, something of a historical precedent to the present conception of binding can already be found in Heidegger's concept of synthesis (1953, §7B).
[23] Admittedly, the first possibility (that a representational constraint should entail a phenomenal one) may appear unrealistic if a state's lack of phenomenal character is considered not to be a hindrance to its having representational content. (For example, dispositional brain-states might be thought to have representational content, presumably without having any phenomenal character.) But even so, there may still be other entailments that would also obviate the need for specific phenomenological constraints in addition to the representational ones. For suppose a given phenomenological constraint P entails a more general constraint C—such as the constraint that the state in question should have any phenomenal character at all—and that the representational constraint R, in conjunction with C, entails P. There would then seem to be no need to impose P in addition to R; the weaker constraint C might well be enough. (However, even with respect to C, it would have to be asked how relevant it is for the brain's proper functioning.)
[24] For example, cf. Ashby et al. (Citation1996), Friedman et al. (Citation1995), Treisman and Schmidt (Citation1982). Illusory conjunctions are errors of feature binding, where perceived features, usually colors and shapes, are wrongly “bound together”. (Thus, a subject presented with a red square and a blue circle will under certain conditions sincerely report that he saw a red circle and a blue square.) The constraints in question would, for example, include findings about the conditions under which certain types of illusory conjunctions tend to occur.
[25] One might recall here David Marr's (Citation1982, ch. 1) distinction between the computational level of investigation and the lower ones of algorithm and implementation, and his recommendation that investigations belonging to these lower levels should be adequately informed by those belonging to the computational level.
[26] These assumptions have also been in the background (or even foreground) of much previous work on mental content; e.g.: Peacocke (Citation1989).
[27] For example, without such an account, there will be no satisfactory answer to the objection raised in note (14).
[28] As another reason, one might be tempted to add that psychologists will have to rely on such a connection when, from a subject's reports on what he perceives, they draw conclusions as to what properties he represents the perceived objects as having. However, I think it is arguable that the justifiability of such conclusions is rather a constraint on any defensible theory of representational content, than a doubtful presupposition that has to be “buttressed”. And insofar as it is such a constraint, of course, it already lends support to that connection.
