Noûs, 39(2): 197–255, June 2005.

Cognitive Architecture, Concepts, and Introspection:

An Information-Theoretic Solution to

the Problem of Phenomenal Consciousness

 

Murat  Aydede

University of Florida

Department of Philosophy

330 Griffin-Floyd Hall

P.O. Box 118545

Gainesville, FL 32611-8545

maydede@phil.ufl.edu

Güven  Güzeldere

Duke University

Department of Philosophy

201 West Duke Building

Box 90743

Durham, NC 27708

guven@duke.edu

 

ABSTRACT.  This essay is a sustained attempt to bring new light to some of the perennial problems in philosophy of mind surrounding phenomenal consciousness and introspection through developing an account of sensory and phenomenal concepts.  Building on the information-theoretic framework of Dretske (1981), we present an informational psychosemantics as it applies to what we call sensory concepts, concepts that apply, roughly, to so-called secondary qualities of objects.  We show that these concepts have a special informational character and semantic structure that closely tie them to the brain states realizing conscious qualitative experiences.  We then develop an account of introspection which exploits this special nature of sensory concepts.  The result is a new class of concepts, which, following recent terminology, we call phenomenal concepts: these concepts refer to phenomenal experience itself and are the vehicles used in introspection.  On our account, the connection between sensory and phenomenal concepts is very tight: it consists in different semantic uses of the same cognitive structures underlying the sensory concepts, such as the concept of red.  Contrary to widespread opinion, we show that information theory contains all the resources to satisfy internalist intuitions about phenomenal consciousness, while not offending externalist ones.  A consequence of this account is that it explains and predicts the so-called conceivability arguments against physicalism on the basis of the special nature of sensory and phenomenal concepts.  Thus we not only show why physicalism is not threatened by such arguments, but also demonstrate its strength in virtue of its ability to predict and explain away such arguments in a principled way.  However, we take the main contribution of this work to be what it provides in addition to a response to those conceivability arguments, namely, a substantive account of the interface between sensory and conceptual systems and the mechanisms of introspection as based on the special nature of the information flow between them.

 

1  INTRODUCTION

The current manifestation of the mind-body problem primarily centers around a deep disagreement between materialists and anti-materialists about the ontological nature of phenomenal consciousness.  Facing the charge that they lack the conceptual resources to understand the phenomenal, materialists are challenged to accept the existence of non-physical properties, or, at a minimum, admit to being completely in the dark regarding how to bridge the “explanatory gap” between the physical and the phenomenal.  Joseph Levine, in his recent book (2001), concludes his lengthy discussion of the explanatory gap with this challenge:

 

What emerges from our discussion is that the explanatory gap is intimately connected to the special nature of phenomenal concepts.  E-type materialists [exceptionalists] try to save materialism from the conceivability argument by arguing that phenomenal concepts are special in some way.  Well, I grant that, but then we have the problem of providing an explanation in physicalistic terms of that very specialness, and we don’t seem to have one.  If we could explain the explanatory gap, then either it would go away or we would just learn to live with it.  But it seems we can’t do that without a good account of phenomenal concepts, and that’s something we don’t have.  We lack both an account of phenomenal properties and phenomenal concepts.  (p. 86)

 

In this essay, we provide both.  The overarching goal of our project is to present a sustained and thorough information-theoretic argument that sheds new light on the problems surrounding phenomenal consciousness, especially the problem of the explanatory gap.

Building on the information-theoretic framework of Dretske (1981), we first develop an informational psychosemantics for what we call sensory concepts, i.e., concepts that apply, roughly, to so-called secondary qualities of objects.  We show that these concepts have a special informational and semantic character that ties them closely to the brain states realizing conscious experiences from which they are acquired.  We then develop an account of phenomenal concepts utilizing this special character of sensory concepts.  Phenomenal concepts are those concepts we use in introspecting our experiences and their qualities.  It is through these concepts that we conceive of the phenomenal character of our experiences.  On our account, sensory and phenomenal concepts turn out to share the same cognitive structures but their semantics are differently anchored.  What makes this possible is the dual informational content of these structures.  In the end, it is the special nature of phenomenal concepts that enables us to meet the anti-materialist challenge.  Further, contrary to widespread opinion, we show that informational psychosemantics contains the resources to satisfy internalist intuitions about phenomenal consciousness in a principled way, while not offending externalist ones.

Our account contributes to what appears to be a growing convergence of views, sometimes loosely grouped under the label of “perspectivalism,”[1] which in recent years have been developed in response to conceivability arguments against physicalism.[2]  In its typical versions, perspectivalism is advanced in three stages.  First, it diagnoses the puzzle involved in attempting to conceive the phenomenal in terms of the physical as a Frege puzzle — namely, as one arising from distinct but co-denotational concepts.  Second, it points out that the Frege case at hand is special in a way that marks it off from standard Frege cases, and that this specialness needs accounting for.  Third, it postulates a group of concepts, typically called “phenomenal concepts,” whose nature is said to be perspectival, a fact that is supposed to reveal what is so special about our epistemic access to the phenomenal qualities of our experiences.

We think that extant perspectivalist accounts are lacking in precisely those respects that are crucial for the acceptability of physicalist responses to the conceivability arguments.  At the heart of the debate is a set of intuitions about the conceivability of a range of scenarios (e.g., zombies, inverted spectra).  The debate centers around what these intuitions show.  The anti-physicalists want to draw a metaphysical conclusion (namely, the falsity of physicalism) from the conceivability of these scenarios.  Thus, the question of whether conceivability, which is an epistemic affair, entails metaphysical possibility becomes one of the crucial issues.  The perspectivalist materialists with whom we join forces in this paper[3] maintain that no such metaphysical conclusion follows, but they tend to restrict their claim only to those cases where the scenarios involve phenomenal consciousness as it relates to the physical world (after all conceivability seems to be our only guide to possibility).  Thus, at a global scale, many perspectivalists grant that given a complete physical description of the world and competence with common concepts that apply to macro phenomena, it is possible in principle to derive a priori all true claims about macro phenomena.  For instance, given complete physical knowledge (augmented with standard indexical information, etc.), it is not conceivable that water should boil at a temperature other than 0º C at sea level.  They also grant to anti-physicalists that facts about phenomenal consciousness (and only these facts) cannot be so derived.  But they refuse to draw any metaphysical conclusion from that.

It is the special nature of phenomenal concepts that is supposed to discharge the heavy explanatory burden the perspectivalists incur by this refusal.  Understandably, the absence of a principled and independently motivated story about phenomenal concepts that would justify taking this “exceptionalist” route has made perspectivalists vulnerable to accusations of special pleading.  On standard accounts, phenomenal concepts turn out to have just those features (simple, primitive, unanalyzable, demonstrative, etc.) that happen to make them immune to absorption into a general logical reduction pattern.  But where these features come from and what independent reasons we have to believe that they work this way are left mostly unexplained.  Phenomenal concepts are claimed to be special, but the critics of physicalism are justified in their demand that their special character needs to be independently grounded in a general and naturalistic account of concept formation and use.  That is, whatever it is about these concepts that are supposed to justify making consciousness an exception to an otherwise perfectly general pattern needs to fall out naturally from a general and independently motivated account of conscious experience and the concepts it gives rise to.

In other words, the account of phenomenal concepts must be part of a more general theory whose explanatory power should go beyond justifying the exception and saving materialism.  Levine (2001) rightly points out that perspectivalists haven’t so far produced an independently motivated general and naturalistic account with sufficient and credible detail.[4]  This work constitutes an attempt to develop such an account — a principled and independently motivated physicalist story, using the resources of information theory, about the special nature of phenomenal concepts, and their role in explaining why there is an explanatory gap of the sort not present in other phenomena.

We derive such an account from a cognitive architectural framework and general informational psychosemantics of sensory concepts together with an information-theoretic account of introspection.  Fundamentally, we agree with Gareth Evans on the philosophical utility of construing human beings as “gatherers, transmitters, and storers of information,” for “these platitudes locate perception, communication, and memory in a system — the informational system — which constitutes the substratum of our cognitive lives” (Evans, 1982: 122).  As such, our work is in the spirit of previous attempts to provide naturalized information-theoretic accounts of thought and reference (Evans 1982, Barwise and Perry 1983, Israel and Perry 1990, Fodor 1987, 1990).  But our greater debt is to Fred Dretske’s seminal work, Knowledge and the Flow of Information (1981), which first brought the resources of information theory into debates in epistemology and philosophy of mind in a ground-breaking way.[5]  As will be apparent in the coming sections, our account also provides an unexpected synthesis of otherwise quite diverse views.  For instance, it has consequences that we believe should satisfy internalists about phenomenal consciousness, while integrating intuitions that motivate both higher-order perception and higher-order thought accounts of consciousness (even though our view does not itself fall under these labels).

We ask the reader to bear with us while we lay the groundwork for our account before we come to the last section (§9) of this rather lengthy essay, where we ultimately respond to the conceivability arguments against physicalism.  In the next section (§ 2), in a somewhat reconstructed form and sometimes using different terminology, we will present the bare bones of an information-theoretic account of concept formation (from sensory experiences) based on Dretske (1981).[6]  We will be substantially modifying and developing this account in the remaining sections: in particular, in §§ 3–8, we will apply this account to sensory concepts and derive from it an account of phenomenal concepts and introspection.  In §§ 6–7, we will pay close attention to bodily sensations (especially, pain) and the sensory concepts they give rise to, because we believe that these sensations and the way we think about them form the microcosm of a more general mechanism of introspection.  Although we will discuss the conceivability arguments in the last section (§9), we submit that by that time the reader will be in a position to see where the real action is.  Indeed, we take the main contribution of this work to be what it provides in addition to a satisfying response to the conceivability arguments, namely, a substantive and detailed account of the interface between sensory and conceptual systems and the mechanisms of introspection based on the special nature of the information flow between them.

 

2    THE ARCHITECTURE OF INFORMATION FLOW IN COGNITION

Information-theoretic psychosemantics postulates an architectural distinction between sensory systems and a central cognitive system controlling the intentional behavior of the organism.  The sensory system has the job of providing information about one’s environment to the cognitive system, and normally affects behavior only indirectly, via intermediary cognitive structures.  Sensory systems hook up with the environment via transducers whose job is to transform the particular forms of energy impinging on the peripheral sensory organs into forms usable by the internal perceptual systems.  The output of transducers and much of the subsequent processing in the sensory and perceptual systems appear to be automatic and unconscious (but see below).  The output of the sensory system is a sensory representation of (some aspects of) the distal layout that is made available to the central cognitive system.

In this framework, the sensory representations are conscious only insofar as the information they contain is available to the central conceptual system, even if the information is not fully put to use.  We will also say that the information contained in sensory representations is available to the organism consciously only insofar as the organism can conceptualize this information, i.e., only insofar as the information can be used in the acquisition or deployment of the relevant concepts.[7]  Hence, we will use “having an experience” (which is generally but not necessarily a conscious affair) and “tokening a sensory representation” interchangeably.[8]  We will come back to this point later on, but for the rest of this paper we will concern ourselves with sensory representations that are conscious in this way.  Hence we will not further discuss those modular (pre- or intra-perceptual) processes whose state-transitions and outputs are not consciously accessible — that is, which do not constitute direct inputs to the central cognitive system.[9]

What are the functional determinants of this architectural distinction?  We have already touched on one: sensory representations don’t normally affect behavior directly.  It is largely the central cognitive system which controls voluntary behavior through motor systems.  So a necessary condition of a cognitive structure’s being conceptual as opposed to sensory is its executive connections to behavior.  A representation is sensory (as opposed to perceptual or conceptual — see § 2.2 below), on the other hand, only if it makes information about one’s environment (internal/bodily as well as external) available to the conceptual system for further processing, which normally also makes the representation (experience) conscious.[10]  Besides this, the most important characteristics underpinning the architectural distinction are to be found in the following five distinctions: vertical vs. horizontal information processing, sensation vs. perception, analog vs. digital encoding of information, extractable vs. non-extractable analog information, and acquisition vs. deployment of concepts.  We explain each of these in the remainder of this section.

 

2.1  Vertical vs. Horizontal Information Processing

First, sensory experiences are supposed to track changes in the environment.  In this they are (non-conceptual) representations whose primary job is to make available to their hosts temporally indexed information about the environment.  The crucial point here is that sensory experiences normally carry information about features of the environment: they are responses to environmental events.  As such their informational value is typically restricted within a time frame sufficient for the organism to act back on the environment on the basis of this information.  In short, sensory representations are normally stimulus-driven (a fortiori not directly voluntary).  We will call this aspect of information processing vertical information processing.

By contrast, central cognitive processes, such as thinking, reasoning, remembering, imagining, and daydreaming are normally horizontal forms of information processing.  By this we mean that they can, and frequently do, occur in the absence of a direct causal (i.e., vertical/informational — see below) relation with the things being thought about.  This is perhaps the most important hallmark of human intentionality.  In contrast to sensory systems, central cognitive systems harbor representational processes defined over concepts that are not directly prompted by what those concepts represent.[11]

Although all concepts can be informationally decoupled from their referents in horizontal processes, most of them can also be used vertically, so that their tokenings carry information about the (instantiation of the) property they apply to.  In this extended sense of a vertical process, experience is the necessary intermediary.[12]

In brief, conceptual representations are the kind of cognitive structures that are capable of being engaged in horizontal processing, whereas sensory representations are not.[13]

 

2.2  Sensation vs. Perception

There is a useful sense in which perception, unlike sensation, is the vertical informational process whereby objects of sensation and their sensible qualities are discriminated and recognized, i.e., categorized or classified under concepts.[14]  For most perceptual and observational concepts, this normally takes the form of recovering the information already (mostly) in the sensory array by computational processes that result in the tokening of a concept applied to the object of perception.  We see this process mainly as one of information extraction by digitalization or abstraction from a rich array of information present in analog form in the experience.  The mechanism underlying the formation of primitive sensory concepts and their vertical deployment is probably hard-wired in concept-using organisms like us.

According to this scheme, then, visual object recognition, for instance, however automatic it may be, is mostly a central process,[15] since it involves categorizing an object under a “visual concept.”  Although the process itself appears to be unconscious and automatic, many features of the output representation (like variation in light intensities, lines, edges, colors, distance, orientation, texture, relative position, etc.), apparently utilized in the extraction process, are also consciously (hence centrally/globally) available.  So perception is a central process in our sense and should be treated as a species of conception.

 

2.3  Analog vs. Digital Encoding of Information

Another determinant of the architecture, most important for our purposes, is captured by a distinction between the ways in which information is coded in the representations.  Here, we follow Dretske’s original characterization (cf. Dretske 1981: Chapter 3):

 

(i)    The most specific information a signal r carries about a source s is the information r carries about s in digital form.

(ii)  If r carries more information about s [or, about t (¹s)] in virtue of carrying this digital information about s, then this extra information is said to be carried by s in analog form.

(iii) Analog information is information nested (nomologically or analytically) in the information carried in digital form.

 

Note that according to this characterization a signal always carries information in both digital and analog form: it’s just that the most specific information is selected as digital.

The cognitive value of a sensory representation lies largely in the information about the distal layout it carries in analog form.  Its digital informational content is the most specific information it carries about this layout, which is very rich not only in detail but also in amount.  The conceptual system is mostly keyed to the information nested in this specific and rich information.  The analogy here between pictures and sensory representations will be helpful.  If we take a color picture of a cubical object, the picture will carry very rich, detailed, and determinate information about the size, texture, and orientation of the object, as well as its position relative to other objects, the illumination conditions, its determinate shades of color and their brightness across its surface, and so forth.  We can think of this very specific and detailed information as expressible by a very long conjunction.  But nested in this most specific information there will be less specific information implied by it, such as the information that the object is (just) cubical, that it has (just) six faces, that it has eight corners, that it is darkly colored, (just) colored, etc.  Normally we are interested in the analog information carried by the picture.  We may be interested merely to know that the object depicted is cubical — discarding the more specific information about its color, size, orientation, etc.  Or, depending on the situation, we may be interested only in its size or orientation.

Similarly with sensory representations.  The conceptual system mostly  exploits the analog information nested in the digital information carried by sensory representations.  In fact, part of what makes a cognitive structure a conceptual representation is the way it digitalizes the analog information contained in the sensory representations.  Concepts are those representations (subject to the above architectural constraints) whose most specific informational content is acquired from information carried (mostly) in analog form by sensory representations.  Concepts (except sensory ones — see below) are designed to selectively respond to and utilize the analog information contained in sensory representations.  So, for instance, even though we cannot sensorially represent a triangle without at the same time representing its determinate size, shape, orientation, etc., we can conceptually represent an object simply as a triangle without representing anything more specific or determinate about it.[16]  Concepts on this scheme are those structures that are acquired from sensory representations, mostly on the basis of the analog information they carry.

In this framework, the semantic content of a concept is identified with the information it carries in digital form.  The informational content of a concept, however, is not unique in the way the semantic content is supposed to be, since a vertical tokening of a concept will carry all the information nested in its digital informational content (i.e., in its semantic content).  So, for example, when you identify a geometrical shape as an isosceles triangle, your identification carries more information about the object nested in its being such a triangle, e.g., that it has (just) three sides, that it has (just) three corners, that it is (just) a geometrical shape, that it has a surface area, etc.  These separate pieces of information are all carried in analog form.

 

2.4  Extractable vs. Non-Extractable Analog Information

Both sensory and conceptual representations carry information in both analog and digital form.  But they encode analog information in fundamentally different ways.  In particular:

 

·       Whereas there is always some analog information sensory representations carry in extractable format, the (primitive) conceptual representations carry all their analog information in non-extractable form.[17]

 

A sensory representation is physically realized in such a way that its complex structure allows the analog information contained in it to be extracted by the conceptual system operating on it.  Of course, what information can be extracted from the sensory representation doesn’t depend solely on its complex informational structure; it also depends on the capabilities and the sophistication of the conceptual system and what other information is available to the system.  But, subject to these constraints, it is necessary for conceptualization that the analog information in the sensory representation is carried in a form that is extractable, and not all information carried in analog form is.

To illustrate, consider the example Dretske uses (1981: 138–39).[18]  It is possible to carry all the information encoded by a picture of a scene with a simple/primitive signal, say a buzzer system.  Suppose the buzzer is activated when and only when a camera attached to the buzzer detects the occurrence of a situation exactly like the one depicted in the picture.  As Dretske notes, computer recognition programs that rely on whole-template matching procedures approximate this kind of transition from one form of coding to another.  Both structures carry exactly the same information, both digital and analog.  However, we will say that the buzzer’s buzzing carries the analog information carried by the picture in a way that is not extractable, whereas the picture carries it in an extractable form.

This distinction needs to be developed in more detail in terms of physical constraints on the structures realizing the representations, but what is intuitively obvious — and all we need for present purposes — is that the representational format which allows for information extraction must consist in a structure complex enough to be the only source for subsequent digitalizations based on it.[19]  The activation of the buzzer, though it carries all the information carried by the picture, is structured in such a way that does not allow for digitalization of the information it carries in analog form.  Primitive conceptual representations are like the buzzer system: although their vertical tokenings carry analog information nested in their digital content, they are structured in such a way that they cannot serve as the sole basis for digitalization of this information.  This is part of the reason why primitive concepts are sometimes characterized as discrete representational structures or symbols.

 

2.5  Acquisition vs. Deployment of Concepts

Although the distinction between the acquisition and deployment of concepts is not a functional determinant of the informational architecture, it is important to keep in mind for clarificatory purposes.  Both acquisition and deployment can be vertical and horizontal in some intuitively extended sense.  So, for instance, we can acquire concepts by reading, or by being talked to, by looking at pictures, by engaging in inference to the best explanation, etc.[20]  This would be horizontal acquisition of concepts.  Also note that for the moment we are using the term “acquisition” in a way that is neutral between triggering and learning.

Now that we have the functional determinants of the architecture of information flow in cognition and the relevant distinctions,[21] we will focus on the nature of the concepts to which this architecture gives rise.

 

3  CONCEPTS AND THEIR SENSORY BASES

Among the concepts directly and immediately acquired from sensory experience are what we will call sensory concepts.  These form a special class of concepts that will be important for what follows.  Intuitively and roughly put (to be qualified in a moment), sensory concepts are those concepts whose digital informational content is also part of the digital informational content of the sensory representations from which they are acquired, so that the abstraction/digitalization distance between the concepts and these experiences is minimal.[22]

The digital informational content of sensory representations is rich along several dimensions.  We can think of these dimensions as presenting determinables such that the resolution of our sensory experiences marks the limit of their most determinate values about which we can gather sensory information.  To the extent that we can separate these dimensions, we can speak of that part of the total digital information content of an experience that belongs to one of these dimensions fixed by the modality of the experience.  So, for instance, under conditions that are optimal for color vision, seeing a ripe tomato will involve a visual experience whose total digital content contains the most specific information about the color of the tomato: it will represent the tomato as having a determinate shade of red, say, red16.  This is part of the total digital content of the visual experience containing information about the color of the object seen.  As mentioned above, we can conceive of this total digital content as being expressed by a very long conjunction detailing all the most specific information it carries.  The particular shade of color that a region in the visual field has, then, would be one of the conjuncts.[23]  Sensory concepts are those concepts that are closest (in terms of abstraction distance) along these different dimensions to the digital informational content of experiences from which they are acquired.

If the property of being red16 is a disjunctive property whose disjuncts are particular spectral reflectances, then the information the sensory representation of the tomato carries about red16 is about this disjunctive property.  Every disjunct would be a particular ratio fixed by the percentage of light that the surface of an object reflects at each of the three characteristic wavelengths determined by the response sensitivity of three retinal cone types.[24]  But whatever feature of sensory representation is responsible for carrying this information, it carries it without revealing its complex and disjunctive character.  For instance, this feature, by carrying information about a surface’s being red16, also carries the analog information that it has a spectral reflectance, or that it (just) reflects light at different wavelengths.  These are nested in the information that the surface is red16.  But these pieces of analog information cannot be recovered or extracted from the signal, i.e., from whatever feature of the sensory representation carries the color information in question.

There is, however, still some abstraction/digitalization — some loss of information — in this process.  This can be explained in terms of a distinction between concepts used in synchronic discriminatory tasks and concepts used in diachronic recognitional or identification tasks.  In fact, we typically reserve the notion of a concept for those cognitive structures involved in the latter sort of task.  Consider the tomato again.  If the conditions are appropriate, it will be possible to discriminate slight variations in the shade of red across the surface of the tomato.  But when the same shades of color are shown to us diachronically we may not be able to discriminate among them: most of the time the best we can do is identify and co-classify them as, say, dark red.  Both kinds of task involve discrimination and categorization of different color stimuli, and so, in this  sense, require conceptual capacities.  In what follows, however, when we talk about sensory concepts, we will have in mind the most specific concepts one can have as revealed by diachronic recognition tasks, which involve memory.  It is clear that the abstraction distance between sensory experiences and the sensory concepts conceived in this way is still minimal, although there is still some information lost.  Notice that in the case of color concepts this distance can be explained entirely in terms of set-theoretic notion of inclusion.  When these concepts are vertically deployed, the information they carry is disjunctive: they say something like “it is either red1 or red2 or red3 or … redn”, where n is finite and redi is the most determinate shade of red one’s visual experiences can carry information about and thus be synchronically discriminated.

It is important to note that the disjuncts here are still colors — determinate shades of red.  This is important because the abstraction process here is not based on information about the constituents of colors (whatever objective properties color experiences/con­cepts detect), which are not themselves colors.  So, for example, if color vision detects sets of individual surface spectral reflectances, color sensations don’t represent them by representing their constituent properties, say, individual reflectances or whatever further properties constitute these reflectances.  Hence, color sensations don’t represent colors as having constituent structure, or as we will say sometimes for convenience, as simple/atomic properties.[25]

Contrast this to the visual representation of shapes.  Our visual system happens to be such that we can’t visually represent a geometrical figure (in such a way that we can then recognize it as what it is, say, as a square) without simultaneously representing the lines, angles, curves, edges, and corners that, in some intuitive sense, constitute the figure.  It is important to note that these constituents are not more determinate instances of the same figure type, so that even the concept of a most determinate geometrical figure of that type will not be minimally close to the sensory base it is directly acquired from — even though these sensory bases are the sole authoritative source of acquisition for such concepts.  We will call such concepts perceptual concepts.  The information necessary and sufficient for the correct application of these concepts, whose abstraction distance is nevertheless not minimal (but shorter than what we will call below  observational concepts), is normally contained in the sensory base from which they are directly acquired.  Typical perceptual concepts in the case of vision include concepts of spatiotemporal relations, geometrical figures, and shapes.

For the sake of completeness, we can distinguish sensory and perceptual concepts from observational concepts like the concept of an apple, a robin, a tree, a lake, and a truck.  These concepts are also typically acquired from an appropriate sensory base, but they need not be, and often are not.  However, the information contained in experiences required in the correct application of these concepts is more impoverished, in the sense that it always underdetermines correct categorization.  In other words, although the information about the denotations of these concepts can be perceptually available, its delivery requires that certain channel conditions external to the sensory systems be in place.  The abstraction distance between these concepts and the sensory bases from which they may be acquired is considerably greater than in the case of sensory and perceptual concepts.  What seems to mark the difference is that (most of) the sensory information used in the acquisition and deployment of observational concepts is typically only contingently related to the objects in their extensions.

It is no accident that thought experiments involving spectrum inversion are carried out in terms of sensory bases of sensory concepts, where the property detected and denoted is represented as simple or atomic.[26]  Although we cannot conceive of inversion with respect to the properties denoted by perceptual concepts (e.g., of shapes) and their sensory bases, there is nothing preventing a differently organized cognitive system from performing this feat.  We can imagine and even construct devices that “sensorially” detect geometrical shapes (quite abstract from our cognitive point of view) by outputting simple and primitive sensory representations.  For instance, we can construct a detector that responds with a green light when it detects a square (any square) and with a red light when it detects a circle (any circle).  Suppose that all the information it uses in making its responses is lost at the final output stage.  When this device, a 2D geometrical shape detector, lights up green, its relevant state carries the information that something it is informationally connected to is square.  If it lights up red, its state carries the information that something is circular.  But even if the “sensory” outputs of the device carry these pieces of information, they are structured in such a way that there is no way to recover any information about the structural relationships holding among the internal constituents of these shapes.  Of course, these “sensory” outputs also carry information about the constituent properties (necessarily so), but only in analog form that is not extractable.  Nor is it possible to extract any topological information that obtains between these different shapes — if the device carries information about the shapes in a spatial array.  For all the device “knows,” whatever is being represented by these colored lights, it is simple and atomic.  There are no computational/formal constraints stemming from the representations themselves that would make the thought experiment of an “inverted shape” unintelligible here.  For all the device “knows,” circles could look exactly the same to it as squares do now, and vice versa.

If this device is also equipped with a central conceptual system that can acquire concepts from such “sensory” representations, the concept of a circle the device directly acquires from its “experiences” will be a sensory (as opposed to perceptual) concept in our sense.  Our concept of a circle is not sensory because the sensory representations from which it is acquired don’t carry the information that something is a circle as part of its total digital informational content so that when our conceptual system digitalizes this piece of information there is always more specific information that is lost but nevertheless available to the central cognitive system for digitalization.[27] Furthermore it is this lost information that seems to be used in the acquisition and vertical deployment of the target concept.  What prevents the abstraction distance from being minimal here is the existence of more specific but used-and-then-discarded information that is nevertheless available to the conceptual system for digitalization (which, subject to some further conditions, makes this used-but-then-discarded information contained in the experience consciously available).

In contrast to our perceptual system, the architecture of this device is such that the abstraction distance between the “sensory” and “conceptual” representations of circles and squares is minimal.  Not surprisingly, we are not such machines.  But it is important to keep in mind that there is no logical necessity in our having the perceptual and cognitive architecture that we do, including the set of particular abstraction distances it gives rise to — although there are most likely evolutionary and ecological reasons for this architecture.[28]

Another way to see what makes sensory concepts so special is to understand the nature of the abstraction distance between them and their sensory bases.  As we have said, this distance is minimal (subject to the qualification we have just introduced), which is what marks these concepts off from the rest.  Following Fodor (1990) and Margolis (1998), we will call the mechanisms that mediate the informa­tional link between the vertical tokenings of a concept and the instantiations of the property it applies to sustaining mechanisms.[29]  The intra-cranial portion of the sustaining mechanisms for sensory concepts is not cognitive: since there is (almost) no loss of information in the acquisition of color concepts, there is nothing further to be made available to the central system for digitalization.  Acquisition of sensory concepts is therefore brute and primitive: to acquire these concepts it is enough to occupy the relevant sensory states for an organism equipped with an appropriate conceptual system — i.e., by an information pick-up system operating on the sensory representations.  This is why the notion of learning is not appropriate for the acquisition of these concepts.  Rather, the preferred term for this, for both empiricists and nativists, is “triggering.”  So one sense in which the abstraction distance is minimal is that the process underlying the acquisition and vertical deployment of sensory concepts does not involve any loss of information that is nevertheless available to the conceptual system for further digitalization.

Contrast this to the intra-cranial sustaining mechanisms for other concepts, which are (partially but essentially) cognitive.  The acquisition and deployment of perceptual concepts may be innate and automatic in some sense, but these still involve a digitalization process with considerable loss of information, information that is still available for digitalization.  When we visually recognize shapes of objects or geometrical figures, most of the information about their spatially distributed and organized constituents (illumination gradients, edges, corners, curves, color, etc.) is still consciously available.  It isn’t that we consciously use this information in the acquisition and deployment of such concepts — this is something our perceptual (as opposed to sensory) systems automatically do for us.  But what is interesting is that even though this process may be automatic and unconscious, most of the information used in the process (which is then discarded) is available to us, to the central cognitive system, and thus is conscious in just that sense.  Because of the importance and centrality of perceptual concepts, their acquisition may still be innately determined — i.e., such concepts may be triggered rather than learned.  We leave this issue open.

The notion of learning seems most appropriately applied to the acquisition of observational (and for that matter, theoretical) concepts.  The sustaining mechanisms for those concepts are heavily cognitive, involving the use and loss of a great amount of information, which is also normally consciously available.[30]  Generally, the more cognitive the sustaining mechanisms of a concept are, the greater the abstraction distance between it and the sensory bases from which it is acquired.

Before we move on to examine what makes sensory concepts special, we would like to make a few observations about what is implied by the architecture of the information flow from the sensory to the conceptual.  If we are right, then there is a deep point to be made about autonomous representational systems:

 

(i)    Such systems are nomologically bound to be hooked up to their environments in a way that at some level of abstraction they will always harbor sensory representations that represent complex physical properties in their environment as simple or atomic, or rather, do not represent them as having internal complexity.

(ii)  Furthermore: necessarily, if an autonomous intentional organism has concepts at all (or a conceptual system, as opposed to just sensory representations), however primitive or sophisticated, then it has some sensory concepts in our sense.

 

One of the most basic truths about autonomous intentional systems is that they have to interact with their environment informationally.  So they have to have information entry mechanisms.  These mechanisms cannot deliver every piece of information in analog form, i.e., in a form that is always nested by some further more specific information.  There will have to be a cut-off point about the most specific information the mechanism can provide about the environment.  If this piece of digital information carries the analog information nested in it in an extractable format, then there will have to be structural features of the output representation carrying the (total) digital information that nest this information.  Then the same question arises about the digital content of these features and its format.  This process cannot go indefinitely.  At some point there will have to be representational features with digital informational content that nests the analog information carried by them in a non-extractable format, at which point the property digitally represented won’t be represented as having internal constituents — if the property has internal constituents (this can be a massively disjunctive property like colors).  As will become clear as we proceed, it is these necessities that partly create the mystery around phenomenal consciousness.

 

4  WHAT MAKES SENSORY CONCEPTS SPECIAL

It is not accidental that the distinction we drew between sensory and perceptual concepts is approximately coextensive with the distinction traditionally drawn between concepts of secondary and primary qualities, respectively.[31]  Secondary qualities are those which are represented in our experiences in a primitive way: sensory representations carry information about them in a way that makes the information carried about their constituents analog but non-extractable. (It is the job of empirical scientific investigation to reveal the complex nature of secondary qualities, and extract the information about their constituents.)  Hence, sensory experiences carry the most specific information about these properties without revealing their internal structure.  This is why the abstraction distance between the concepts of secondary qualities and their sensory bases is minimal; equivalently, this is why the acquisition of these concepts is non-cognitive and brute.

Sensory concepts apply, in the first instance, to the objects of perception, to whatever it is that our sensory experiences represent.[32]  This is so despite the fact that they are directly and immediately acquired from sensory representations.  The flow of information required for their acquisition (and vertical deployment) necessitates the presence of sensory intermediaries that carry information about the properties denoted by these concepts.  Indeed, this is one of the main differences between sensory and observational concepts.[33]  There is an asymmetry in their acquisition: while sensory concepts are necessarily acquired from the experiences sensorially representing the properties they denote, observational concepts are different.  Observational concepts are typically acquired from experiences representing their denotations, but this is not necessary.  We can acquire them “horizontally,” i.e., by sensory means (speech perception, seeing pictures, reading books/newspapers, inference, etc.) that are only very indirectly related to, and hence don’t carry information about, their denotations.

There is a deep reason for this asymmetry which we haven’t touched on so far but will be very important for what follows: the information about the secondary qualities contained in experiences cannot be completely digitalized by the conceptual system, whereas the conceptual system can completely digitalize the information contained in experiences about the properties denoted by observational concepts.[34]  “Complete digitalization” is a technical term introduced by Dretske that expresses a necessary condition for a piece of information to count as the semantic content of a concept.  Recall that the semantic content of a concept is the most specific information its vertical tokenings carry about the objects it applies to, which is equated with its digital informational content.  But Dretske eventually refines this definition by requiring that the semantic content be that piece of information which is completely digitalized.  Here is the definition (1981: 185):

 

Structure S has the fact that t is F as its semantic content [i.e., S is the concept of an F] =definition

(a)    S carries the information that t is F and

(b)    S carries no other piece of information, r is G, which is such that the information that t is F is nested (nomically or analytically) in r’s being G.

 

Condition (b) ensures that if S carries the information that t is F, it does so not by carrying information about any intermediary which nests the information that t is F.  When the two conditions are satisfied S carries the information that t is F in completely digital form, or equivalently, S is said to completely digitalize the information that t is F, which then becomes S’s semantic content.  More intuitively, the intention is to rule out those cases where concepts carry the most specific distal information about an object by carrying information about their proximal causes, in our case their sensory bases.[35]  So, for instance, the concept ROBIN, when acquired from experiences that carry information about robins, should not carry information about the structure of sensory representations that give rise to ROBIN.[36]  Since we are working in a naturalistic framework, if concepts carried information about sensory representations from which they are acquired, this information would be information about the instantiation of certain neurophysiological properties (or disjunctive sets of such properties) realizing these sensations.  Hence our concepts would be selectively responding to such properties in the first place.  And this would imply either that our concepts represent neurophysiological conditions, or that our sensory concepts have dual semantic content, and therefore are systematically ambiguous.

Interestingly, Dretske does not make a point about the empirical impossibility of complete digitalization; nor does he talk about the fact that complete digitalization is routinely violated in the case of sensory concepts.[37]  If sensory representations of secondary qualities are realized by a more or less homogeneous set of neurophysiological properties, or by a finite disjunction of such properties, then vertical tokenings of sensory concepts carry information about their distal causes (instantiations of secondary qualities) by carrying information about the instantiations of these proximal physical properties.  Whether or not experiences of such qualities are physically realized in a homogeneous way is ultimately an empirical question, but we think that there is enormous empirical as well as a priori evidence that this is the case — certainly intrapersonally, and most probably interpersonally.[38]  What is important for our purposes is the  claim that the neurophysi­ological realization bases of sensory representations of such qualities are not indefinitely and arbitrarily varied, but consist of a finite disjunctive set of physical properties, and are more or less homogeneous in just this sense.  We think that this claim is true, but we stand ready to be corrected by future empirical evidence.

There are also overwhelmingly strong engineering reasons for this claim: whenever you make an architectural distinction between a sensory buffer and a conceptual system that extracts information about the distal layout from this buffer (and whose behavior is causally sensitive to what this buffer contains), there will be a need to correlate the information carried by concepts and the elements of the buffer in such a way that matches up with the distal layout.  If the only way the conceptual system carries information about the distal properties is through a physically realized sensorium, then it had better be the case that the same elements of this sensorium carried the same information, at least in the case of secondary qualities where the abstraction distance is minimal.  Otherwise, the informational efforts of the conceptual system will be fooled.  From an engineering perspective, it is unclear how such an architectural design can be constructed without making the realization bases of those sensory representations more or less physically homogeneous (i.e., not arbitrarily varied), at least within a single system.[39]

Notice that in the case of observational concepts there is no real problem about complete digitalization.  There are indefinitely many ways robins, trucks, etc. can affect our sensory receptors, and thus many ways in which they can be represented in experience.  In such cases, the standard information-theoretic remedy is to say that these concepts track their distal causes without tracking proximal sensory representations, since the alternative is to say that they track a massively (probably open-ended) disjunctive proximal property.  We believe that the former is indeed more plausible than the latter.  But if so, we can now see better why sensory representations carrying information about properties denoted by observational concepts are not necessary for acquiring such concepts.  These concepts, though observational, are modality-neutral (amodal), and to that extent not perspectival.  But that is not to say that their cognitive sustaining mechanisms don’t involve sensory/perceptual channels and concepts; they do.  It is to say, however, that the sustaining mechanisms involved provide information only (mostly) contingently related to the denotation of these concepts.

It is the failure of complete digitalization that makes sensory concepts special by giving them a perspectival and quasi-indexical character.  Their acquisition, semantics, and vertical deployment are essentially host-unique in two senses:

 

(i)    It matters essentially for whose cognitive system these cognitive structures function as concepts.

(ii)  They track features of the environment (instantiations of secondary qualities however objectively understood) essentially by tracking something about their host, namely, the sensory experiences from which they are directly acquired.

 

In other words, these are concepts which a properly functioning conceptual system cannot normally acquire unless suitably hooked up to a properly functioning sensory/perceptual information delivery system of the same host that has actually delivered the necessary information, i.e., carried information about the properties denoted by these sensory concepts.[40]  We also want to emphasize that their acquisition is direct and immediate, by which we mean this: their sustaining/acquisi­tion mechanisms are not cognitive, but primitive and brute; that is, they don’t involve the exploitation of consciously available information that is then discarded in the digitalization process.  This is roughly to say that the abstraction distance involved in their acquisition is minimal.  In the context of our discussion above, this implies that no information about the internal constituents of the properties denoted by sensory concepts is available in an extractable format: they don’t represent their denotations as having a complex internal structure.  All these points about sensory concepts will be crucially important later on, when we criticize conceivability arguments against physicalism.[41]

 

5    FIXING THE SEMANTIC CONTENT OF SENSORY CONCEPTS

What justifies the claim that, despite the failure of complete digitalization, the semantic content of a sensory concept, say RED, is the secondary quality, redness,  possessed by the objects of the sensory experiences from which we directly acquire it?[42]  Irrespective of what semantic content our theories assign to these concepts, there should be no doubt about what their semantic contents are: they are the qualities that our experiences represent the external objects as having.  Our experiences place these qualities in the world of objects external to our bodies.  So do our sensory concepts.  Given this, the question before us is how to reconcile a Dretskean informational semantics with the failure of complete digitalization.  For even if we rightly want to be able to say that RED represents redness despite the failure of complete digitalization, what justifies rejecting the option, which seems to be a consequence of the theory, that the semantic content of RED is the experience of redness, i.e., E-red, realized by a certain set of neurophysiological properties?

Here is another way of putting the problem.  Informational semantics starts with the information carried by a structure on its way to working out how to determine its semantic content (SC).  We have seen that Dretske wants to assign the completely digitalized informational content of a concept (C) as its semantic content: in other words,

 

·       the semantic content is the most specific information carried by C about a source o such that there is no separate structure e such that C carries the most specific information about o by carrying the most specific information about e.