A graded, dynamic perspective
Axel Cleeremans
Cognitive Science Research Unit
Université Libre de Bruxelles CP 122Avenue F.-D. Roosevelt, 501050 Bruxelles — BELGIUMEmail: axcleer@ulb.ac.be
Luis Jiménez
Facultad de PsicologíaUniversidad de Santiago
15706 Santiago de Compostela — SPAINEmail: jimenez@usc.es
May, 2001
Implicit learning: A graded, dynamic perspective
2
1. IntroductionWhile the study of implicit learning is nothing new, the field as a whole has come toembody — over the last decade or so — ongoing questioning about three of the mostfundamental debates in the cognitive sciences: The nature of consciousness, the natureof mental representation (in particular the difficult issue of abstraction), and the roleof experience in shaping the cognitive system. Our main goal in this chapter is to offera framework that attempts to integrate current thinking about these three issues in away that specifically links consciousness with adaptation and learning. Ourassumptions about this relationship are rooted in further assumptions about the natureof processing and of representation in cognitive systems. When considered together,we believe that these assumptions offer a new perspective on the relationshipsbetween conscious and unconscious processing and on the function of consciousnessin cognitive systems.
To begin in a way that reflects the goals of this volume, we can ask the question:“What is implicit learning for?” In asking this question, one presupposes that implicitlearning is a special process that can be distinguished from, say, explicit learning or,even more pointedly, from learning tout court. The most salient feature attributed toimplicit learning is of course that it is implicit, by which most researchers in the areaactually mean unconscious. Hence the question \"What is implicit learning for?\" is infact a way of asking about the function of consciousness in learning that specificallyassumes that conscious and unconscious learning have different functions. The centralidea that we will develop in this chapter is that conscious and unconscious learningare actually two different expressions of a single set of constantly operating graded,dynamic processes of adaptation. While this position emphasizes that conscious andunconscious processing differ only in degree rather than in kind, it is nevertheless notincompatible with the notion that consciousness has specific functions in the cognitiveeconomy.
Indeed, our main conclusion will be that the function of consciousness is to offerflexible adaptive control over behavior. By adaptive here, we do not mean simply thepossibility for an agent to select one course of action among several possibilities.This, as dozens of computer programs routinely demonstrate, can be achieved withoutconsciousness. Instead, we assume that genuine flexibility necessarily involvesphenomenal consciousness (subjective experience), to the extent that successfuladaptation in cognitive systems seems to make it mandatory that behavioral changesbe based on the rewarding or punishing qualia they are associated with. There would
Implicit learning: A graded, dynamic perspective
3
be no point, for instance, in avoiding dangerous behavior were it not associated withfeelings of danger. Learning is thus necessarily rooted, we believe, in the existence ofat least some primitive ability for the cognitive agents to experience the consequencesof their behavior and to recreate these experiences independently of action. Theseprimitive experiences can then, through more elaborate learning and developmentalprocesses, become integrated into increasingly complex structures that includerepresentations of the self, that is, into a set of representations and processes thatenable an agent to entertain a third-person perspective on itself, or, in other words, tolook upon itself as though it were another agent. We surmise that any information-processing system that is sufficiently complex to make such processes possible shouldbe characterized as conscious — albeit we may never find out unless this systemexhibits the only sort of consciousness that we know of first-hand, that is, humanconsciousness. We will not discuss this important epistemic debate any further shortof noting (1) that it actually is what the Turing Test is about (see French, 2000, forfurther discussion of the Turing Test), and (2) that it is perfectly possible to developsimulations of some behavior that successfully mimics adaptation without requiringqualia, but then, presumably, only at a level of description that would fail to pass moreelaborate testing.
Our primary goal in this chapter will thus be to outline a novel framework withwhich to think about the relationships between learning and consciousness. In section2, we propose to define learning as “a set of philogenetically advanced adaptationprocesses that critically depend on an evolved sensitivity to subjective experience soas to enable agents to afford flexible control over their actions in complex,unpredictable environments”. We continue by discussing the implications of such adefinition of learning on current debates about (1) the nature of phenomenalexperience (section 3) and about (2) the functions of consciousness in cognitivesystems (section 4). In section 5, we turn to an overview of our own proposal, andcontinue by briefly illustrating how our framework can be used to understand diversephenomena in domains such as priming, implicit learning, automaticity and skillacquisition, or development (section 6). We conclude the chapter (section 7) byconsidering issues that the framework does not address. We should add that thischapter is by no means intended to offer a complete overview of all relevantphenomena and theories, but rather to convey the flavor of what we believe to be analternative framework in which to consider some of the central issues in the domain ofimplicit learning.
Implicit learning: A graded, dynamic perspective
4
2. Adaptation, adaptive changes, and learningMounting evidence suggests not only that the brain is far more plastic than previouslythought, but also that the effects of learning can be tracked all the way down to theorganization of local connectivity. To wit: Expert string players exhibit larger-than-normal areas of the somatosensory cortex dedicated to representing input from thefingering digits (Elbert et al., 1995). Likewise, not only is posterior hippocampus, — aregion of the brain involved in episodic and spatial memory — enlarged inexperienced taxi drivers compared to subjects who do not have extensive experiencein memorizing complex maps, but the observed size differences further depend on theamount of driving experience (Maguire et al., 2000). There is also considerableevidence that the brain can recover in various flexible ways after trauma, and evensuggestions that the very organization of the somatosensory cortex (the famousPenfied homonculus) depends on pre-natal sensory experience (Farah, 1998). Morerecently, suggestive evidence for neurogenesis was also found in humans (Eriksson etal., 1998) — a finding that overturned decades of unquestioned — but, as it turns out,erroneous — assumptions about the lack of regenerative cellular processes in the adultbrain. These often spectacular findings all reassert that adaptation plays a fundamentalrole in cognition, and that its effects can be traced all the way down to the manner inwhich specific neural circuits are organized.
Given this plethora of new findings hinting that the brain constantly adapts to theenvironment that it is immersed in, what can we say about the relationships betweenlearning and consciousness? Should we consider processes of adaptation in general tobe distinct from processes of learning? Is it the case, as some authors contend (seePerruchet & Vinter, this volume; Shanks & St. John, 1994) that learning is alwaysaccompanied by conscious awareness? One can ask the question in another way: Whyshould behavior always be available to conscious control? It might seem particularlyadaptive for complex organisms to be capable of behavior that does not requireconscious control, for instance because behavior that does not require monitoring ofany kind can be executed faster or more efficiently than behavior that does requiresuch control. Reflexes such as withdrawing one's hand from a fire are good instancesof behaviors that have presumably evolved to the point that they have beenincorporated in the functional architecture of an organism's central nervous systemand cannot be controlled any longer (or perhaps, only with extensive training on self-control techniques).
Implicit learning: A graded, dynamic perspective
5
The relative accessibility of different actions to conscious awareness suggests thatan important distinction between adaptation in general and learning is, precisely, theextent to which consciousness accompanies each. Learning, according to manystandard definitions (e.g., Anderson & Memory, 1995; Klein, 1991; Tarpy, 1997),constitutes a subset of philogenetically advanced adaptation processes that arecharacteristic of so-called “cognitive systems”, and through which relativelypermanent and generally adaptive changes in the behavior or dispositions of theorganism arise as the result of their previous \"experiences\" with the environment inwhich they are immersed. From such a definition, it follows that the distinctionbetween learning phenomena and the superordinate class of adaptation phenomena towhich they belong depends on the “cognitive” status of the systems in which suchlearning occurs, and on the ability of these systems to enjoy a particular kind ofsensitivity — \"experience\". Thus, however many reasons there might be to consideradaptation and learning as fundamentally rooted in the same mechanisms, we do notthink that learning can simply be equated with adaptation. Adaptation, indeed, is avery broad concept. When taken to its limit, it might be used to refer to any dynamicrelationship between an object and its environment through which (1) the objectchanges its states and dispositions (2) as a result of its prior sensitivity to theenvironment and (3) in a way that continuously modifies this sensitivity. It should beclear that by this definition, even inanimate objects such as rocks, thermostats orcomputer programs all exhibit patterns of adaptation. Indeed, erosion in rocks, theswitch of a relay in a thermostat, or the occurrence of specific digital states incomputers, can all be characterized as adaptive “responses” to changingenvironmental conditions, to the extent that they modify the systems’ futuresensitivity to the reoccurrence of the same environmental conditions. In livingsystems, these processes of adaptation are further subject to continuous evolution on aspecies basis through the laws of natural selection.
Is it reasonable to consider such processes as processes of learning? Consider againstandard definitions of learning. What, exactly, in these definitions, does “experience”refers to? Should our “experiences” as human beings be considered as similar to thoseof stones and amoebas? Certainly not. However, the literature about learning is ingeneral conspicuously prone to conflate the term \"experience\" with any other kind ofphenomenally neutral sensitivity that produces relatively permanent and adaptivechanges in the responses of a system. For instance, even though neither machines norneurovegetative systems are generally considered to be endowed with subjectiveexperience, there is at least one journal that is entirely devoted to \"Machine Learning\".It is also relatively easy to find articles in psychological journals in which the changes
Implicit learning: A graded, dynamic perspective
6
produced in our neurovegetative systems in response to their environment areanalyzed as examples of learning (e.g., Ader & Cohen, 1985).
While this conflation between “experience” and “mere sensitivity” has had themerit of emphasizing that there is a continuity between the processes of change thatoccur in different natural or artificial systems, it also blurs the distinction betweenlearning and adaptation phenomena in general. In so doing, it has also furthercontributed to doing away with the distinction between cognitive and non-cognitivesystems. Dennett (1996), in particular, has made this conflation completely explicit byassuming that the differences between cognitive and non-cognitive systems (e.g.,between most animals and plants) might only be in the eye of the beholder. Indeed,according to Dennett, the main difference between animals and plants is that we tendto adopt an intentional stance when analyzing animals’ behavior, but do not do sowhen it comes to understand the dynamics of plant adaptation. As he boldly puts it,there is no reason to dispute the claim that plants should be considered as extremelyslow animals whose \"experiences\" are overlooked because of our “temporal scale”chauvinism (Dennett, 1996), or that libraries should be taken as cognitive systems thatuse researchers as tools to reproduce themselves (Dennett, 1991).
While this conclusion strikes many of us as bluntly absurd, perhaps its absurdityshould be taken as an indication that we need to revisit the notion of “experience” and,in so doing, attempt to carefully delineate what it entails. Indeed, if learning is afundamental element of what it takes for a system to be “cognitive” (e.g., Dretske,1988), it might also be the case that the nature of the phenomenal states upon whichlearning operates is essential to distinguish it from other processes of adaptation. Thisanalysis thus forces us to look into the nature of phenomenal experience in somedetail. That is what we attempt to do in the next section.
3. ConsciousnessWhat is consciousness? While it would be foolish to even attempt to answer thisquestion in this chapter, it might nevertheless be useful to offer guidelines about thesorts of explanations we are looking for, and about which of these are relevant to thestudy of implicit learning. In the following, we briefly discuss three aspects ofconsciousness that often tend to be overlooked in discussions of implicit learning: Thefact that consciousness is not a unitary phenomenon, the fact that consciousness isgraded, and the fact that consciousness is dynamic.
Implicit learning: A graded, dynamic perspective
7
First, consciousness is not a unitary concept, but instead includes differentdimensions. Block (1995), for instance, distinguishes between access consciousness,phenomenal consciousness, monitoring consciousness and self consciousness.Everybody agrees that the most problematic aspect of consciousness is phenomenalconsciousness, or subjective experience, that is, the fact that information processing isaccompanied by qualia — elements of conscious imagery, feelings or thoughts thattogether appear in our mind to form a coherent impression of the current state ofaffairs.
In the specific context of research about implicit learning, the central question isthus: Can changes in behavior occur without correlated changes in subjectiveexperience, and are these changes best characterized as mere adaptation or aslearning? This, at it turns out, is also one of the central questions in the ongoing“search for the neural correlates of consciousness” that has been the focus of so muchrecent empirical research about consciousness in the cognitive neurosciences. In anexcellent overview, Frith, Perry and Lumer (1999) have suggested to organizeparadigms through which to study the “neural correlates of consciousness” in ninegroups resulting from crossing two dimensions: (1) three classes of psychologicalprocesses involving respectively knowledge of the past, present, and future —memory, perception, and action —, and (2) three types of cases where subjectiveexperience is incongruent with the objective situation — cases where subjectiveexperience fails to reflect changes in either (1) the stimulation or in (2) behavior, and(3) cases where subjective experience changes whereas stimulation and behaviorremain constant.
The paradigmatic example of the third situation is binocular rivalry, in which anunchanging compound stimulus consisting of two elements each presented separatelyand simultaneously to each eye produces spontaneously alternating completeperceptions of each element. By asking participants to indicate which stimulus theyperceive at any moment, one can then hope to establish which regions of the brainexhibits activity that correlates with subjective experience and which do not, in asituation where the actual stimulus remains unchanged. Frith et al. go on bydelineating many other relevant empirical paradigms involving both normal subjectsas well as patients suffering from a variety of neuropsychological syndromes. Whilereviewing these different paradigms in detail goes far beyond the scope of thischapter, it is interesting to note that implicit learning, in their analysis, constitutes oneexample of cases where subjective experience remains constant while behaviorchanges. The study of implicit learning is thus highly relevant to the study ofconsciousness in general.
Implicit learning: A graded, dynamic perspective
8
In addition to the well-known difficult challenges involved in designing empiricalparadigms suitable for the exploration of differences between conscious andunconscious processing (see Cleeremans, 1997 for an overview of these issues), thestudy of consciousness also notoriously involves a great deal of conceptual issues. Inthis respect, it is worth pointing out that current theories of consciousness indeedmake sometimes very contrasted assumptions about its underlying mechanisms. Forinstance, Farah (1994) proposed to distinguish between three types of neuroscientificaccounts of consciousness: “Privileged Role” accounts, “Integration” accounts, and“Quality of Representation” accounts. “Privileged Role” accounts take their roots inDescartes’ thinking and assume that consciousness depends on the activity of specificbrain systems whose function it is to produce subjective experience. “Integration”accounts, in contrast, assume that consciousness only depends on processes ofintegration through which the activity of different brain regions can be synchronizedor made coherent so as to form the contents of subjective experience. Finally,“Quality of Representation” accounts assume that consciousness depends onparticular properties of neural representations, such as their strength or stability intime.
In a recent overview article (Atkinson, Thomas, & Cleeremans, 2000, see alsoO'Brien & Opie, 1999), we proposed to organize computational theories ofconsciousness along two dimensions: (1) A process vs. r epresentation dimension,which opposes models that characterize consciousness in terms of specific processesoperating over mental representations, with models that characterize consciousness interms of intrinsic properties of mental representations, and (2) A specialized vs . non - specialized dimension, which contrasts models that posit information-processing systems dedicated to consciousness with models for which consciousness can beassociated with any information-processing system as long as this system has therelevant properties. Farah’s three categories can be subsumed in this analysis in thefollowing manner: “Privileged Role” models, which assume that some brain systemsplay a specific role in subtending consciousness, are specialized models that can beinstantiated either through vehicle or through process principles. “Quality ofRepresentation”, models, on the other hand, are typical vehicle theories in that theyemphasize that what makes some representations available to conscious experienceare properties of those representations rather than their functional role. Finally,Farah’s “Integration” models are examples of non-specialized theories, which canagain be either instantiated in terms of the properties of the representations involvedor in terms of the processes that engage these representations.
Implicit learning: A graded, dynamic perspective
9
Atkinson et al.’s analysis thus offers four broad categories of computationalaccounts of consciousness:
(1) Specialized vehicle theories , which assume that consciousness depends on theproperties of the representations that are located within a specialized system inthe brain. An example of such accounts is Atkinson and Shiffrin’s (Atkinson& Shiffrin, 1971) model of short-term memory, which specifically assumesthat representations contained in the short-term memory store (a specializedsystem) only become conscious if they are sufficiently strong (a property ofrepresentations).(2) Specialized process theories , which assume that consciousness arises fromspecific computations that occur in a dedicated mechanism, as in Schacter’sCAS (Conscious Awareness System) model (Schacter, 1989). Shacter’s modelindeed assumes that the CAS’s main function is to integrate inputs fromvarious domain specific modules and to make this information available toexecutive systems. It is therefore as specialized model in that it assumes thatthere exist specific regions of the brain whose function it is to make itscontents available to conscious awareness. It is a process model to the extentthat any representation that enters the CAS will become available to consciousawareness in virtue of the processes that manipulate these representations, andnot in virtue of properties of those representations themselves.(3) Non-specialized vehicle theories include any model that posits that availabilityto consciousness only depends on properties of representations, regardless ofwhere in the brain these representations exist. O’Brien & Opie’s“connectionist theory of phenomenal experience” (O'Brien & Opie, 1999) isthe prototypical example of this category, to the extent that it specificallyassumes that any stable neural representation will both be causally efficaciousand form part of the contents of phenomenal experience.(4) Non-specialized process theories , finally, are theories in which it is assumedthat representations become conscious whenever they are engaged by certainspecific processes, regardless of where these representations exist in the brain.Most recent proposals fall into this category. Examples include Tononi andEdelman’s “dynamic core” model (Tononi & Edelman, 1998); Crick andKoch’s idea that synchronous firing constitutes the primary mechanismsthrough which disparate representations become integrated as part of a unifiedconscious experience (Crick & Koch, 1995), or Grossberg’s characterization
Implicit learning: A graded, dynamic perspective
10
of consciousness as involving processes of “resonance” through whichrepresentations that simultaneously receive bottom-up and top-downactivation become conscious because of their stability and strength(Grossberg, 1999).
While most recent neuro-computational models of consciousness fall into the lastcategory, several proposals also tend to be somewhat more hybrid, instantiatingfeatures and ideas from several of the categories described by Atkinson et al. Baars'influential \"Global Workspace\" model (Baars, 1988), for instance, incorporatesfeatures from specialized process models as well as from non-specialized vehiclestheories, to the extent that the model assumes that consciousness involves aspecialized system (the global workspace), but also characterizes conscious states interms of the properties associated with their representations (i.e. global influence andwidespread availability) rather than in terms of the processes that operate on theserepresentations. Likewise, Dehaene and Naccache’s recent “neural workspace”framework (Dehaene & Naccache, 2001) assumes that consciousness depends (1) onthe existence of a distributed system of long-range connectivity that links manydifferent specialized processing modules in the brain, and (2) on the simultaneousbottom-up and top-down activation of the representations contained in the linkedmodules. Thus, this model acknowledges both the existence of specific, dedicatedmechanisms to support consciousness as well as the specific properties ofrepresentations (e.g., their strength or stability) brought about by specific processes(e.g., resonance).
These various tentative accounts of the neural or computational mechanisms ofconsciousness are highly relevant to the study of implicit learning because any theoryof the mechanisms through which implicit learning occurs necessarily also has tomake corresponding assumptions about the mechanisms of consciousness. As we shallsee in section 4, however, most existing theories of implicit learning tend to be rathermute about their implications with respect to the study of consciousness. Indeed, mostof the debate in the psychological literature about the relationships between consciousand unconscious processing has been dedicated to addressing methodological ratherthan conceptual issues. While these methodological debates are of central importance,we also believe that addressing the conceptual issues is essential.
A second central aspect of conscious experience — and one that is also particularlyrelevant for behavioral studies of implicit cognition, is that consciousness is not an all-or-none process or property, but that it affords many degrees and components.Conscious experience, however unified it appears to us, is not a single thing. Any
Implicit learning: A graded, dynamic perspective
11
theory of consciousness therefore has to answer questions about how the variouselements of conscious experience are integrated with each other so as to form aunified whole, and about how to best think about the relative complexity of differentsorts of conscious experiences. In other words: How does one go from the simpleexperiences that a snail might enjoy of its surroundings to the considerably morecomplex experience produced by your reading these words? How does one accountfor the differences between the sort of consciousness that infants undoubtedly possessto the sort of verbally rich consciousness that adults enjoy? Process and vehicletheories of consciousness make very different assumptions about these questions. ForO’Brien & Opie (1999), for instance, the graded character of conscious experience isreadily accommodated by vehicle theories, to the extent that properties ofrepresentations such as strength or stability in time can easily be mapped ontocorresponding degrees or components of conscious awareness. This mapping issomewhat more delicate for what we have called process theories, even though at firstsight they appear to offer an appealing set of conceptual principles with which tounderstand how conscious experience can increase in complexity throughdevelopment or learning.
Dienes & Perner (1999) have recently pursued this goal in their theory of explicitand implicit knowledge, and “higher-order thought” (HOT) theories of consciousnessin general can be described as relying on this principle (e.g., Rosenthal, 1986, 1997).However, what is harder to accept from such accounts of subjective experience is thatits phenomenal character could be brought about in the first place from a series ofcomputational processes performed on otherwise non-phenomenal representations.Indeed, and however much one might disagree with the specific way in which thisthought experiment was framed, Searle’s Chinese Room argument showed us twentyyears ago that the phenomenal properties of experience just seem not to be the sort ofstuff that one might expect to obtain by mere shuffling of formal representations orsymbols, no matter how convoluted, recurrent, or complex the relation among thesesymbols may turn out to be (Searle, 1980; 1992; 1999). Neither semantics norphenomenal experience can emerge out of syntax. Symbols need to be grounded.Hence, if this intuition is right, a pure process theory could never tell us the last wordin accounting for the first principles of consciousness.
Vehicle theories, it therefore appears, appear to be the best candidates to accountfor the emergence of the first elements of subjective experience which, throughprocesses of learning, development and socialization, subsequently provide theappropriate foundations for the emergence of more elaborated forms of consciousness.It must be made clear at this point that by \"vehicle theories\" we refer to any theory
Implicit learning: A graded, dynamic perspective
12
that assumes that experience is not merely a relational or syntactic property that couldbe realized through any representational vehicle, but that claim instead that experiencearises in a specific medium (e.g., neural) and as a result of processes that are proper tothis medium1.
For the sake of discussion, let us simply accept that phenomenal experience arisesas the result of some neural processes. What, then, might be the functions fulfilled byphenomenal experience? What is it about experience that makes it play a special rolein distinguishing between learning and mere adaptation? These questions are in factquestions about a third aspect of consciousness, that is, its dynamical character. Mostdiscussions of consciousness tend to analyze it as a static property of some processesor representational states. However, it is obvious that consciousness is a phenomenonthat is highly dynamical: What I am aware now I might be unaware of at the nextmoment. Likewise, what I am aware of at some point in time when learning a newskill is not identical with what I will be aware of after I have mastered the skill. Thus,we therefore believe that processes of change are central to our understanding ofconsciousness, and that an analysis of its possible functions should therefore be rootedin an analysis of the role that learning and adaptation play in shaping action.
4. The function of consciousness: Commander Data meets the ZombiesThe findings briefly overviewed at the beginning of section 2 raise the question ofwhat the role of consciousness might be in adaptation and learning. We concluded thata significant difference between adaptation and learning is whether or notconsciousness is involved. In this section, we attempt to reflect upon the function thatconsciousness might have in information processing. In so doing, we suggest thatmost existing theories of the relationships between conscious and unconsciousprocessing have simply failed to give consciousness a clear functional role.
In a recent overview article, Dehaene and Naccache (2001) conclude that “Thepresent view associates consciousness with a unified neural workspace through whichmany processes can communicate. The evolutionary advantages that this systemconfers to the organism may be related to the increased independence that it affords.”(p. 31). Dehaene and Naccache thus suggest that consciousness allows organisms tofree themselves from acting out their intentions in the real world, relying instead onless hazardous simulation made possible by the neural workspace. While we certainlyagree with this conclusion, it begs the question of how consciousness came to playthese functions in the first place. Are there any adaptive or evolutionary causes that
Implicit learning: A graded, dynamic perspective
13
would favor the emergence of unifying control systems characterized by consciousstates, and that could go beyond what local adaptive processes can do by forcing largeparts of the nervous system to work together in a coherent direction for some fractionsof seconds? How can these coherent, resonant, synchronous, reverberant, or otherwiseconscious states of the system come to reflect the most adaptive representation of thecurrent situation, given that “what is most adaptive” continuously changes?
As discussed by Perruchet and Vinter (this volume), the answers to these questionsare intimately related to the dynamics between learning and consciousness: On theone hand, phenomenal consciousness provides the cognitive system with an adapted,global representation of the current situation so that learning mechanisms operate onthe best possible representations. On the other hand, learning changes theserepresentation in increasingly adaptive ways. From this perspective then, the centralfunction of consciousness is to offer flexible, adaptive control over behavior.
This complex, dynamical relationship between consciousness and learning has,however, often tended to be overlooked in classical models of cognition. As argued inCleeremans (1997) and also in Jiménez and Cleeremans (1999), this is most likely dueto the fact that classical models of cognition (the “Computational Theory of Mind”,see Fodor, 1975) take it as a starting point that cognition is symbol manipulation. Aswe will try to highlight in the next few paragraphs, we surmise that one takescognition to be exclusively and exhaustively about symbol manipulation, then thereare but a few conceptual possibilities with which to think about differences betweenconscious and unconscious states.
Cognitive scientists concerned with the relation between consciousness andcognition generally tend to oscillate between two extreme (and admittedly caricatural)positions, which we have dubbed “Commander Data” and “Zombie” theories ofcognition. Star Trek’s character Data is an android whose bodily and cognitiveinnards are fully transparent to himself. Except in rare circumstances (whichsystematically tend to be described as the result of some sort of dysfunction), Data isthus capable of describing in uncanny detail each and every aspect of its internalstates: How much force he is applying when attempting to pry open a steel door, howmany circuits are currently active in his positronic brain, or the number of times overthe last ten years he smelled a particular scent, and in which circumstances he did so,etc. Commander Data theorists likewise assume that cognition is fully transparent,that is, (1) that whatever knowledge is expressed through behavior is alsotransparently available to introspection, and (2) that consciousness reigns supreme andallows access, with sufficient effort or attention, to all aspects of our inner lives. This
Implicit learning: A graded, dynamic perspective
14
perspective is what Broadbent described as the “common sense” view of cognition,according to which “people act by consulting an internal model of the world, adatabase of knowledge common to all output processes, and manipulating it to decideon the best action” (Broadbent, Fitzgerald, & Broadbent, 1986, p. 77).
In contrast, the famed philosophical zombies (Chalmers, 1996) are perfectlyopaque, and in this sense instantiate absolutely implicit beings: Whatever internalknowledge currently influences their behavior can neither be explicit nor consciousbecause, by definition, they lack conscious experience. Zombie theorists thus take itas a starting point that consciousness has an epiphenomenal character: There is azombie within you and, while you may not be aware of its existence, it could in factbe responsible for most of your actions. It is capable of processing all the informationyou can process in the same way that you do, with one crucial difference: “All is darkinside” (Chalmers, 1996, p. 96); your zombie is unconscious. From this perspectivethen, cognition is inherently opaque, and consciousness, when present, offers but avery incomplete and imperfect perspective on internal states of affairs.
Needless to say, both of these perspectives are profoundly unsatisfactory. On theone hand, Zombie perspectives (ZP) ascribe no role whatsoever to consciousness ininformation processing, threaten to rob us of free will, and — because it is absurd todeny consciousness altogether — are ultimately forced to assume the existence ofequally powerful conscious and unconscious systems. On the other hand, CommanderData perspectives (CDP), by assuming that all of cognition is conscious, paradoxicallylikewise depict consciousness as epiphenomenal. Crucially, both perspectives assumethat consciousness does not change cognition in any principled way, and hence thatconsciousness plays no functional role beyond that of a epiphenomenon thataccompanies either a functionally redundant subset of (ZP) or all (CDP) cognitiveevents.
On the face of the deeply counterintuitive flavor of both perspectives, it seemssurprising to see that the past few years have witnessed the appearance of severalbroad theoretical proposals that intentionally or inadvertently endorse either of theseperspectives. Some of these proposals are based on empirical evidence, and argue thatthere is in fact no evidence for unconscious influences in cognition. Thus for instance,Holender (1986), based on an extensive review of the subliminal perception literature,found no evidence for the existence of unconscious priming. Holender (1992) furtherproposed that many congruency effects observed in priming experiments can beaccounted by conflicts between conscious contents, that is, without appealing to theeffects of unconscious influences. Likewise, Shanks and StJohn (1994), expanding on
Implicit learning: A graded, dynamic perspective
15
the perspective offered by Brewer (1976), concluded their target article dedicated toimplicit learning by the statement that “Human learning is almost invariablyaccompanied by conscious awareness” (p. 394).
Other proposals are more conceptual in nature. For instance, O’Brien and Opie(O'Brien & Opie, 1999) propose that the contents of phenomenal consciousnessinclude all stable neural states, and that it is only those stable states that are “causallyefficacious”, that is, susceptible to influence further processing and, ultimately,behavior. Perruchet and Vinter (1998, this volume), consider that unconsciousinfluences on behavior should be ascribed exclusively to noncognitive, neuralprocesses and state that “Mental life […] is co-extensive with consciousness”(Perruchet, personal communication, see also Dulany, 1997). Finally, Dienes andPerner’s (1999) recent “theory of implicit and explicit knowledge”, while carefullydelineating the various ways in which knowledge can be cast as implicit or explicit,also seems to take it as a starting point that causally efficacious knowledge is alwaysexplicit in some sense , that is, at least at the specific level that is needed to account forthe observed behavioral effects, and hence ends up, we believe, inadvertently paintinga picture of cognition in which the implicit again plays no functional role in cognition.It should be pointed out that if the emphasis of these theories on the \"transparent\"character of cognition can be seen as a normal swing of the conceptual pendulum,there is nevertheless something paradoxical about the emergence of such proposals ata time when the importance of unconscious processing in cognition finally appears tohave gained some form of recognition in dozens of articles, books and conferences.The debate, we believe, is not so much rooted in equivocal empirical findings, butrather in the deep conceptual problems associated with the notion of unconsciousrepresentation. Hence, defenders of the claim that cognition can be unconscious oftensuccumb to some version of the ZP, while defenders of the opposite view can often betaken to endorse some variant of the CDP. Crucially, we believe that both thesegeneral frameworks are in fact based on the classical assumption that cognitioninvolves symbol manipulation , and hence that their only way to separate conscious from unconscious cognition is to assume that unconscious cognition is just likeconscious cognition, but only minus consciousness (Searle, 1992).
In the next section, we would like to sketch out an alternative, subsymbolic,framework through which to think about the relationship between learning andconsciousness — one that we believe offers a clear function to consciousness by
Implicit learning: A graded, dynamic perspective
16
linking it with adaptability in cognitive systems, while at the same time leaving openthe possibility for adaptive changes to occur without consciousness.
5. The frameworkIf our central assumption that the function of consciousness is to offer adaptivecontrol over behavior is correct, then consciousness is necessarily closely related toprocesses of learning, because one of the central consequences of successfuladaptation is that conscious control is no longer required over the correspondingbehavior. We therefore believe that it makes sense to root accounts of consciousnessin accounts of how change occurs in cognitive systems.
Like Perruchet & Vinter (this volume), we assume that there is a dynamicrelationship between consciousness and learning such that (1) awareness of aparticular state of affairs triggers learning and (2) that this learning in turn changes thecontents of subjective experience so as to make these contents more adapted.However, and this is an important departure from Perruchet & Vinter’s framework,we also assume that learning has additional obligatory indirect effects that can fail toenter awareness. In other words, learning is not just about modifying consciousexperience, as Perruchet & Vinter seem to assume. Thus, when I learn about cats, Ialso indirectly learn about dogs and other animals, because the correspondingrepresentations are all linked together by virtue of being embedded in distributedrepresentational systems. These indirect effects of conscious learning need notthemselves be conscious, particularly if they are weak.
We will return to these issues in the general discussion. At this point, we wouldlike to introduce the set of assumptions that together form our framework. In thefollowing, we present these assumptions in four groups: Assumptions aboutinformation processing (P1-4), about representation (R1-3), about learning (L1-3) and,finally, about consciousness (C1-5).
5.1 Assumptions about information processingConsistently with well-known ideas in the connectionist literature (e.g., Rumelhart &McClelland, 1986), we will assume the following without further discussion:
Implicit learning: A graded, dynamic perspective
17
P1.The cognitive system is best viewed as involving a large set of interconnected
processing modules organized in a loose hierarchy. Each module in turn consistsof a large number of simple processing units connected together.
P2.Long-term knowledge in such systems is embodied in the pattern of connectivity
between the processing units of each module and between the modulesthemselves.
P3.Dynamic, transient patterns of activation over the units of each module capture
the results of information processing conducted so far.
P4.Processing is graded and continuous: Connected modules continuously influence
each other’s processing in a graded manner that depends on the strength of theconnection between them and on the strength of the activation patterns that theycontain.5.2. Assumptions about representationRepresentation is one of the most difficult issues to think about in the cognitivesciences because it is often delicate to delineate exactly which states should beproperly taken to be representational (see Dienes & Perner, this volume, for a detaileddiscussion of representation). In the following, and in contrast to purely dynamicalapproaches, we take the perspective that representations are necessary as mediatingstates through which the intermediate results of processing can be captured, therebymaking it possible for complex tasks to be decomposed into modular components.R1.Representations consist exclusively of the transient patterns of activation that
occur in distributed memory systems
This assumption is a central one in our framework because it contrasts with otherrecent proposals (e.g., Dienes & Perner, this volume). In particular, we do notthink that the knowledge that is embedded in the pattern of connectivity betweenunits of a module or between modules themselves is representational in the samemanner that patterns of activation are. Indeed, while such knowledge can beanalyzed as representational from a third-person perspective (because theconnection between two units, for instance, can be described as representing the fact that the units’ activity are correlated), it is never directly available to thesystem itself. In other words, such knowledge is knowledge “in the system” rather“for the system” (see Clark & Karmiloff-Smith, 1993). Knowledge embedded inconnections weights can thus only be expressed dynamically, over the course ofsome processing, when the corresponding representations form over a given set ofprocessing units. These representations can then in turn influence furtherprocessing in other modules. Importantly, and in contrast to thoroughly classical
Implicit learning: A graded, dynamic perspective
18
approaches in cognitive science, the extent to which representations can influenceprocessing in such systems never depends on representations being interpreted bya “processor”.
R2.Representations are graded: They vary on several dimensions that include
strength, stability in time, and distinctiveness
Patterns of activation in neural networks and in the brain are typically distributedand can therefore vary on a number of dimensions, such as their stability in time,their strength. or their distinctiveness. Stability in time refers to how long arepresentation can be maintained active during processing. There are manyindications that different neural systems involve representations that differ alongthis dimension. For instance, the prefrontral cortex, which plays a central role inworking memory, is widely assumed to involve circuits specialized in theformation of the enduring representations needed for the active maintenance oftask-relevant information. Strength of representation simply refers to how manyprocessing units are involved in the representation, and to how strongly activatedthese units are. As a rule, strong activation patterns will exert more influence onongoing processing than weak patterns. Finally, distinctiveness of representationrefers to the extent of overlap that exists between representations of similarinstances. Distinctiveness has been hypothesized as the main dimension throughwhich cortical and hippocampal representations differ (McClelland, McNaughton,& O'Reilly, 1995; O'Reilly & Munakata, 2000), with the latter becoming activeonly when the specific conjunctions of features that they code for are activethemselves.
In the following, we will collectively refer to these different dimensions as“quality of representation” (see also Farah, 1994) For our purposes, the mostimportant notion that underpins these different dimensions is that representations,in contrast to the all-or-none prepositional representations typically used inclassical theories, instead have a graded character which enables any particularrepresentation to convey in a natural manner the extent to which what it refers tois indeed present. A second important aspect of this characterization ofrepresentational systems in the brain is that representations are complex,distributed objects that systematically tend to involve many processing units.R3.Representations are dynamic, active, and constantly causally efficacious.
This assumption simply states that memory traces, far from being staticpropositions waiting to be accessed by some process, instead continuouslyinfluence processing regardless of their strength, stability, or distinctiveness. This
Implicit learning: A graded, dynamic perspective
19
assumption is again central in any connectionist account of cognition. Indeed, ittakes its roots in McClelland’s analysis of cascaded processing (McClelland,1979), which, by showing how modules interacting with each other need not“wait” for other modules to have completed their processing before starting theirown, demonstrated how stage-like performance could emerge out of suchcontinuous, non-linear systems. Thus, even weak, poor-quality traces, in ourframework are capable of influencing processing, for instance through associativepriming mechanisms, that is, in conjunction with other sources of stimulation. Strong, high-quality traces, in contrast have generative capacity , in the sense thatthey can influence performance (i.e., determine responses) independently of theinfluence of other constraints, that is, whenever their preferred stimulus is present.5.3. Assumptions about learningHaving put in place our assumptions about processing and representation, we nowfocus on learning mechanisms. We assume the following:
L1.Adaptation is a mandatory consequence of information processing
Every form of neural information processing produces adaptive changes in theconnectivity of the system, through mechanisms such as Long Term Potentiation(LTP) or Long Term Depression (LTD) in neural systems, or hebbian learning inconnectionist systems. An important aspect of these mechanisms is that they aremandatory in the sense that they take place whenever the sending and receivingunits or processing modules are co-active. O’Reilly and Munakata (2000) havedescribed hebbian learning as instantiating what they call model learning . Thefundamental computational objective of such unsupervised learning mechanismsis to enable the cognitive system to develop useful, informative models of theworld by capturing its correlational structure. As such, they stand in contrast withtask learning mechanisms, which instantiate the different computational objectiveof mastering specific input-output mappings (i.e., achieving specific goals) in thecontext of specific tasks through error-correcting learning procedures. Regardlessof how these two classes of learning mechanisms can be combined, the importantpoint to remember in the context of this framework is that model learningoperates whenever information processing takes place, whereas task learning onlyoperates in specific contexts defined by particular goals.L2.Learning is adaptation that specifically involves high-quality representations.
We assume that learning consists specifically of those adaptation processes thatinvolve high-quality, strong, stable representations. One way to characterize this
Implicit learning: A graded, dynamic perspective
20
notion is to appeal to another distinction offered by O’Reilly & Munakata (2000)— that between weight-based and activation-based processing. According toO’Reilly & Munakata, “Activation-based processing is based on the activation,maintenance, and updating of active representations to influence processing,whereas weight-based processing is based on the adaptation of weight values toalter input/output mappings” (p. 380). The main advantage of activation-basedprocessing is that it is faster and more flexible than weight-based processing.Speed and flexibility are both salient characteristics of high-level cognition.O’Reilly & Munakata further speculate that activation-based processing is one ofthe central characteristics of the frontal cortex, and suggest that this region of thebrain has evolved specifically to serve a number of important functions related tocontrolled processing, such as working memory, inhibition, executive control, andmonitoring or evaluation of ongoing behavior. To serve these functions,processing in the frontal cortex is characterized by mechanisms of activemaintenance through which representations can remain strongly activated for longperiods of time so as it make it possible for these representations to biasprocessing elsewhere in the brain.
O’Reilly and Munakata point out that a major puzzle is to understand how thefrontal cortex comes to develop what they call a “rich vocabulary of frontalactivation-based processing representations with appropriate associations tocorresponding posterior-cortical representations” (p. 382). Our framework doesnot solve this difficult chicken-and-egg problem, but simply suggests that earlylearning or development, which involve mostly weight-based processing,progressively results in the emergence of the strong, high-quality representationsthat allow activation-based processing and the ensuing flexibility to take place.Language and linguistic representations in general undoubtedly play a major rolein making activation-based processing possible.
L3.Learning has both direct and indirect effects.
Learning not only has direct effects, (i.e., changing the subjective experience thatcorresponds to the processing of a particular event and modifying the system’sresponse to that event), but it also has indirect effects on how (functionally orphysically) similar events are processed. This is again a natural consequence ofthe assumption that memory systems in general involve distributed,superpositional representations, such that all representations share manyprocessing units, and such that all processing units are involved in manyrepresentations. In such representational systems, changes to any particularrepresentation that might arise from learning will necessarily have indirect effects
Implicit learning: A graded, dynamic perspective
21
on related representations. Importantly, these indirect effects are mediated bychanges in the connection weights shared by the different representations in agiven module; in other words, they do not involve direct, simultaneousmodification of the corresponding representations. These indirect effects are thus,in our framework, not necessarily accompanied by awareness, because to beaccompanied by awareness, their origin and magnitude would have to beidentifiable by the agent.5.4 Assumptions about consciousnessSo far, we have spelled out a number of assumptions about information processing,representation, and learning. We are now ready to introduce our assumptions aboutconsciousness and its relationship to adaptation and learning processes. The centralideas that we would like to explore are (1) that the extent to which a particularrepresentation is available to consciousness depends on its quality, (2) that learningproduces, over time, higher-quality, adapted representations, and (3) that the functionof consciousness is to offer necessary control over those representations that arestrong enough to influence behavior, yet not sufficiently adapted that their influencedoes not require control anymore.
Insert Figure 1 about here
Figure 1 aims to capture these ideas by representing the relationships between qualityof representation (X-axis) on the one hand and (1) potency, (2) availability to control,(3) availability to subjective experience. We discuss the figure at length in thefollowing section. Let us simply note here that the X-axis represents a continuumbetween weak, poor-quality representations on the left and very strong, high-qualityrepresentations on the right., and that principle R3 (“Representations are constantlycausally efficacious”) is captured by the curve labeled “potency”, which assumes thatall representations, even weak ones, can influence behavior to some extent. Thegeneral form of the relationship between quality of representation and potency isassumed to be non-linear.
Two further points are important to keep in mind with respect to Figure 1. First,the relationships depicted in the Figure are intended to represent availability to somedimension of behavior or consciousness independently of other considerations. Many
Implicit learning: A graded, dynamic perspective
22
potentially important modulatory influences on the state of any particular module arethus simply not meant to be captured neither by Figure 1, nor by our framework as wepresent it here. Second, the figure is intended to represent what happens in each ofmany processing modules involved in any particular cognitive task. Thus, as hinted byassumptions P1-P4, at any point in time, there will be many such modules active, eachcontributing to some extent to behavior and to conscious experience; each modulatingthe activity of other modules. With these caveats in mind, let us now turn to our fiveassumptions about consciousness and its relationship with learning:
C1.Consciousness involves two dimensions: Subjective experience and control
As argued by many, and most cogently by Ned Block, consciousness involves atleast two separable aspects, namely access consciousness (A-consciousness) andphenomenal consciousness (P-consciousness). For Posner and Rothbart (Posner &Rothbart, 1998), awareness of the sensory world and voluntary control are the twomost important aspects of consciousness. According to Block (1995), “Aperceptual state is access-conscious roughly speaking if its content — what isrepresented by the perceptual state — is processed via that information processingfunction, that is, if its content gets to the Executive system, whereby it can beused to control reasoning and behavior.” (p. 234). In other words, whether a stateis A-conscious is defined essentially by the causal efficacy of that state; the extentto which it is available for global control of action. Control refers to the ability ofan agent to control, to modulate, and to inhibit the influence of particularrepresentations on processing. In our framework, control is simply a function ofpotency, as described in assumption C3. In contrast, P-consciousness refers to thephenomenal aspects of subjective experience: A state is P-conscious to the extentthat there is something it is like to be in that state. While the extent to whichpotency (i.e., availability to access consciousness), control, and subjectiveexperience (i.e., availability to phenomenal consciousness) are dissociable isdebatable, our framework suggests that these three aspects of consciousness areclosely related to each other.C2.Availability to consciousness correlates with quality of representation
This assumption is also a central one in our framework. It states that explicit,conscious knowledge involves higher quality memory traces than implicitknowledge. \"Quality of representation\designates several properties of memory traces, such as their relative strength inthe relevant information-processing pathways, their distinctiveness, or theirstability in time. Our assumption is consistent with the theoretical positionsexpressed by several different authors over the last few years. O’Brien & Opie
Implicit learning: A graded, dynamic perspective
23
(1999) have perhaps been the most direct in endorsing a characterization ofphenomenal consciousness in terms of the properties of mental representations indefending the idea that “consciousness equals stability of representation”, that is,that the particular mental contents that one is aware of at some point in timecorrespond to those representations that are sufficiently stable in time. Mathis &Mozer (1996) have also suggested that consciousness involves stablerepresentations, but have defined stability more technically than O’Brien & Opiehave, specifically by offering a computational model of priming phenomena inwhich stability literally corresponds to the state that a so-called dynamic“attractor” network reaches when the activations of a subset of its units stopschanging and settle into a stable, unchanging state.
A slightly different perspective on the notion of “quality of representation” isoffered by authors who emphasize not stability, but strength of representation asthe important feature through which to characterize availability to consciousness.One finds echoes of this position in the writings of Kinsbourne (1997), for whomavailability to consciousness depends on properties of representations such asduration, activation, or congruence. Importantly, for both O’Brien & Opie and forKinsbourne, the contents of subjective experience never depend onrepresentations entering a particular system in the brain — that is, consciousnessis conceived as essentially decentralized: Any region of the brain can contribute tothe contents of subjective experience so long as its representational vehicles havethe appropriate properties.
In Figure 1, we have represented the extent to which a given representation isavailable to the different components of consciousness (phenomenalconsciousness, access-consciousness/potency, and control) as functions of a singleunderlying dimension expressed in terms of the quality of this representation.Availability to access-consciousness is represented by the curve labeled“potency”, which expresses the extent to which representations can influencebehavior as a function of their quality. We simply assume that high-quality,strong, distinctive representations are more potent than weaker representations.“Availability to control processes” is represented by a second curve, so labeled.We simply assume that both weak and very strong representations are difficult tocontrol, and that maximal control can be achieved on representations that arestrong enough that they can begin to influence behavior in significant ways, yetnot so strong that have become utterly dominant in processing. Finally,availability to phenomenal experience is represented by the third curve, obtainedby convolving the other two. The underlying intuition, discussed in the context of
Implicit learning: A graded, dynamic perspective
24
assumption C4, is that which contents enter subjective experience is a function ofboth availability to control and of potency.
C3.Developing high-quality representations takes time
This assumption states that the emergence of high quality representations (seeassumption C2) in a given processing module takes time, both over training ordevelopment, as well as during processing of a single event. Figure 1 can thus beviewed as representing not only the relationships between quality ofrepresentation and their availability to the different components of consciousness,but also as a depiction of the dynamics of how a particular representation willchange over the different time scales corresponding to development, learning, orwithin-trial processing.
Both skill acquisition and development, for instance, involve the long-termprogressive emergence of high-quality, strong memory traces based on earlyavailability of weaker traces. Likewise, the extent to which memory traces caninfluence performance at any moment (e.g., during a single trial) depends both onavailable processing time, as well as on overall trace strength. We envision theseprocesses of change as operating on the connection weights between units in aconnectionist network. They can involve either task-dependent, error-correctingprocedures, or unsupervised procedures such as hebbian learning. In either case,continued exposure to exemplars of the domain will result in the development ofincreasingly congruent and strong internal representations that capture more andmore of the relevant variance. Although we think of this process as essentiallycontinuous, we distinguish three stages in the formation of such internalrepresentations (each depicted as separate regions in Figure 1): Implicitrepresentations, explicit representations, and automatic representations.
The first region, labeled “implicit cognition” in Figure 1, is meant tocorrespond to the point at which processing starts in the context of a single trial,or to some early stage of development or skill acquisition. In either case, thisstage is characterized by weak, poor-quality representations. A first importantpoint, embodied in assumption R3, is that representations at this stage are alreadycapable of influencing performance, as long as they can be brought to bear onprocessing together with other sources of constraints, that is, essentially throughmechanisms of associative priming and constraint satisfaction. A secondimportant point is that this influence is best described as \"implicit\relevant representations are too weak (i.e., not distinctive enough) for the systemas a whole to be capable of exerting control over them: You cannot control what
Implicit learning: A graded, dynamic perspective
25
you cannot identify as distinct from something else. One might even speculatethat what enables you to take control of an internal state is precisely the fact that itis capable of triggering responses in and of itself — a speculation that linkscontrol with action in a very direct way.
The second region of Figure 1 corresponds to the emergence of explicitrepresentations, defined as representations over which one can exert control. Inthe terminology of attractor networks, this would correspond to a stage duringlearning at which attractors become better defined — deeper, wider, and moredistinctive. It is also at this point that the relevant representations acquiregenerative capacity, in the sense that they now have accrued sufficient strength tohave the potential to determine appropriate responses when their preferredstimulus is presented to the system alone. Because awareness is partially tied tocontrol in our framework, one would thus also be aware both of these internalrepresentations and of their influence on our behavior. Because one is aware ofthese representations, one can then also possess metaknowledge about them, andrecode them in various different ways, for instance, as linguistic propositions.The third region involves what we call automatic representations, that is,representations that have become so strong that their influence on behavior cannot longer be controlled (i.e., inhibited). Such representations exert a mandatoryinfluence on processing. Importantly, however, one is aware both of possessingthem (that is, one has relevant metaknowledge) and of their influence onprocessing (see also Tzelgov, 1997), because availability to conscious awarenessdepends on the quality of internal representations, and that strong representationsare of high quality. In this perspective then, one can always be conscious ofautomatic behavior, but not necessarily with the possibility of control over thesebehaviors.
In our framework, skill acquisition, and development therefore involve acontinuum at both ends of which control over representations is impossible ordifficult, but for very different reasons: Implicit representations influenceperformance but cannot be controlled because they are not yet sufficientlydistinctive and strong for the system to even know it possesses them. This mightin turn be related to the fact that, precisely because of their weakness, implicitrepresentations cannot influence behavior on their own, but only in conjunctionwith other sources of constraints. Automatic representations, on the other hand,cannot be controlled because they are too strong, but the system is aware both oftheir presence and of their influence on performance.
Implicit learning: A graded, dynamic perspective
26
C4.The function of consciousness is to offer flexible, adaptive control over behavior
Our framework gives consciousness a central place in information processing, inthe sense that its function is to enable flexible control over behavior. Crucially,however, consciousness is not necessary for information processing, or foradaptation in general, thus giving a place for implicit learning in cognition. Webelieve this perspective to be congruent with theories of adaptation and optimalityin general.
Indeed, another way to think about the role of learning in consciousness is toask: \"When does one need control over behavior?\". Control is perhaps notnecessary for implicit representations, for their influence on behavior isnecessarily weak (in virtue of the fact that precisely because they are weak, suchrepresentations are unlikely to be detrimental to the organism even if they are notparticularly well-adapted). Likewise, control is not necessary for automaticrepresentations, because presumably, those representations that have becomeautomatic after extensive training should be adapted (optimal) as long as theprocesses of learning that have produced them can themselves be assumed to beadaptive. Automatic behavior is thus necessarily optimal behavior, except,precisely, in cases such as addiction or in laboratory situations where theautomatic response is manipulated to be non-optimal, such as in the Stroopsituation. Referring again to Figure 1, this analysis thus suggests that therepresentations that require control are the explicit representations that correspondto the central region of Figure 1: Representations that are strong enough that theyhave the potential to influence behavior in and of themselves (and hence that oneshould really care about, in contrast to implicit representations), but notsufficiently strong that they can be assumed to be already adapted, as is the casefor automatic representations. It is for those representations that control is needed,and, for this reason, it is of these representations that one is most aware of.Likewise, this analysis also predicts that the dominant contents of subjectiveexperience at any point in time consists precisely of those representations that arestrong enough that they can influence behavior yet weak enough that they stillrequire control. Figure 1 reflects these ideas by suggesting that the contents ofphenomenal experience depend both on the potency of currently activerepresentations as well as on their availability to control. Since availability tocontrol is inversely related to potency for representations associated withautomatic behavior, this indeed predicts weaker availability to phenomenalexperience of “very strong” representations as compared to “merely strong”
Implicit learning: A graded, dynamic perspective
27
representations. Such “automatic representations” therefore form what Mangan(1993) has called the “fringe of consciousness”. In other words, suchrepresentations can become conscious if appropriate attention is directed towardstheir contents — as in cases where normally automatic behavior (such as walking)suddendly becomes conscious because the normal unfolding of the behavior hasbeen interrupted (e.g., because I’ve stumbled upon something) — but they are notnormally part of the central focus of awareness nor do they require consciouscontrol.
While the dominant contents of subjective experience can thus be viewed asreflecting the activity of the topmost module in the constantly evolving loosehierarchy of processing modules involved in any particular aspect of informationprocessing, it is also important to note that we assume, in contrast to the positionexpressed by Perruchet & Vinter, that complex representations depend on thecontinued activation of their more elementary components. In other words, whilelearning certainly results in the elaboration of progressively more complexrepresentations, it neither prevents their components from contributing tosubjective experience nor does it eliminate their influence on ongoing processing.This therefore opens the door for the continued — but attenuated, indirect —expression of the representations associated with these lower-level modules.C5.Learning shapes conscious experience
This assumption, which we adapt from Perruchet & Vinter (this volume) is acorollary of assumption C4: If the function of consciousness is to offer flexible,adaptive control over behavior, then its contents — the way it reflects the world— should necessarily be shaped by learning so that, at any moment, thesecontents tend to reflect precisely those aspects of the situation that most requirecontrol. This assumption allows us to relate two central aspects of consciousnessthat have often been considered as independent: subjective experience and controlof action —or phenomenal vs. access consciousness.5.5. Ways for knowledge to be implicitIn our framework, we emphasized quality of representation as a central dimensionthrough which to account for which representations are likely to enter consciousawareness. It should be clear, however, that we take quality of representation as anecessary , but not sufficient condition, for conscious awareness. In particular, our framework remains mute with respect to the fate of the high-quality, strongrepresentations that characterize explicit, conscious cognition, short of claiming that it
Implicit learning: A graded, dynamic perspective
28
is these representations that are most likely to form the contents of subjectiveexperience. Whether these representations actually enter conscious experience is yet adifferent story — one in which processes of attention and processes of integrationundoubtedly play a central role. In this respect, our framework is not inconsistent withrecent proposals that emphasize the importance of such processes in the formation ofsubjective experience. One such recent proposal has been put forward by Dehaeneand Naccache (2001). These authors, based on Baars’ notion of “global workspace”(Baars, 1988), propose that conscious awareness depends on the extent to which thecontents of the many domain-specific unconscious processing modules that make upour brain can be made accessible globally through specific, dedicated, long-distanceneural pathways that interconnect the modules and specific regions of the brain (i.e.,essentially prefrontal cortex, anterior cingulate and other regions connected to both).Availability to the global workspace thus depends on both bottom-up (i.e., inputstrength) and top-down (i.e. attention) factors. When these two conditions exist, thecontents of those modules that connect to the neural workspace would then enter inthe stable, resonant, or synchronous states that are assumed to correlate withconscious awareness.
Kanwisher (2001) also discusses the conditions under which particularrepresentations will enter conscious awareness, and notes that activation strengthalone, while perhaps necessarily, is certainly not sufficient. Like Dehaene andNaccache, Kanwisher suggests that awareness also depends on “informationalaccess”, that is, on the fact that other parts of the brain/mind have access to therelevant representations. Kanwisher also suggests the accessibility can change overtime as a result of practice — a point that we very much agree with —, and that animportant further factor in determining availability to consciousness is what she callsthe “type/token” distinction, that is, the fact that awareness of some perceptualattribute not only requires a strong corresponding representation, but also“individuation of that perceptual information as a distinct event” (p. 107). In otherwords, the relevant representation has to be accompanied by relevant metaknowledge— a point discussed in detail by Dienes and Perner (1999).
Our own framework leaves open four distinct possibilities for knowledge to beimplicit.
First, we assume that the knowledge that is embedded in the connection weightswithin and between processing modules can never be directly available to consciousawareness and control. This is simply a consequence of the fact that we assume thatconsciousness necessarily involves representations (patterns of activation over
Implicit learning: A graded, dynamic perspective
29
processing units). Because weight-based knowledge is not representational in thisspecific sense, it follows that it can never directly contribute to the contents ofconscious experience. This knowledge will, however, shape the representations thatdepend on it, and its effects will therefore detectable — but only indirectly, and onlyto the extent that these effects are sufficiently marked in the correspondingrepresentations.
Second, we assume that to enter conscious awareness, a representation needs to bea sufficiently high-quality in terms of strength, stability in time, or distinctiveness.Weak representations are therefore poor candidates to enter conscious awareness.This, however, as we repeatedly emphasized, does not necessarily imply that theyremain causally inert, for they can influence further processing in other modules, evenif only weakly so. Note that this aspect of our framework differs both from theassumptions put forward by O’Brien and Opie (1999) and from those embodied inPerruchet and collaborators (Perruchet, Vinter, & Gallego, 1997; Perruchet & Vinter,1998; this volume).
Third, a representation can be strong enough to enter conscious awareness, butfailed to be recognized as relevant to the particular situation that is currentlyunfolding. This case corresponds almost exactly with Kanwisher’s “type/token”distinction, and also with aspects of Dienes & Perner’s analysis of the differencesbetween implicit and explicit knowledge. Conscious contents, indeed, have to belinked together in a coherent manner before they can be made available globally forconscious report and for the control of action. One should therefore be very careful indistinguish between cases involving awareness of the intention of initiating somebehavior, awareness of the fact that the behavior is taking place, awareness of thecauses of the behavior, and awareness of the effects of the behavior. There are thusmany opportunities for a particular conscious content to remain, in a way, implicit, notbecause its representational vehicle does not have the appropriate properties, butbecause it fails to be integrated with other conscious contents. Dienes & Perner (thisvolume) offer an insightful analysis of the different ways in which what we call high-quality representations can remain implicit.
Finally, a representation can be so strong that its influence can be no longer becontrolled. In theses cases, it is debatable whether the knowledge should be taken asgenuinely unconscious, because they certainly can become fully conscious as long asappropriate attention is directed to them, but the point is that such very strongrepresentations can trigger and support behavior without conscious intention andwithout the need for conscious monitoring of the unfolding behavior.
Implicit learning: A graded, dynamic perspective
30
Another way to think about these different ways for knowledge to be implicit is toconsider the various mechanisms of change suggested by O’Reilly & Munakata.Recall that these authors distinguish between weight-based processing and activation-based processing. Weight-based processing in turn involves model learning(subserved by hebbian-like learning mechanisms) and task learning (subserved byerror-correcting learning procedures). From the perspective developed here,activation-based processing and learning will always tend to be associated withawareness — even though it might often occur that conscious contents fail to beassociated with relevant metaknowledge and therefore remain implicit. Modellearning, in contrast, corresponds to the clearest case of implicit learning, to the extentthat it is assumed to be a mandatory consequence of information processing. Suchlearning therefore never depends on the intentions or goals of the agent, and itseffects, because they are very gradual, can be expressed in behavior before theybecome available to awareness. Task learning, by contrast, is necessarily intentional,and therefore more likely to shape representations in ways that are directly consistentwith the current goals of the agent.
6. ImplicationsIn this section, we offer a necessarily brief and sketchy set of examples where wehave found our framework helpful in terms of understanding empirical phenomenasuch as priming, skill acquisition, automaticity, development, or the interpretation ofdissociations in neuropsychology. This short review also gives us the opportunity tolink our framework with similar previous accounts of these phenomena and to furthercontrast our own proposal with other positions.6.1. Priming.In a recent paper, Becker et al. (1997) describe an attractor, neural-network model ofboth short- and long-term priming effects that accounts for a large variety of primingphenomena as the result of an automatic process of incremental learning that is basedon the same information processing and representational principles that we have justoutlined. Becker and colleagues showed that semantic priming can be construed as theautomatic deepening of the basin of attraction \"of the semantic space for both primesand related targets, and that this effect should primarily manifest itself on semantic-retrieval tasks\" (p. 1062). Their model accurately predicts that performing a semantictask on a target is influenced by having previously performed a similar task on a
Implicit learning: A graded, dynamic perspective
31
semantically related prime, even if a number of intervening words are presentedbetween prime and target. Importantly, it also predicts low or null priming effectswhen long-term semantic effects are tested through a lexical decision task (see alsoJoordens & Becker, 1997) or when the processing task performed with the primes isnot semantic (Friedrich, Henik, & Tzelgov, 1991; Kaye & Brown, 1985; Smith,Theodor, & Franklin, 1983). These results, as well as the successful simulation work,are compatible with our own assumptions in that they suggest (1) that learning isassumed to be a mandatory consequence of processing, (2) that the effects of learningare particularly focused on those representational features that are relevant to theprocessing task (and which therefore produce specific experiences); and (3) that theseeffects are not limited to their most direct consequences —in this case, the episodicrecollection that a prime has been presented— but may also produce a host of indirect(priming) effects that are not necessarily mediated by conscious recollection of itscause.
6.2. Implicit learningIf priming can be cast as a form of implicit learning, as Becker et al. (1997) suggest, itseems that implicit learning can likewise be depicted as a form of complex relationalpriming. Indeed, while our framework emphasizes that learning results fromconscious experience, it also makes it clear that the effects of learning need not belimited to modifying conscious experience. In particular, two important assumptionsembodied in our framework are (1) that adaptation occurs as a mandatoryconsequence of processing, and (2) that learning has both direct and indirect effects.Three consequences of these assumptions are (1) that learning can occur withoutintention to learn, (2) that the changes resulting from learning can remain unconsciousat the time of learning, and (3) that such changes can influence subsequent processingeven in the absence of awareness that this is so.
Because of the intricate methodological issues involved, it has proved ratherdifficult to gather supporting evidence for any of these three claims. It is alwaysdifficult to assess exactly what participants in an experiment involving learning areactually intending to do. Implicit learning studies have often tried to circumvent thisproblem by exposing participants to very complex settings in which learning wouldnot be expected to improve through an intentional orientation, but this strategy has notbeen frequently used (see Jiménez, Méndez, and Cleeremans, 1996, for one example).However, indirect evidence that people can effectively learn without intending to doso has been obtained through some recent experiments that use dual-cue paradigms, inwhich the existence of a perfect and explicit predictor of the relevant stimulus
Implicit learning: A graded, dynamic perspective
32
dimension can be taken to prevent the deliberate search for more complexcontingencies (Cleeremans, 1997; Jiménez & Méndez, in press). The fact that learningof these complex contingencies can be obtained even under these dual-cue conditionsprovides us with a clear indication that this learning proceeds regardless ofparticipants' intention to learn. Importantly, however, these results should not be takenas indicating that learning is completely unselective. Indeed, several recentexperimental results (Jiménez & Méndez, 1999; 2001; see also Jiang & Chun, inpress) convincingly indicate that learning is selectively obtained for those particularfeatures that are relevant to the task(s) at hand and, hence, that learning is deeplymodulated by the attentional variables that ultimately determine the learner'sexperiences.
As we have repeatedly stated, the fact that learning depends on the consciousexperiences of the learner does not necessarily entail that all learning should beconscious at the moment of learning, or that they should be conscious to produce anyeffect on performance. The unconscious nature of the knowledge acquired duringtraining on a sequence learning task has been examined by us in previous studies (e.g.,Jiménez et al., 1996) and it has been recently investigated by Destrebecqz andCleeremans (in press) by adapting Jacoby’s process dissociation procedure.
In a typical sequence learning situation (see Clegg, DiGirolamo, & Keele, 1998),participants are asked to react to each element of a sequentially structured visualsequence of events in the context of a serial reaction time task. On each trial, subjectssee a stimulus that appears at one of several locations on a computer screen and areasked to press as fast and as accurately as possible on the key corresponding to itscurrent location. Unknown to them, the sequence of successive locations follows arepeating pattern (Nissen & Bullemer, 1987), and participants learn this pattern, asshowed by a progressive decrease of their reaction times, that increase dramaticallywhen the sequential structure of the material is modified (Cohen, Ivry, & Keele, 1990;Curran & Keele, 1993; Reed & Johnson, 1994).
This learning, however, often fails to be expressed through verbal reports,(Willingham, Nissen, & Bullemer, 1989; Curran & Keele, 1993) — a dissociation thathas led many authors to consider learning in this situation to be implicit. However,many of the relevant studies have been criticized on methodological grounds thatwould be too long to review in this chapter (but see Cleeremans, Destrebecqz, &Boyer, 1998 for a detailed overview). Suffice it to say that many of the relevantmethodological difficulties stem from the fact that most empirical paradigms through
Implicit learning: A graded, dynamic perspective
33
which implicit learning has been studies have assumed that one take specific taskswith either implicit or explicit processing.
To overcome these methodological difficulties, Destrebecqz & Cleeremans (inpress) sought to adapt Jacoby’s process dissociation procedure (e.g., Jacoby, 1991) tothe study of sequence learning. Subjects were first trained, under the incidentallearning conditions typical of implicit learning studies, on a second-order conditionalsequence. This training occurred under two conditions defined by the length of theresponse-to-stimulus interval (RSI): One group of participants was trained with astandard RSI of 250 msec, and another was trained with an RSI of 0 msec. For thislatter group, the next stimulus therefore appeared on the screen as soon as theprevious one had been responded to. Consistently with the ideas embodied inassumption C3 above, we hoped that reducing the time available for processing wouldselectively impair the development of strong, explicit representations of the linksbetween the temporal context set by previous elements of the sequence and thelocation of the next stimulus.
To find out about participants’ explicit knowledge of the material, Destrebecqz andCleeremans asked them to perform two generation tasks and a recognition task. Thegeneration task was adapted from Jacoby’s PDP, and consisted of both an inclusiontask as well as an exclusion task. In inclusion, participants had to generate a sequenceof 96 elements that resembled the training sequence. They were told to base theirsequence either on conscious recollection or to guess. Both conscious andunconscious processes can therefore contribute to performance in inclusion. Inexclusion, participants were again told to generate a sequence of 96 elements, but thistime they were told to produce a sequence that was as different as possible from thetraining sequence. By assumption, the only way participants can perform thisexclusion task successfully is by recollecting the location of the next stimulus and byselecting another location. Failure to exclude can thus be interpreted as reflecting theinfluence of implicit knowledge. In this condition thus, and in contrast to whathappens in inclusion, conscious and unconscious components of performance actagainst each other. Finally, in recognition, participants were presented with 24sequences of three elements, only 12 of which had actually been part of the trainingsequence. For each, they had to indicate the extent to which they believed it was partof the training sequence on a 6-points scale.
The results indicated that while both groups of participants exhibited some explicitknowledge of the material through the inclusion task, only people trained with an RSIof 250 msec were able to perform successfully in the exclusion task. People trained
Implicit learning: A graded, dynamic perspective
34
with an RSI of 0 msec indeed continued to generate material from the trainingsequence in spite of instructions to the contrary. Further, only participants trained witha 250 msec RSI were able to perform above chance on the final recognition task.When applied to these data, our framework suggests the following interpretation:People trained with an RSI were given more opportunities to develop and linktogether high quality memory traces than people in the no RSI condition. Becauseawareness depends in part on the quality of stored memory traces, the former willtherefore tend to acquire more explicit knowledge than the latter. Importantly, “noRSI” participants do acquire relevant knowledge about the sequence — but in theform of weaker memory traces that are only capable of influencing responses whencontextual information is simultaneously available. This knowledge can thus beexpressed in the SRT task as well as in the generation tasks because in both cases,responses can be determined based jointly on an external stimulus (self-generated inthe case of the generation tasks, or produced by the experimental software in the SRTtask) and the relevant memory traces. Because these traces are weak and becausecontrolled processing (and hence awareness) requires high-quality traces to beavailable, their influence on performance remains undetected and controlledresponding made difficult. The relevant sequential knowledge therefore cannot beinhibited when the generation task is performed under exclusion conditions. Similarly,during recognition, weak memory traces do not allow successful discriminationbetween old and novel sequences in the absence of perceptual and motor fluency, aswas the case in Destrebecqz & Cleeremans’s study.
6.3 Skill acquisition and automaticity Skill acquisition refers to extended periods of exposure to a particular domain duringwhich learning occurs. It might involve learning how to use musical instrument,learning to master a particular athletic skill, or learning natural language. In ourframework, skill acquisition thus involves a graded continuum expressed in terms ofthe relative strength of the underlying representations. This continuum involves weak,implicit representations when learning starts, and very strong, high-qualityrepresentations when training ends.
Automaticity has often been associated with lack of availability to consciousexperience, but some authors (i.e., Tzelgov, 1997) have proposed that the definingfeature of automatic behavior should simply be its ballistic properties, that is, the factthat once initiated, execution of the behavior can no longer be inhibited until
Implicit learning: A graded, dynamic perspective
35
completed. We very much agree with this position: In our framework, the strongrepresentations associated with automatic behavior are available to subjectiveexperience and form what one could call, along with Mangan, the “fringe” ofconsciousness. In other words, such representations can become conscious ifappropriate attention is directed towards their contents — as in cases where normallyautomatic behavior (such as walking) suddendly becomes conscious because thenormal unfolding of the behavior has been interrupted (e.g., because I’ve stumbledupon something) — but they are not normally part of the central focus of awarenessnor do they require conscious control. This is reflected in our framework by assumingthat the contents of phenomenal experience depend both on the potency of currentlyactive representations as well as on their availability to control. Since availability tocontrol is inversely related to potency for automatic representations, this indeedpredicts weaker availability to phenomenal experience of very strong representationsas compared to merely strong representations.
Our framework also predicts that very strong representations are left in place; thatis, they become active whenever their preferred stimulus is present. This suggests thatwhat happens over the course of learning a skill is that additional novel ways ofinhibiting or otherwise modulating the effects of these very strong representations arefound through processes of learning. Consider what happens when you learn to playthe piano, for instance. As Karmiloff-Smith (1992) points out, one goes from effortfulprogramming of every movement to a stage where entire sequences of movements canbe executed all at once, and then to a later stage where genuine flexibility is achieved.Our suggestion here is that subjective experience at each stage simply reflects thecontents of the processing modules that currently contain the most abstractrepresentation of the stimulus. Ability to control the influence of the contents oflower-level modules is thus progressively lost during skill acquisition, butimportantly, these contents are still constitutive of subjective experience, if onlythrough their role in supporting higher-level representations.6.4 DevelopmentThe notion that development involves continuous changes in the strength or quality ofunderlying representations is central in many accounts of various relevant phenomena.For instance, McClelland and Jenkins (McClelland & Jenkins, 1991)’s connectionistmodel of developmental changes in the balance beam task is rooted in the idea thatexperience at solving balance beam problems results in the progressive differentialstrengthening of the internal representations of the weight and distance information.The relatively systematic sequence of stages observed through development in the
Implicit learning: A graded, dynamic perspective
36
emergence of mastery on this task, which exhibits various patterns of ability to solvespecific problems, can thus simply be accounted for by competition between theinformation-processing pathways corresponding to each dimension. In other words,this account of the emergence of skilled behavior at mastering balance-beam tasks isentirely strength-based.
Munakata and collaborators (Munakata, et al., 1997) have likewise proposed anovel account of the development of object permanence during infancy (see alsoMareschal, Plunkett, & Harris, 1999) in which the notion of strength of representationalso plays a central role. Classical theories of object permanence assume that at someearly point during development, children “acquire the concept” that objects continueto exist when out of view. The crucial point here is that this knowledge is assumed tobe of a conceptual nature: Children are taken to be constantly developing explicittheories about their environment, and their theories can be described as consisting of aset of all-or-none beliefs about the way the world works. In stark constrast, Munakataet al. suggested that the progressive emergence of appropriate anticipatory responsesin situations where a moving object temporarily disappears behind a screen canemerge simply out of the operation of prediction-driven mechanisms such asinsantiated in the Simple Recurrent Network (Elman, 1990; see also Cleeremans,Servan-Schreiber, & McClelland, 1989). Most importantly, Munakata et al. showedhow the model, when trained on such a prediction task, progressively developsstronger, higher-quality representations of the object while it is hidden, and how thisprogressive strengthening of the model’s internal representations can be related to thedevelopment of knowledge about object permanence. Another important aspect of thiswork was the demonstration that the very same principles — strength of internalrepresentation — could account for observed dissociations between differentmeasures (e.g., looking times vs. reaching behavior) of children’s ability to exhibitknowledge of object permanence. It is interesting to note that in many ways, thedebates elicited in the developmental literature by the empirical findings related toobject permanence mirror those taking place in the field of implicit learning aboutwhether or not subjects are best described as “knowing the rules of the grammar”.A central idea that both of these models illustrate is that continuous changes alongone dimension can exert non-linear effects on the overall behavior of the system wheninteractions between several dimensions are considered. In other words, all-or-nonebehavior can be rooted in continuous, graded changes in some relevant underlyingdimension.
Implicit learning: A graded, dynamic perspective
37
Finally, with respect to the development of explicit, conscious representations incognitive systems, our framework can also be linked in interesting ways with theprocesses of representational redescription envisioned by Karmiloff-Smith(Karmiloff-Smith, 1992) as the main process of change during development. A crucialclaim embodied in the assumptions that underpin the notion of representationalredescription is that learning is success-driven, that is, behavioral mastery of aparticular skill does not constitute a signal for learning to stop but rather a signal forfurther learning to occur — on the internal representations through which mastery wasachieved. Representational description, according to Karmiloff-Smith, is a “…process of ‘appropriating’ stable states to extract the information they contain, whichcan then be used more flexibly for other purposes” (p. 25). Thus, representationschange over development in such a manner that previously implicit dimensions of theproblem — which are sufficient to achieve behavioral success — progressivelybecome explicit and hence available for global control of action and for verbal report.Finally, our framework is also congruent with the idea that modules, in general, are aproduct of learning and development rather than their starting point.
7. Disscussion: What we leave behind In this chapter, we have attempted to outline a framework that offers a clear functional
role to consciousness by linking conscious awareness with adaptation in general, andwith learning in particular. We have argued that if we take consciousness as the onlymechanism through which flexible control can be achieved over action, then it followsthat learning should be the most important factor that determines the contents ofconscious experience. Learning thus shapes consciousness, and consciousness, in turn,reflects the adapted appreciation of the dynamics of the current situation that isnecessary to make flexible control over action possible (see also Perruchet & Vinter,this volume). Our framework as it stands, however, does not address how the contentsof consciousness are shaped by experience; it merely suggests the conditions underwhich representations are most likely to become part of conscious experience, and,importantly for our purposes, it also roots the emergence of conscious awareness intothoroughly subsymbolic mechanisms.
Further, our framework does not assume that there exists a strong distinctionbetween conscious and non-conscious aspects of cognition. Rather, it assumes thatconscious and unconscious aspects of cognition are simply that — aspects of a singleset of underlying neural mechanisms. Again, this position does not deny — far from it— that there are qualitative differences between conscious and unconscious
Implicit learning: A graded, dynamic perspective
38
computations, but simply emphasizes that such differences are rooted in the non-linearproperties of otherwise graded, continuous representation and processing systems.The most important implication of these assumptions in the context of implicitlearning research is that our framework leaves open the possibility for change to occurwithout intention and without concurrent awareness that change is taking place.“What we leave behind”, then, is a large set of unanswered questions about the fateof what we have called “explicit representations” — those representations that weassume constitute the best candidates to form the contents of phenomenal experience.However, we hope to have convinced readers (1) that understanding conscious(symbolic) cognition necessarily involves rooting this understanding in an analysis ofimplicit (subsymbolic) cognition, and (2) that understanding processes of learning isfundamental for any theory of consciousness. In this respect, the study of implicitlearning has a bright future, for it is through the development of sensitive paradigmsthrough which to explore the differences between conscious and unconsciouscognition that one can best contribute to the search for the neural correlates ofconsciousness.
Implicit learning: A graded, dynamic perspective
39
Footnotes1
This does not necessarily imply that artificial consciousness is not possible, butsimply that the relevant processes cannot consist simply of symbol manipulation.
Implicit learning: A graded, dynamic perspective
40
AcknowledgementsAxel Cleeremans is a research associate with the National Fund for ScientificResearch (Belgium). This work was supported by a grant from the Université Libre deBruxelles in support of IUAP program P4/19, by grant HPRN-CT-2000-00065 fromthe European Commission, and by DGES Grant PB97-0525 from the Ministerio deEducación y Cultura (Spain) to Luis Jiménez.
Implicit learning: A graded, dynamic perspective
41
References Ader, R., & Cohen, N. (1985). CNS-immune system interactions: Conditioning Phenomena.Behavioral and Brain Sciences , 8 , 379-394.Anderson, J. R., & Memory, L. a. (1995). Learning and Memory . New York: Wiley.Atkinson, A., Thomas, M., & Cleeremans, A. (2000). Consciousness: Mapping the theoreticallandscape. Trends in Cognitive Sciences, 4 (10), 372-382.Atkinson, R. C., & Shiffrin, R. M. (1971). The control of short-term memory. Scientific American, 224 ,82-90.
Baars, B. (1988). A Cognitive Theory of Consciousness . Cambridge: Cambridge University Press.Becker, S., Moscovitch, M., Behrmann, M., & Joordens, S. (1997). Long-term semantic priming: Acomputational account and empirical evidence. Journal of Experimental Psychology: Learning, Memory and Cognition, 23 , 1059-1082.Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences,18 , 227-287.Broadbent, D. E., Fitzgerald, P., & Broadbent, M. H. P. (1986). Implicit and explicit knowledge in thecontrol of complex system. British Journal of Psychology, 77 , 33-50.Chalmers, D. J. (1996). The conscious mind: In search of a fundamental theory : Oxford UniversityPress.
Clark, A., & Karmiloff-Smith, A. (1993). The cognizer’s innards: A psychological and philosophicalperspective on the development of thought. Mind and Language, 8 , 487-519.Cleeremans, A. (1997). Principles for implicit learning. In D. C. Berry (Ed.), How implicit is implicit learning? (pp. 195-234). Oxford: Oxford University Press.Cleeremans, A., Destrebecqz, A., & Boyer, M. (1998). Implicit learning : News from the front. Trends in Cognitive Sciences, 2 , 406-416.Cleeremans, A., Servan-Schreiber, D., & McClelland, J. L. (1989). Finite state automata and simplerecurrent networks. Neural Computation, 1 , 372-381.Clegg, B. A., DiGirolamo, G. J., & Keele, S. W. (1998). Sequence learning. Trends in Cognitive Science, 2 , 275-281.Cohen, A., Ivry, R. I., & Keele, S. W. (1990). Attention and structure in sequence learning. Journal of Experimental Psychology : Learning, Memory and Cognition, 16 , 17-30.Crick, F., & Koch, C. (1995). Are we aware of neural activity in primary visual cortex? Nature, 375 ,121-123.
Curran, T., & Keele, S. W. (1993). Attentional and nonattentional forms of sequence learning. Journal of Experimental Psychology : Learning, Memory and Cognition, 19 , 189-202.Dehaene, S., & Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: Basicevidence and a workspace framework. Cognition, 79 , 1-37.Dennett, D. C. (1991). Consciousness Explained . Boston, MA.: Little, Brown & Co.Dennett, D. C. (1996). Kinds of Minds: Towards an understanding of Consciousness : Basic Books.Destrebecqz, A., & Cleeremans, A. (in press). Can sequence learning be implicit? New evidence withthe Process Dissociation Procedure. Psychonomic Bulletin & Review .Dienes, Z., & Perner, J. (1999). A theory of implicit and explicit knowledge. Behavioral and Brain Sciences, 22 , 735-808.Dienes, Z. & Perner, J. (this volume). A theory of the implicit nature of implicit learning.Dretske, F. (1988). Explaining behavior . Cambridge, MA: MIT Press.Dulany, D. E. (1997). Consciousness in the explicit (deliberative) and implicit (evocative). In J. D.Cohen & J. W. Schooler (Eds.) Scientific approaches to the study of consciousness. (pp. 179-212).Mahwah, NJ.: Lawrence Erlbaum Associates.
Elbert, T., Pantey, C., Wienbruch, C., Rockstroh, B., & Taub, E. (1995). Increased cortical
Science, 270 , 305-307.representation of the fingers of the left hand in string players. Implicit learning: A graded, dynamic perspective
42
Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14 , 179-211.Eriksson, P. S., Perfilieva, E., Björk-Eriksson, T., Alborn, A.-M., Nordborg, C., Peterson, D. A., &Gage, F. H. (1998). Neurogenesis in the adult hippocampus. Nature Medicine, 4 (11), 1313-1317.Farah, M. J. (1994). Visual perception and visual awareness after brain damage: A tutorial overview. InC. Umiltà & M. Moscovitch (Eds.), Attention and Performance XV: Conscious and nonconscious information processing (pp. 37-76). Cmabridge, MA: MIT Press.Farah, M. J. (1998). Why does the somatosensory homonculus have hands next to face and feet next togenitals: A hypothesis. Neural Computation, 10 (8), 1983-1985.French, R. M. (2000). The Turing test: The first 50 years. Trends in Cognitive Sciences, 4 (3), 115-122.Friedrich, F. J., Henik, A., & Tzelgov, J. (1991). Automatic processes in lexical access and spreadingactivation. Journal of Experimental Psychology: Human Perception and Performance , 17 , 792-806.Frith, C., Perry, R., & Lumer, E. (1999). The neural correlates of conscious experience : Anexperimental framework. Trends in Cognitive Sciences, 3 , 105-114.Fodor, J. (1975). The Language of Thought. New York, NY: Harper & Row Publishers Inc.Grossberg, S. (1999). The link between brain learning, attention, and consciousness. Consciousness and Cognition, 8 , 1-44.Holender, D. (1986). Semantic activation without conscious activation in dichotic listening, parafovealvision, and visual masking : A survey and appraisal. Behavioral and Brain Sciences, 9 , 1-23.Holender, D. (1992). Expectancy effects, congruity effects, and the interpretation of response latencymeasurement. In J. Alégria, D. Holender, J. J. d. Morais, & M. Radeau (Eds.), Analytic approaches to human cognition (pp. 351-375). Amsterdam: Elsevier.Jacoby, L. L. (1991). A process dissociation framework: Separating automatic from intentional uses ofmemory. Journal of Memory and Language , 30 , 513-541.Jiang, Y., & Chun, M. (in press). Selective attention modulates implicit learning. Quarterly Journal of Experimental Psychology. Jiménez, L., Méndez, C., & Cleeremans, A. (1996). Comparing direct and indirect measures ofsequence learning. Journal of Experimental Psychology: Learning, Memory, and Cognition , 22 , 948-969.
Jiménez, L., & Cleeremans, A. (1999). Fishing with the wrong nets: How the implicit slips through therepresentational theory of mind. Behavioral and Brain Sciences, 22 , 771.Jiménez, L., & Méndez, C. (1999). Which attention is needed for implicit sequence learning? Journal of Experimental Psychology: Learning, Memory, and Cognition , 25 , 236-259.Jiménez, L. & Méndez, C. (in press). Implicit sequence learning with competing explicit cues.Quarterly Journal of Experimental Psychology (A) .Joordens, S., & Becker, S. (1997). The long and short of semantic priming effects in lexical decision.Journal of Experimental Psychology: Learning, Memory, and Cognition , 23 , 1083-1105.Cognition, 79 , 89-113.Kanwisher, N. (2001). Neural events and perceptual awareness. Karmiloff-Smith, A. (1992). Beyond modularity : A developmental perspective on cognitive science . Cambridge: MIT Press.
Kaye, D. B., & Brown, S. W. (1985). Levels and speed of processing effects on word analysis. Memory and Cognition, 13 , 425-434.Kinsbourne, M. (1997). What qualifies a representation for a role in consciousness? In J. D. Cohen & J.W. Schooler (Eds.), Scientific Approaches to Consciousness (pp. 335-355). Mahwah, NJ: LawrenceErlbaum Associates.Klein, S. B. (1991). Learning: Principles and Applications : McGraw Hill.Maguire, E. A., Gadian, D. G., Johnsrude, I. S., Good, C. D., Ashburner, J., Frackowiak, R. S., & Frith,C. D. (2000). Navigation-related structural change in the hippocampi of taxi drivers. Proceedings of the National Academy of Sciences of the U.S.A . , 10 , 1073.Mangan, B. (1993). Taking phenomenology seriously: The \"fringe\" and its implication for cognitiveresearch. Consciousness and Cognition , 2 , 89-108.Mareschal, D., Plunkett, K., & Harris, P. (1999). A computational and neuropsychological account of
Developmental Science, 2 , 306-317.object-directed behaviours in infancy. Mathis, W. D., & Mozer, M. C. (1996). Conscious and unconscious perception: A computationaltheory. Paper presented at the Proceedings of the Eighteenth Annual Conference of the Cognitive Science Society, Hillsdale, N.J.
Implicit learning: A graded, dynamic perspective
43
McClelland, J. L. (1979). On the time-relations of mental processes : An examination of systems incascade. Psychological Review, 86 , 287-330.McClelland, J. L., & Jenkins, E. (1991). Nature, nurture, and connectionism: Implications forconnectionist models of development. In K. v. Lehn (Ed.), Architectures for Intelligence — The Twenty-second (1988) Carnegie Symposium on Cognition . Hillsdale, NJ: Lawrence ErlbaumAssociates.
McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementarylearning systems in the hippocampus and neocortex: Insights from the successes and failures ofconnectionist models of learning and memory. Psychological Review, 102 , 419-457.Munakata, Y., McClelland, J. L., Johnson, M. H., & Siegler, R. S. (1997). Rethinking infantknowledge: Toward an adaptive process account of successes and failures in object permanencetasks. Psychological Review, 10 (4), 686-713.Nissen, M. J., & Bullemer, P. (1987). Attentional requirement of learning: Evidence from performancemeasures. Cognitive Psychology, 19 , 1-32.O'Brien, G., & Opie, J. (1999). A connectionist theory of phenomenal experience. Behavioral and Brain Sciences, 22 , 175-196.O'Reilly, R. C., & Munakata, Y. (2000). Computational Explorations in Cognitive Neuroscience: Understanding the Mind by Simulating the Brain . Cambridge, MA.: MIT Press.Perruchet, P., & Vinter, A. (1998). Learning and development: The implicit knowledge assumptionreconsidered. In M. A. Stadler & P. A. Frensch (Eds.), Handbook of implicit learning (Vol. 15, pp.495-531): Sage Publications.
Perruchet, P., Vinter, A., & Gallego, J. (1997). Implicit learning shapes new conscious percepts andrepresentations. Psychonomic Bulletin and Review, 4 , 43-48.Perruchet, P. & Vinter, A. (this volume). The self-organizing consciousness: A framework for implicitlearning.
Posner, M. I., & Rothbart, M. K. (1998). Attention, self-regulation, and consciousness. Philosophical Transactions of the Royal Society B, 353 , 1915-1927.Reed, J., & Johnson, P. (1994). Assessing implicit learning with indirect tests: Determining what islearned about sequence structure. Journal of Experimental Psychology: Learning, Memory andCognition, 20 , 585-594.Rosenthal, D. (1986). Two concepts of consciousness. Philosophical Studies , 94 , 329-359.Rosenthal, D. (1997). A theory of consciousness. In N. Block, O. Flanagan, & G. Güzeldere (Eds.), The Nature of Consciousness: Philosophical Debates. Cambridge, MA: MIT Press. Rumelhart, D. E., & McClelland, J. L. (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Volume 1: Foundations . Cambridge, MA: MIT Press.Schacter, D. L. (1989). On the relations between memory and consciousness: Dissociable interactionsand conscious experience. In H. L. Roediger and F. I. M. Craik (Eds.), Varieties of Memory and Consciousness: Essays in Honour of Endel Tulving (pp. 355-389). Mahwah, NJ: Lawrence ErlbaumAssociates.
Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences , , 417-457.3The rediscovery of the mind . Cambridge, MA.: MIT Press.Searle, J. R. (1992). Searle, J.R. (1999). Chinese Room Argument. In R.A. Wilson and F.C. Keil (Eds.) The MIT Encyclopedia of the Cognitive Sciences (pp.115-116). Cambridge, MA: The MIT press.Shanks, D. R., & St. John, M. F. (1994). Characteristics of dissociable human learning systems.Behavioral and Brain Sciences, 17 , 367-447.Smith, M. C., Theodor, L., & Franklin, P. E. (1983). The relationship between contextual facilitationand depth of processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 9 ,697-712.
Tarpy, R. M. (1997). Contemporary Learning Theory and Research : McGraw Hill.Science, 282 (5395), 1846-1851.Tononi, G., & Edelman, G. M. (1998). Consciousness and complexity. Tzelgov, J. (1997). Automatic but conscious: That is how we act most of the time. In R. S. Wyer (Ed.),The automaticity of everyday life (Vol. X, pp. 217-230). Mahwah, .N.J.: Lawrence ErlbaumAssociates.
Implicit learning: A graded, dynamic perspective
44
Willingham, D. B., Nissen, M. J., & Bullemer, P. (1989). On the development of proceduralknowledge. Journal of Experimental Psychology : Learning, Memory and Cognition, 15 , 1047-1060.
Implicit learning: A graded, dynamic perspective
45
Figure CaptionsFigure 1: Graphical representation of the relationships between quality ofrepresentation (X-axis) and (1) potency, (2) availability to control, (3) availability tosubjective experience. See text for further details.
Implicit learning: A graded, dynamic perspective
46
Implicit CognitionPrimingExplicit CognitionAutomaticityAvailability to P-ConsciousnessPotencyAvailability to A-ConsciousnessAvailability to Control ProcessesQUALITY OF REPRESENTATION (stability, distinctiveness, strength
因篇幅问题不能全部显示，请点此查看更多更全内容