variedades de emergentismo

8/11/2019 Variedades de Emergentismo

1/11

Evoluti on and Cognit ion

49

1999, V ol. 5, No. 1

1. Introduction

In addition to having var-ious technical uses, theterm emergence also hasa use in ordinary lan-guage. Thus, sometimespeople use the expressionthe emergence of x just

to mean that x has ap-peared or that x has comeup. The term emergenceis used in this way inbook titles such as TheEmergence of Symbols(B

ATES

1979) and TheEmergence of Probabili-ty (H

ACKING

1975). I willnot focus on this ordinary use of emergence inwhat follows. Instead, I will focus on the technicaluses of emergence.

In most technical uses, emergent denotes a sec-ond order property of certain first order properties(or structures), namely, the first order properties thatare emergent. However, it is controversial what thecriteria are by which emergent properties are to bedistinguished from non-emergent properties. Somecriteria are very strong, so that few, if any, propertiescount as emergent. Other criteria are inflationary inthat they count many, if not all, system properties asemergent. One of the consequences of this contro-versy is a great confusion about what is really meantby an emergent property, when this term is used in

such different disciplines as theories of self-organiza-tion, philosophy of mind, dynamical systems theo-ry, or connectionism.

Therefore, the second section of this article is in-tended to discuss, in a systematic way, several theo-ries and concepts of emergence of different strengths.It will be shown that the weaker versions are compati-ble with property reductionism. In contrast, strongerversions are incompatible with property reduction-ism. Also, the important distinction between syn-chronic and diachronic theories of emergence is de-veloped within this section.

In the third section, Iexamine the different con-cepts of emergence, distin-guished in section 2, asthey apply to particularproperties. It will becomeevident that the conceptsperform very differentjobs, and so, one needs to

be clear about which con-cept of emergence onewants to employ. Forexample, within philoso-phy of mind, particularlywithin the qualia debatethere is a need for a strongnotion of emergence,while in discussions of

emergent properties of connectionist nets one can becontent with weaker notions of emergence.

2. Weak emerge ntism,synchronic emergentism, and diachronicemergentism

There are three theories among the different variet-ies of emergentism deserving particular interest:

synchronic

emergentism, diachronic

emergentism,and aweak

version of emergentism. For synchronicemergentism the timeless relationship between asystems property and its microstructure, i.e. thearrangement and the properties of the systemsparts, is in the center of interest. For such a theory,

a property of a system is taken to be emergent, i f itis irreducible

, i.e., if it is not reducible to thearrangement and the properties of the systemsparts. In contrast, diachronic emergentism ismainly interested in predictability

of novel proper-ties. For such a theory, those properties are emer-gent that could not have been predicted in princi-ple before their first instantiation. By the way,these two stronger versions of emergentism are notindependent of each other, since irreducible prop-erties are eo ipso unpredictable in principle beforetheir first appearance. Hence, synchronically emer-

Achim Stephan

Varieties of Emergentism

In dif ferent discipl in es such as phil osophy of m ind,cognit i ve science, and t heori es of sel f-organi zati on theterm emergence has di ff erent jobs to perform . There-fore, vari ous concepts of emergence are developed andexam ined. W hi le weaker versions are compati blewi th property reducti oni sm, stronger versions are not.W ithi n philosophy of mind, particularly withi n the

qual ia debate there is a need for a strong noti on ofemergence, whi le in di scussions of emergent proper-ti es of connecti oni st nets one can do wi th weaker no-ti ons of emergence.

Emergence, reducti on, connecti oni sm, qual ia .

Abstract

Key words


2/11


50

1999, V ol. 5, No. 1

Achim Stephan

gent properties are diachronically emergent, too,but not vice versa.

Both stronger versions of emergentism are basedon a common weak theory, which at the presentpervades emergentist theorizing mainly in connec-tionism and theories of self-organization. Its three

basic featuresthe thesis of

physical moni sm

, thethesis of

systemic (or

collective

) properties

, and the the-sis of

synchroni c determini sm

are compatible withreductionist approaches without any problems. Thestronger versions of emergentism can be developedfrom weak emergent ism

by adding further theses.

Weak em ergentism

The first feature of contemporary theories of emer-gencethe thesis of

physical monism

is a thesisabout the nature of systems that have emergent

properties (or structures). The thesis says that thebearers of emergent properties (or structures) con-sist of material parts only. According to the thesis,all possible candidates for emergent properties, suchas, e.g., being alive or being in a mental state, areinstantiated only by material systems with a suffi-ciently complex physical microstructure. Itexcludes all vitalistic

positions which hold thatproperties like being alive can be instantiated onlyby a compound consisting of an organism and some

supernatural

entity, e.g. an entelechy

or an lan vi tal

.

1

Thus, all substancedualistic positions are rejected;

for they base having cognitive states on supernatu-ral bearers such as ares cogi tans

.

2

Hence, the thesis ofphysical monism denies that there are any superna-tural components responsible for a systems havingemergent properties. Particularly, this means thatliving or cognitive systemswhether artificial ornaturalconsist of the same parts as lifeless objectsof nature. There is no reason to suppose that thereare some specifi c components

that belong just tothose systems which are alive or able to cognize, butare missing in systems which are lifeless or unable tocognize. Instead, it is nothing but specific

constellati-

ons

of physicochemical processes that show vitalbehavior or have mental qualitites.(i)

Physical moni sm

. Entities existing or cominginto being in the universe consist solely ofmaterial parts. Likewise, properties, disposi-tions, behaviors, or structures classified asemergent are instantiated by systems consist-ing exclusively of physical parts.

Embracing a naturalistic

position, emergentists sub-scribe to a scientific position, but in so doing, theydo not subscribe to reductionism.

While the first thesis puts the discussion of emer-gent properties and structures within the frame-work of a physicalistic naturalism, the second thesisdelimits the type of properties that are possible can-didates for emergents. It is the thesis of systemicproperties.

This thesis is based on the assumption that gener-al

properties of complex systems fall into two differ-ent classes:

3

(i) properties which some of the sy-stems parts also have, and (ii) properties that noneof the systems parts have. Examples of the first classare properties such as being extended and having avelocity; sometimes such properties are called here-ditary properties (however, hereditary is not usedin a biological sense). Examples of properties in thesecond class are walking, reproducing, breathingand having a sensation of pain. These properties arecalled systemic or collective properties.

(ii)

Systemi c propert ies

. Emergent properties are sys-temic properties. A property is a systemic prop-erty if and only if a system possesses it, but nopart of the system possesses it.

Sometimes systemic properties are characterized asnovel properties, only by virtue of being systemic.However, this does not attribute any temporaldimension; instead it characterizes a timeless sys-tematic

relationship: in comparison to the proper-ties of the systems parts, the systems properties arenew. Thus, one could, if one liked, distinguishbetween diachronic

and synchronic

novelties. How-

ever, I prefer to characterize systematically

novelproperties as does the thesis of

systemic properties

,only diachronic

novelties in time should be charac-terized by a thesis of

novelty

(see below).It should be uncontroversial that both artificial

and natural systems with systemic properties exist.Those, who would deny their existence would haveto claim that al l

of a systems properties are heredi-tary properties, that is to say, that they are instanti-ated already by some of the systems parts. Count-less examples refute such a claim.

While the first thesis restricts the type of parts out

of which systems having emergent properties maybe built, and while the second thesis characterizesin more detail the type of properties that might beemergent, the third thesis specifies the type of rela-tionship that holds between a systems micro-struc-ture and its emergent properties as a relationship of

synchronic determi nati on

:(iii)

Synchron ic determi nat ion

. A systems propertiesand dispositions to behave depend nomologi-cally on its micro-structure, that is to say, on itsparts properties and their arrangement. There


3/11


51

1999, V ol. 5, No. 1

Vari eti es of Emergent ism

can be no difference in the systemic propertieswithout there being some differences in theproperties of the systems parts or their arrange-ment.

4

Anyone who denies the thesis of the systems prop-erties synchronic determination either has to admit

properties that are not bound to the properties andarrangement of its bearers parts, or she has to sup-pose that some other factors, in this case non-natu-ral factors, are responsible for the differentdispositions of systems that are identical in theirmicrostructure. She would have to admit, for exam-ple, that there may exist objects that have the sameparts in the same arrangement as diamonds, butwhich lack the diamonds hardness, that may havehardness 2 instead of hardness 10 on the M

OHS

-scale. This seems to be implausible. Equally beyondthought is that there may exist two micro-identical

organisms, one is viable and the other not. In thecase of mental phenomena, opinions may be morecontroversial; but one thing seems to be clear: any-one who believes, e.g., that two creatures identicalin micro-structure could be such that one is color-blind while the other can distinguish colors in theordinary way, does not hold a naturalisticphysical-istic position.

5

W eak

emergentism as sketched so far comprisesthe minimal conditions for emergent properties. Itis the common base for all stronger theories ofemergence. Moreoverand this is a reason for dis-

tinguishing it as a theory in its own rightit is heldnot only by some philosophers (e.g., B

UNGE

andV

OLLMER

), but also by cognitive scientists (e.g.,H

OPFIELD

, R

OSCH

, V

ARELA

, and R

UMELHART

) in exact-ly its weak form. The three features of weak emer-gentism(i) the thesis of

physical m onism

, (ii) thethesis of

systemi c properti es

, and (iii) the thesis of

syn-

chronic determination

, however, are compatiblewith contemporary reductionist approaches with-out further ado. Some champions of weak

emergent-ism credit the compatibility of emergence and re-ducibility as one of its merits compared to stronger

versions of emergentism.

Synchronic em ergentism

We come to the essential features of more ambitioustheories of emergence, the theses of irreducibility

(or

non-deducibility

) and of unpredictability

of certainsystemic properties. These theses are closely con-nected: Irreducible systemic properties are eo ipsounpredictable, in principle, before their first appear-ance. But besides irreducible properties, there also

seem to be properties that cant be predicted beforetheir first appearance on other grounds. Therefore,the thesis of unpredictability is more complex thanthe thesis of irreducibility of systemic properties.Thus, it is reasonable to start with a discussion of thethesis of irreducibility, which is easier to analyze.

B

ROAD

s attempt to explicate a (strong) theory ofemergence may count as downright classical; it reads:

Put in abstract terms the emergent theory assertsthat there are certain wholes, composed (say) ofconstituents A

, B

, and C

in a relation R

to each oth-er; that all wholes composed of constituents of thesame kind as A

, B

, and C

in relations of the samekind asR

have certain characteristic properties; that

A

, B

, and C

are capable of occuring in other kinds ofcomplex where the relation is not the same kind as

R

; and that the characteristic properties of the whole

R(A, B, C) cannot, even in theory, be deduced

from

the most complete knowledge of the properties ofA

,

B

, and C

in isolation or in other wholes which arenot of the formR(A, B, C) (1925, p61).

According to B

ROAD

s definition a systemic prop-erty, which is supposed to be nomologically depen-dent on its systems micro-structure (by the thesis ofsynchronic determination), is called irreducible

andtherefore emergent

, if and only if it cannot be de-duced from the arrangement of its systems partsand the properties they have isolated or in other(more simple) systems.

6

Although, prima facie, it looks as if B

ROAD

s pro-

posal gives us a clear and distinct explication ofwhat it is for a systemic property to be irreducible(or non-deducible), a further look reveals that twodifferent kinds of irreducibility having quite differ-ent consequences are concealed. As we will see, onetype of irreducibility seems to imply downwardcausation, while the other seems to imply epiphe-nomenalism. The failure to keep apart the two kindsof irreducibility has muddled the recent debateabout the emergence of properties.

To make things clearer, I shall first discuss when asystemic property is reducible

. For this to be the case,

two conditions must be fulfilled: The first is thatfrom the behavior of the systems parts alone itmust follow that the system has some property P

.The second condition demands that the behaviorthe systems parts show when they are part of thesystem follows from the behavior they show in iso-lation or in simpler systems than the system inquestion. I f both conditions are fulfil led, the behav-ior of the systems parts in other contexts revealswhat systemic properties the actual system has.That is to say, those properties are reducible. Since


4/11


52

1999, V ol. 5, No. 1

Achim Stephan

both conditions are independent from each other,two totally different possibilities for the occurrenceof irreduciblesystemic properties will result: (a) a sys-temic property Pof a system Sis irreducible, if it doesnotfollow, even in principle, from the behavior ofthe systems parts that Shas property P; and (b) a

systemic property Pof a system Sis irreducible, if itdoes notfollow, even in principle, from the behav-ior of the systems parts in simpler constellationsthan Show they will behave in S.

Thus, a necessary requirement for a systemicproperty to be reducible is that its being instantiat-ed has to follow from the behavior of its bearersparts. In other words: From the behavior of the sys-tems parts it should follow that the system has allcharacteristic features that are essential for havingthe systemic property. BROAD, for example, takesthis condition, which is enclosed in the first criteri-

on for reducibility, to be always fulfil led in the caseof the characteristic properties of chemical com-pounds and viable organisms. Their propertiesmight be irreducible only by violation of the secondcriterion, what means that from the behavior of thesystems parts in other (simpler) systems it wouldnot follow how they will behave in the actual sys-tem. In contrast, he claims that the irreducibility ofsecondary qualities and phenomenal qualities re-sults already from a violation of the first condition,since they were neither adequately characterizableby the macroscopic nor by microscopic behavior of

the systems parts, even in principle. For, when wesay that a certain object is red or a chemical sub-stance has the smell of liquid ammonia, we do notmean that the corresponding systems parts behaveormovein a certain way. No progress in the sciencescould change this state of affairs in any way.7BROADhas il lustrated the fundamental distinction between(behaviorally) analyzableand unanalyzableproper-ties by pointing to characteristic properties of orga-nisms and secondary qualities, respectively.

If secondary and phenomenal qualities are notanalyzable,8even in principle, then there is no pros-

pect that an increase of scientific knowledge willclose the gap between physical processes and sec-ondary qualities or between physiological processesand phenomenal states of consciousness (qualia),respectively.

We can now specify more exactly the feature ofirreducibility which is central for synchronicemer-gence. Its first variant is based on the behavioralunanalyzability of systemic properties. It reads:(a ) Unanalyzability. Systemic properties which are

not behaviorally analyzablebe it micro- or

macro-scopicallyare (necessarily) irreduc-ible.

However, even if secondary and phenomenal quali-ties belong to the class of unanalyzable properties, itdoes not follow that the specific behavior of the sys-tems parts upon which those qualities supervene is

itself not deducible from the behavior those partsshow isolated or in other (simpler) systems. The irre-ducibility which results from a violation of the firstcriterion of reducibility does not imply, by itself, aviolation of the second criterion of reducibility.

On the other side, however, even analyzable sys-temic properties can be irreducible and thereforeemergent. This is the case when the second criterionof reducibility will be violated, i.e., when the behav-ior of the systems parts does not follow from theirbehavior in other (simpler) constellations. BROADthinks that such examples of irreducible behavior

might occur in chemical compounds and also in or-ganisms.9His central idea is that the parts of a genu-inely novel structure, such as, e.g., an organism incomparison to any inorganic compound, might be-have in a way that is not deducible from the partsbehavior in other structures. Implicitly, that meansthat the actual behavior of parts that interact inwholes does not result from their behavior in pairs.10

If the behavior of some systems parts is irreduciblein this respect, then all properties that depend no-mologically on the behavior of the systems parts(for example, reproduction) are irreducible too.

Thus, we can specify more precisely the secondvariant of a systemic propertys irreducibility. It isbased on the non-deducibility of the behavior of thesystems parts:(b ) Irreducibil it y of t he components behavior. The

specific behavior of a systems componentswithin the system is irreducible if it does notfollow from the components behavior in iso-lation or in other (simpler) constellations.

A violation of the second criterion of reducibility,which is manifested in the irreducibility of the com-ponents behavior, does not imply, however, a vio-

lation of the first criterion of reducibility. Systemicproperties that cannot be reduced because thebehavior of the systems parts is irreducible mightnevertheless be behaviorally analyzable. Hence, thetwo criteria of reducibility as well as those irreduc-ibili ties that are based on the violation of these crite-ria are independent of each other. Summarizing, weget from (a ) and (b ) the following modified versionof systemic property irreducibility:(iv) Irreducibility. A systemic property is irreducible

if (a) it is neither micro- nor macro-scopically


5/11

Evoluti on and Cognit ion 53 1999, V ol. 5, No. 1


behaviorally analyzable, or if (b) the specific be-havior of the systems components, over whichthe systemic property supervenes, does not fol-low from the components behavior in isola-tion or in other (simpler) constellations.

Thus, we have to distinguish two totally different

types of irreducibility of systemic properties.Equally different seem to be the consequences thatresult from them. If a systemic property is irreduc-ible because the behavior of the systems parts, overwhich the property supervenes, is itself irreducible,this seems to imply that we have a case of down-ward causation. For, if the components behavior isnot reducible to their arrangement and the behaviorthey show in other (simpler) systems or in isolation,then there seems to exist some downward causalinfluence from the system itself or from its structureon the behavior of the systems parts. To be sure, if

there would exist such instances of downward cau-sation this would not amount to a violation ofsome widely held assumptions, such as, for exam-ple, the principle of the causal closure of the physi-cal domain. Within the physical domain, we wouldjust have to accept additional types of causal influ-ences besides the already known basal types ofmutual interactions.

In contrast, the occurrence of unanalyzable prop-erties does not imply any kind of downward causa-tion. Systems that have unanalyzable propertiesthat depend nomologically on their bearers micro-

structures need not be constituted in a way thatamounts to the irreducibility of their componentsbehavior. Nor is implied that the systems structurehas a downward causal influence on the systemsparts. All the more, there is no reason to assume thatunanalyzable properties themselves exert a causalinfluence on the systems parts. Rather it is to ask,how unanalyzable properties might have any causalrole to play at all. Since they are not behaviorallyanalyzablethat is to say, they neither seem to cor-respond to any mechanism nor do they seem to re-sult from any mechanism, it is hard to see how

they could be causally effective themselves. I f, how-ever, one can not see howunanalyzable propertiesmight play a causal role, then, it seems, such proper-ties are epiphenomena.

Diachronic emergentism

All diachronic theories of emergence have at bot-tom a thesis about the occurrence of genuine novel-tiesproperties or structuresin evolution. This

thesis excludes at the same time all preformationistpositions.(v) Novelty. In the course of evolution exemplifica-

tions of genuine novelties occur again andagain. Already existing building blocks will de-velop new constellations; new structures will

be formed that constitute new entities withnew properties and behaviors.

However, bare addition of the thesis of novelty doesnot turn a weak theory of emergence into a strongone, since reductive physicalism remains compati-ble with such a variant of emergentism. Only theaddition of the thesis of unpredictability, in princi-ple, of novel properties will lead to stronger forms ofdiachronicemergentism.

A short consideration shows that systemic prop-erties can be unpredictable in principle for two fun-damentally different reasons: (i) they can be unpre-

dictable because the micro-structure of the system,which exemplifies the property for the first time inevolution, is unpredictable. For, if the micro-struc-ture of a newly emerging system is unpredictable, soare the properties which depend nomologically onit. (ii) However, a property can be unpredictableeven though the novel systems micro-structure ispredictable. That is the case if the property itself isirreducible: For, if systemic properties are irreduc-ible, then they are unpredictable before their firstappearance. However, this does not preclude thatfurther occurrences of such properties might be pre-

dicted adequately.Since in the second case criteria for being unpre-

dictable are identical with those for being irreduc-ible, this notion of unpredictability will offer notheoretical gains beyond those afforded by the no-tion of irreducibility.11Let us focus, therefore, on thefirst case:unpredictabi li ty of structure. This version ofunpredictability passed almost unnoticed in classi-cal li terature on emergentism during the 1920s, butbecause of strong interest in dynamical systems andchaotic processes this notion gains considerable sig-nificance.

The structure of a new formed system can be un-predictable for several reasons. Thus, belief in an in-deterministic universe implies that there will benovel, unpredictable structures. However, from anemergentist perspective it would be of no interest, ifa new structures appearance would be unpredict-able only because its coming into being is not deter-mined, not to mention that most emergentistsclaim, anyway, that the development of new struc-tures is governed by deterministic laws. But still de-terministic formings of new structures can beunpre-


6/11


Achim Stephan

dictable in principle, if they are governed by lawswhich are attributed to deterministic chaos.

An essential outcome of the theory of chaos isthat there existeven very simplemathematicalfunctions, whose own behavior cannot be predict-ed. Only the rise of experimental mathematics on

highly efficient computers has revealed, for exam-ple, the properties of various logistic functions.Their intra-mathematical unpredictability has to dowith an aperiodic behavior of these functions, bywhich marginally different initial values of somevariable can lead to radically distinct trajectories ofthe functions.

A standard example is the logistic function f(x) =mx(1 x) for 0 x1. For a parameter mwith 0 m4 the logistic function maps the interval [0, 1] ontoitself. Of particular interest is, how parameter mex-ercises an influence on the long term behavior of

the function when iterated repeatedly. For 0 m1the situation is obvious. All initial values of the vari-able xlet the function f(x) approximate the value 0after sufficiently many iterations, thus, the origin isthe attractor. For 1


7/11



ability in principle of systemic properties is entirelycompatible with their being reducible to the micro-

structure of the system that instantiates them.

Synopsis

Figure 1 depicts the logical relationships that holdbetween the different versions of emergentism.

W eak diachronic emergentism results from weakemergentismby adding a temporal dimension in theform of the thesis of novelty. Both versions are com-patible with reductive physicalism. Weak theoriesof emergence are used today mainly in cognitive sci-ences, particularly for characterization of systemic

properties of connectionist nets, and in theories ofself-organization. Synchronic emergentism resultsfrom weak emergentism by adding the thesis of irre-ducibility. This version of emergentism is importantfor the philosophy of mind, particularly for debat-ing nonreductive physicalism and qualia. It is notcompatible with reductive physicalism any more.Strong diachroni cemergentismonly differs from syn-chronic emergentism because of the temporal di-mension in the thesis of novelty. In contrast, struc-ture emergentism is entirely independent ofsynchronic emergentism. It results from weak emer-

gentism by adding the thesis of structureunpre-dictability. Although structure emergentism em-phasizes the boundaries of prediction withinphysicalistic approaches, it is compatible with re-ductive physicalism, and so it is weaker than syn-chronic emergentism. Theories of deterministicchaos in dynamical systems can be acknowledged asa type of structure emergentism. Likewise, its per-spective is important for evolutionary research. Incomparison to the above mentioned versions ofemergentism the synthetic position of strong diach-

roni c structu re emergent ismhas no equivalent in re-cent discussion. Most important from a theoretical

point of view areweak emergent ism, synchronic emer-gentism, and diachroni c structure emergent ism.

3. Eme rge nce in the qualia debate andin connectionism

Let us turn to some specific cases to consider howthe different concepts of emergence are applicable.We will see that in one of the central debateswithin philosophy of mind, namely the debateabout the nature of qualia, there is need for astrong notion of emergence such as that of syn-

chronic emergentism. In contrast, in connection-ism a weaker notion of emergence is employed.

Emergence in the deb ate over whether

qua lia are ph ysical

In recent debate about qualia, NAGEL, BLOCK, JACK-SON, LEVINE, CHALMERSand MCGINN, among others,have argued in one way or another that qualitativemental phenomena are not reducible to physical orfunctional states, respectively. If their argumentssucceed, they imply emergentist or substancedual-

istic positions. Most interesting and powerfulseems to be LEVINEs so-called explanatory gapargument, which I will consider closely in the fol-lowing.

LEVINE starts with comparing two statements,namely (i) pain is the firing of C-fibers, and (ii)heat is the motion of molecules.13The decisive dif-ference between the two identity statements, ac-cording to LEVINE(1983), is that the second is fu l lyexplanatory, while the first is not. The second iden-tity statement is assumed to be fully explanatory,

+noveltyweak

emergentism

weakdiachronic

emergentism +unpredictability

+irreducibility

diachronicstructure

emergentism

+irreducibility

synchronicemergentism

strongdiachronic

emergentismstrong diachronic

structure emergentism

+irreducibility

+novelty

Figure 1 : Logical relationships that hold between the different versions of emergentism.


8/11


Achim Stephan

because knowledge of natural laws helps us under-stand why the motion of molecules has exactly thecausal role usually ascribed to heat. Thereby, it ispresupposed that the macro-physical concept ofheat can be fully explicated by heats causal role. Inother words: statement (ii) is fully explanatory be-

cause some systems property of heat is reducible inrespect to the motions of molecules, and so far it isnot emergent. On the other hand, the reason forstatement (i) not being fully explanatory is, LEVINEsays, that the notion of pain is not exhausted bythe causal role of pain. And, what is true for painseems to be true for all other phenomenal states.

Lets see why pain and other qualia are so resist-ent to explanatory reduction. According to LEVINE,reduction that is explanatory requires two stages:Stage 1 involves the (relatively? quasi?) a prioriprocess of working the concept of the property to

be reduced into shape for reduction by identifyingthe causal role for which we are seeking the under-lying mechanisms. Stage 2 involves the empiricalwork of discovering just what those underlyingmechanisms are (1993, p132).

If one claims that a reduction that is explanatoryis impossible in principle, as is claimed for qualia,that does not imply a failure of the second task.What is implied is a failure, in principle, of the firsttask. Or, as LEVINEputs it: to the extent that there isan element in our concept of qualitative characterthat is not captured by features of its causal role, to

that extent it will escape the explanatory net of aphysicalistic reduction (1993, p134). Thus, LEVINEssynchronicqual iaemergent ismis based on two theses:

(1) The reduction of a systemic property Pis ex-planatory if and only if the realization base ex-hausts exactly the causal role which is constitutiveof P.

(2) Phenomenal properties (or states) are not ful-ly graspable by the features of their causal role.14

Several responses are possible to LEVINEs claim:(i) one tries to avoid emergentism by showing ei-ther how properties that are not fully graspable via

their causal roles, might yet be reducible explana-torily, or by showing that phenomenal qualitiesare graspable adequately via causal roles. For bothvariants, however, there exist nothing more thanfirst attempts or statements of intentions. (ii) Oneaccepts both theses and, thereby, an emergentistposition. But then, additional questions arise: Towhat type of psychophysical theories does syn-chronic qualiaemergentism belong? Is it still aphysicalistic position, or rather a kind of propertydualism?

Anyway, if LEVINEs explanatory gapargumentsucceeds then qualia are extremely good candi-dates for emergent properties in the (strong) syn-chronic sense.

Emergence in con nectionism

In the last decade, connectionism has received greatattention in cognitive science. Its core idea is toassume a network of elementary units that have acertain level of activation. Units are connected witheach other. Units whose activation exceeds a certainthreshold, can activate or inhibit other unitsaccording to certain weights that specify prevailingconnections.

To see to what extent emergentist considerationsare relevant for connectionism, I shall first examinemore closely the parameters that specify a connec-

tionist net. Each net is determined essentially bythree factors: (i) by the number of units and con-necting links which hold between them; (ii) by thefunction that determines the level of activation foreach unit; and (iii) by the rule that determines howconnection weights will change.

In each case the number of units and the linksbetween them are fixed; they make up, so to speak,the skeleton of a network which is static under or-dinary circumstances: neither the number of unitsnor the structure of their links will change. A sys-tems actual dynamic results from the possibility of

modifying the weights of its internal connectinglinks. From a macroscopic point of view, these con-tinuous processes of accomodation can be seen aslearning procedures. Thus, a connectionist netlearns by locally determined changes of its con-nections weights, and not by adding some furtherpropositions to its data base. When a net is fed withvarious inputs after its practice time, it will calcu-late its outputs or answers with stable weights ingeneral.

Behavior and properties of connectionist netsgive rise to emergentist considerations in many

ways. Three aspects should be discerned: first, con-nectionist nets obviously have systemic properties,that is they have properties their parts do not have.Thus, a nets properties are at least weakly emergent.However, it remains to be determined whether sys-temic properties of nets are emergent in a strongsense, namely synchronically emergent. Secondly,many systemic properties of nets seem to be emer-gent in a phenomenological way. By this, I meanthat the properties appear, or come into being bythemselves if nets get adequate stimuli. These facts


9/11



of the case are referred to in English by the wordemergent in its ordinary use. No specific theory ofemergence is implied by this usage. Since some con-nectionists make use of the word emergent in itsordinary use intermingled with a more technicaluse, it is important to tell the notions apart. Eventu-

ally, connectionist nets develop during their train-ing phasein a somewhat mini-evolutionary pro-cesstheir soft structures, by which I understandthe specific distribution of link weights. Thus, thequestion arises whether this formation of structureis an interesting case of structure emergence.

Lets examine first the relationship between glo-bal net properties and their realization base, name-ly the nets structure and its parts properties. Con-siderations of connectionist nets architectures andtheir modes of operation reveal that only trainednets show typical macroscopic properties such as

rule following, schemata formation, or patternrecognition. Untrained nets do not have those(cognitive) properties, they have only the disposi-tion to acquire them. Macroscopic properties oftrained nets supervene upon both their given hardstructure, and their acquired soft structure. Theyare fully reducible to the organization of the net inconsideration, the properties of its units (namelytheir activation formula), and the properties oflinks consisting between its units (nameley distri-bution of weights, and formula for changingweights). If these quantities are known, the output

behavior of any net can be predicted exactly andexplained. It is obvious that a nets parts, namely itsunits and the links between them, do not have anyof those macroscopic (cognitive) properties. So far,these properties a net acquires by training are typi-cal systemic properties. However, since they are notirreducible systemic properties, but are completelydeducible from a nets structure, and its parts andlinks properties, a nets systemic properties aremerely weakly emergent. They are not synchroni-cally emergent.

RUMELHARTand MCCLELLANDdiscuss in great de-

tail net properties which they call emergent, andstress that connectionist approaches do not implyreductionistic, but interactionistic positions.

We are simply trying to understandthe essenceof cognition as a property emerging from the inter-actionsof connected units in networks. We certainlybelieve in emergent phenomena in the sense of phe-nomena which could never be understood or pre-dicted by a study of the lower level elements in iso-lation. [...] This is the case in many fields. Knowingabout the individuals tells us little about the struc-

ture of the organization, but we cant understandthe structure of the higher level organizations with-out knowing a good deal about individuals and howthey function. This is the sense of emergence we arecomfortable with (1986, p128).

However, RUMELHART and MCCLELLANDs claim

that connectionism amounts to a non-reductive po-sition results from an inappropriate strong notionof reduction. Both authors, one must know, take asystems properties as reduced only when eithersome of the systems parts have these properties al-ready, or when the systems property can be reducedto some linear interactions of the parts. Yet, almostno systemic property would be reducible accordingto these criteria. At the same time, RUMELHARTandMCCLELLANDconcede that a nets behavior is com-pletely intelligible, if one takes into account all in-teractions between its units, that is to say, if one

considers a systems hard and soft structure. Hence,even from their point of view, connectionism is notan instance of synchronic emergentism.

In discussions of connectionist nets behavior or-dinary use and technical use of the notion of emer-gence often gets intermixed. For example, RUMEL-HART and MCCLELLAND reiterate that macroscopicproperties emerge from micro-level interactions:[M]any of the constructs of macrolevel descrip-tions such as schemata, prototypes, rules, produc-tions, etc. can be viewed as emergingout of interac-tions of the microstructure of distributed models

(1986, p125; italics are mine). By this characteriza-tion, they lay stress upon the fact that a nets sys-temic properties come into being from the complexbehavior of the nets components: rules and sche-mata can be available without being explicitly fedinto the system. Here, emergence is not discussed ina technical sense, rather the term is used in its ordi-nary sense to describe a systems abilities to acquiresystemic properties by self-organizational processes.The temporary manifestation of schemata that werealready latently in the links weights is interpreted,then, as an emergent property of the net. Connec-

tionists, thus, mainly point to emergent rules oremergent schemata to demarcate their positionfrom classical representationalism, accordingly towhich all rules and schemata have to be fed in ex-plicitely (see HORGAN/TIENSON1996).

There is a further feature of connectionist netsthat provokes emergentist considerations: duringthe phase of training or learning a net runs througha mini-evolutionary processlink weights areadapted such that the net is enabled to handle thetasks it is supposed to master. Within this time the


10/11


Achim Stephan

so-called soft structure of the net develops.15Onlywhen the links weights are adjusted adequatelydoes a net has available desired macro-properties,that is, only then can it develop schemata, recognizepatterns, or make use of rules. Those are not imple-mented explicitely as they are in symbol manipulat-

ing devices, but are extracted from given material.However, this is not a case of genuine structureemergence: not only does the distribution of weightsresult from deterministic principles, the changes ofweights are even calculable exactly, if we know thelearning rule, the activation formula, the units ini-tial activation, the intial weights, and the inputs.

Even though we therefore should not speak ofstructure emergence in connectionist networks, re-garding their soft structure, nets show a tremendousplasticity, when compared with other objects, evenwhen compared with other dynamical systems.

Chemical compounds, to give an example, have nodegrees of freedom to change their internal structure.In this respect connectionist nets differ clearly fromseeming analogic cases such as diamonds, whichwere referred to by RUMELHARTand MCCLELLANDtoexplain emergent systems properties (1986, p128).Diamonds are not dynamical systems that realizeonly after a certain number of reiterated steps theproperty of being hard. The diamonds property ofbeing hard is always manifest,it does not emerge.

To sum up, connectionist

nets do not instantiate anystronger type of emergence.Neither are the nets proper-ties synchronically emer-

gent, nor is the formation of a nets soft structure acase of structure emergence. In the weak sense inwhich macroscopic properties of nets are emergent,all systemic properties of complex systems are emer-gent. A difference to many other systems exists atbest in the plasticity of nets and in their capacity to

develop in a training phase by themselves adequateattractors to cope with given tasks. Correspondingmacroscopic properties will become manifest onlytemporarily during treatments. This second orderproperty might justifiably be characterized as phe-nomenological emergence. Only some dynamicalsystems have this property.

4. Conclusion

In particular, I have distinguished three versions ofemergentism: weak emergentism, (strong) synchro-

nic emergentism, and diachronic structure emer-gentism. Synchronic emergentism results from weakemergentism by adding the thesis of irreducibility. Itturned out that this version of emergentism isimportant for the philosophy of mind, particularlyfor debating qualia. However, this does not establishthat qualia are emergent phenomena in the strongsense. But they are good candidates for being so.Connectionism, on the other hand, does not treat

any phenomenon as synchro-nically emergent, nor does ittreat novel structures as candi-

dates for structure emergence.Phenomena discussed by con-nectionism are emergent onlyin the weak sense.16

Notes

1 Supernatural properties are meant to be hyperphysical,i.e., as independent from (physical) nature and their laws.

2 In the history of emergentism, however, there were theo-ries of emergence that did not claim the thesis of physical

monism; instead, they tookmentalorneutralbuilding blocksas fundamental (cf. BROAD1925, pp610653). Anyway, thethesis of physical monism is not questioned by main streamdebate today.

3 General properties are properties of a general type, such ashaving a weight, or being liquid; they are not specific prop-erties, such as having a weight of 154.5 pounds or beingliquid by a temperature of 1200 C.

4 In recent debate, the thesis of synchroni c determin ati onissometimes stated in a less stronger version as the thesis ofmereological supervenience, which claims that a systemssproperties (or dispositions) supervene on its parts proper-ties and their arrangement. Then, too, there is no differencein the systemic properties without differences in the parts

properties or their arrangement. The thesis of mereologicalsupervenience, however, is weaker than the thesis of syn-chronic determination, since it does not claim the depen-denceof the systems properties from its micro-structure, itonly claims their covariance.

5 However, similar considerations hold for propositional at-

titudes only, as long as one does not subscribe to external-ism, that is to say, if one does not claim that, e.g., thecontent of a belief depends essentially on the nature of thereferents of the believers thoughts and concepts.

6 Properties that might be ascribed to a systems part in iso-lation are, according to BROAD, properties that depend es-sentially on the micro-structure of the part, while externalfactors, such as the parts arrangement and its neighboringparts, can be seen as almost irrelevant for the parts havingthese properties (cf. 1919, pp112f).

7 However, whether reference to linguistic usage might an-swer questions concerning reducibility in a definite way iscontroversial. Particularly, CHURCHLANDhas opposed argu-ments of the BROADIANstyle (see 1988, pp29ff).

Achim Stephan, Insti tut fr Phi losophie derUni versitt (TH), D -76128 Karl sruhe, Ger-many. Email : astephan@uni -bremen.de

Authors ad dress


11/11



8 Properties that are called unanalyzable for simplicity here,might be analyzable in other ways than by behavioral fea-tures. A certain smell, for example, might be analyzed as amixture of the smells of musk and fish-meal. This, however,would not be an analysis based on concepts of motion andbehavior.

9 BROADhas also examined in abstracto under what condi-

tions the behavior of systems components can be irreduc-ible (cf. 1919, pp113f). Recently, BECHTELhas held a simi larposition: although studying the properties of amino acidsin isolation may reveal their primary bonding properties,it may not reveal to us those binding properties that giverise to secondary and tertiary structure when the aminoacids are incorporated into protein molecules (1988,p95).

10 Scenarios of this kind are discussed already by M ILLandFECHNER(cf. STEPHAN1999, sections 6.1. and 7.3.).

11 A difference in extension between both notions could re-sult only in respect to those properties, which, althoughreducible, are not predictable before their first appearance.A reducible property is unpredictable before its first appear-ance if the behavior of the systems components upon

which it supervenes does not follow from their behavior in

those systems that exist at the time of prediction. The no-tion of unpredictabili ty widened in this way would depend,however, in a very contingent way on the chronologicalorder of systems coming into being in evolution. Therefore,such a notion should not be important in qualifying thenotion of emergence.

12 LAPLACEhimself assumed that his calculator knows all laws

governing nature, but he took those laws to be laws of N EW-TONIANphysics. During his time, nobody knew anythingabout aperiodic chaotic processes.

13 Talk about firing of C-fibers has its firm place in literatureabout the qualiaproblem. It is, however, only a fill-in foran adequate neurophysiological analysis of a pains mate-rial base.

14 Notice the analogies between LEVINEs theses and the crite-ria for irreducibility in the section above.

15 The developing soft structure depends on two factors, thenets hard structure, and the given inputs, that is externalinfluences on it.

16 A longer version of this paper was published under the titleVarieties of Emergence in Artificial and Natural Systems(1998). For more detailsboth historically and systemati-

callycf. my book on Emergenz (1999).

References

Alexander, S. (1920 ) Space, Time, and Deity. Two Volumes.London: Macmillan.

Bates, E. (1979 ) The Emergence of Symbols. New York: Aca-demic Press.

Bechtel, W. (19 88) Philosophy of Science. An Overview forCognitive Science. Hillsdale, NJ: Lawrence Erlbaum Asso-ciates.

Broad, C. D. (1919) Mechanical Explanation and its Alterna-

tives. Proceedings of the Aristotelian Society 19: 86124.Broad, C. D. (1925) The Mind and i ts Place in Nature. Lon-don: Kegan Paul, Trench, Trubner & Co.

Churchland, P. (1988) Matter and Consciousness. RevisedEdition. Cambridge, MA, London: MIT Press.

Hacking, I. (1975) The Emergence of Probability. Cambridge:Cambridge University Press.

Horgan, T./Tienson, J. (1996 ) Connectionism and the Philos-

ophy of Psychology. Cambridge, Ma., London: MIT Press.Levine, J. (1 983 ) Materialism and Qualia: The Explanatory

Gap. Pacific Philosophical Quarterly 64: 354361.Levine, J. (1 993 ) On Leaving Out What Its Like. In: Davies,

M./Humphreys, G. W. (eds) Consciousness. Oxford, Cam-bridge, MA: Blackwell, pp. 121136.

Rumelhart, D. E./M cClelland, J. L.(198 6) PDP Models andGeneral Issues in Cognitive Science. In: Rumelhart, D. E./McClelland, J. L./the PDP Research Group, Parallel Distrib-uted Processing. Explorations in the Microstructure of Cog-

nition. Vol. 1. Cambridge, MA, London: MIT Press, pp.110146.Stephan, A. (1998) Varieties of Emergence in Artificial and

Natural Systems. Zeitschrift fr Naturforschung 53c: 639656.

Stephan, A. (1999) Emergenz. Von der Unvorhersagbarkeitzur Selbstorganisation. Dresden, Mnchen: Dresden Uni-versity Press.

variedades de emergentismo

Documents