Varieties of Emergence in Artificial and Natural Systems* Achim Stephan Institut für Philosophie der Universität Karlsruhe (TH), Postfach 6980, D-76128 Karlsruhe at present: Hanse-Wissenschaftskolleg, Fischstr. 31, D-27749 Delmenhorst Z. Naturforsch. 53c, 639-656 (1998); received April 20, 1998 Emergence, Reduction, Connectionism, Qualia, Synergetics In different disciplines such as philosophy of mind, dynamical systems theory, and connec­ tionism the term ‘emergence’ has different jobs to perform. Therefore, various concepts of emergence are developed and examined. While weaker versions are compatible with prop­ erty reductionism, stronger versions are not. Within philosophy of mind, particularly within the qualia debate there is a need for a strong notion of emergence, while in discussions of emergent properties of connectionist nets or of dynamical systems one can do with weaker notions of emergence.

Introduction quences of this controversy is a great confusion In addition to having various technical uses, the about what is really meant by an ‘emergent prop­ term ‘emergence’ also has a use in ordinary lan­ erty’, when this term is used in such different disci­ guage. Thus, sometimes people use the expression plines as theories of self-organization, philosophy ‘the emergence of x’ just to mean that x has ap­ of mind, dynamical systems theory, or connec­ peared or that x has come up. The term ‘emer­ tionism. gence’ is used in this way in book titles such as Therefore, the second section of this article is “The Emergence of Symbols” (Bates, 1979) and intended to discuss, in a systematic way, several “The Emergence of Probability” (Hacking, 1975). theories and concepts of emergence of different Of course, one could speak about ‘the emergence strengths. It will be shown that the weaker ver­ of animals with brains’ or about ‘the emergence of sions are compatible with property reductionism. robots’ in this sense of ‘emergence’. However, I In contrast, stronger versions are incompatible will not focus on this ordinary use of ‘emergence’ with property reductionism. Also, the important in what follows. Instead, I will focus on the techni­ distinction between synchronic and diachronic cal uses of ‘emergence’. theories of emergence is developed within this sec­ In most technical uses, ‘emergent’ denotes a sec­ tion. ond order property of certain first order properties In the third section, I examine the different con­ (or structures), namely, the first order properties cepts of emergence, distinguished in section 2 , as that are emergent. However, it is controversial they apply to several natural and artificial systems. what the criteria are by which emergent properties It will become evident that the concepts perform are to be distinguished from non-emergent prop­ very different jobs, and so, one needs to be clear erties. Some criteria are very strong, so that few, if about which concept of emergence one wants to any, properties count as emergent. Other criteria employ. For example, within philosophy of mind, are inflationary in that they count many, if not all, particularly within the qualia debate there is a system properties as emergent. One of the conse- need for a strong notion of emergence, while in discussions of emergent properties of connec- tionist nets or of dynamical systems one can be * This communication is a contribution to the workshop content with weaker notions of emergence. on “Natural Organisms, Artificial Organisms, and ‘Emergent’ is not only attributed to properties Their Brains” at the Zentrum für interdisziplinäre in a strict sense, but also to dispositions, behavior, Forschung (ZiF) in Bielefeld (Germany) on March 8 -1 2 , 1998. and structures. To simplify my presentation, I will use the concept of a property in a wide sense to Reprint requests to Dr. A. Stephan, Nauheimer Str. 39, D-65428 Rüsselsheim. apply to dispositions (e.g., being breakable) and E-mail: [email protected] behavior: Of a system which behaves such and

0939-5075/98/0700-0639 $ 06.00 © 1998 Verlag der Zeitschrift für Naturforschung, Tübingen • www.znaturforsch.com. D

Dieses Werk wurde im Jahr 2013 vom Verlag Zeitschrift für Naturforschung This work has been digitalized and published in 2013 by Verlag Zeitschrift in Zusammenarbeit mit der Max-Planck-Gesellschaft zur Förderung der für Naturforschung in cooperation with the Max Planck Society for the Wissenschaften e.V. digitalisiert und unter folgender Lizenz veröffentlicht: Advancement of Science under a Creative Commons Attribution Creative Commons Namensnennung 4.0 Lizenz. 4.0 International License. 640 A. Stephan • Varieties of Emergence in Artificial and Natural Systems such, one can always say it has the property (or the bearers of emergent properties (or structures) disposition) to behave such and such. I will, how­ consist of material parts only. According to the ever, discuss emergent structures separately. As we thesis, all possible candidates for emergent proper­ will see, they are particularly important for the in­ ties, such as, e.g., being alive or being in a mental terpretation of dynamical systems and connec- state, are instantiated only by material systems tionist nets. with a sufficiently complex physical microstruc­ ture. It excludes all vitalistic positions which hold that properties like being alive can be instantiated Weak Emergence, Synchronic Emergence, only by a compound consisting of an organism and and Diachronic Emergence some supernatural entity, e.g. an entelechy or an There are three theories among the different elan vital} Thus, all substance-dualistic positions varieties of emergentism deserving particular in­ are rejected; for they base having cognitive states terest: synchronic emergentism, diachronic em er­ on supernatural bearers such as a res cogitans? gentism, and a weak version of emergentism. For Hence, the thesis of physical monism denies that synchronic emergentism the timeless relationship there are any supernatural components responsi­ between a system’s property and its microstruc­ ble for a system’s having emergent properties. Par­ ture, i.e. the arrangement and the properties of the ticularly, this means that living or cognitive sys­ system’s parts, is in the center of interest. For such tems - whether artificial or natural - consist of a theory, a property of a system is taken to be the same parts as lifeless objects of nature. There emergent, if it is irreducible, i.e., if it is not reduci­ is no reason to suppose that there are some spe­ ble to the arrangement and the properties of the cific components that belong just to those systems system’s parts. In contrast, diachronic emer­ which are alive or able to cognize, but are missing gentism is mainly interested in predictability of in systems which are lifeless or unable to cognize. novel properties. For such a theory, those proper­ Instead, it is nothing but specific constellations of ties are emergent that could not have been pre­ physico-chemical processes that show vital be­ dicted in principle before their first instantiation. havior or have mental qualitites. By the way, these two stronger versions of emer­ gentism are not independent of each other, since (i) Physical monism. Entities existing or coming irreducible properties are eo ipso unpredictable in into being in the universe consist solely of principle before their first appearance. Hence, syn- material parts. Likewise, properties, disposi­ chronically emergent properties are diachronically tions, behaviors, or structures classified as emergent, too, but not vice versa. emergent are instantiated by systems con­ Both stronger versions of emergentism are sisting exclusively of physical parts. based on a common ‘weak’ theory, which at the present pervades emergentist theorizing mainly in Embracing a naturalistic position, emergentists connectionism and theories of self-organization. subscribe to a scientific empiricist position, but in Its three basic features - the thesis of physical mo­ so doing, they do not subscribe to reductionism. nism , the thesis of systemic (or collective) proper­ While the first thesis puts the discussion of ties, and the thesis of synchronic determinism - are emergent properties and structures within the compatible with reductionist approaches without framework of a physicalistic naturalism, the se­ any problems. The stronger versions of emer­ cond thesis delimits the type of properties that are gentism can be developed from weak emergentism by adding further theses. ^Supernatural’ properties are meant to be hyperphysi­ cal, i.e., as independent from (physical) nature and Weak emergentism their laws. 2 In the history of emergentism, however, there were The first feature of contemporary theories of theories of emergence that did not claim the thesis of emergence - the thesis of physical monism - is a physical monism; instead, they took mental or neutral building blocks as fundamental (cf. Broad, 1925, thesis about the nature of systems that have emer­ p. 610-653). Anyway, the thesis of physical monism is gent properties (or structures). The thesis says that not questioned by main stream debate today. A. Stephan • Varieties of Emergence in Artificial and Natural Systems 641 possible candidates for emergents. It is the thesis While the first thesis restricts the type of parts of systemic properties. out of which systems having emergent properties This thesis is based on the assumption that gene­ may be built, and while the second thesis charac­ ral properties of complex systems fall into two dif­ terizes in more detail the type of properties that ferent classes:3 (i) properties which some of the might be emergent, the third thesis specifies the system’s parts also have, and (ii) properties that type of relationship that holds between a system’s none of the system’s parts have. Examples of the micro-structure and its emergent properties as a first class are properties such as being extended relationship of synchronic determination : and having a velocity; sometimes such properties (iii) Synchronic determination. A system’s proper­ are called hereditary properties (however, 'heredi­ ties and dispositions to behave depend nomo- tary' is not used in a biological sense). Examples logically on its micro-structure, that is to say, of properties in the second class are walking, re­ on its parts’ properties and their arrange­ producing, breathing and having a sensation of ment. There can be no difference in the sys­ pain. These properties are called systemic or col­ temic properties without there being some lective properties. differences in the properties of the system’s

parts or their arrangement . 6 (ii) Systemic properties. Emergent properties are systemic properties. A property is a systemic Vollmer illustrates this case with an example property if and only if a system possesses it, from chemistry: The system ‘graphite’ consists of

but no part of the system possesses it .4 carbon atoms arranged in honeycomb layers. However, a tetrahedral arrangement of the same Sometimes systemic properties are charac­ atomic parts constitutes a different system, e.g., terized as ‘novel’ properties, only by virtue of be­ ‘diamond’, with different systemic properties. By ing systemic .5 However, this does not attribute any choosing different atomic building blocks, but the temporal dimension; instead it characterizes a same structure type, again we obtain another sys­ ‘timeless’ systematic relationship: in comparison to tem, e.g. ‘silicon’. In each of these examples the the properties of the system’s parts, the system’s systemic properties are nomologically dependent properties are ‘new’. Thus, one could, if one liked, on properties of the system’s parts and their ar­ distinguish between diachronic and synchronic rangement (cf. to Vollmer, 1988, p. 93). Anyone novelties. However, I prefer to characterize sys­ who denies the thesis of the system’s properties tematically novel properties as does the thesis of synchronic determination either has to admit ‘free systemic properties, only diachronic novelties in floating’ properties, i.e., properties that are not time should be characterized by a thesis of novelty bound to the properties and arrangement of its (see below). bearer’s parts, or she has to suppose that some It should be uncontroversial that both artificial other factors, in this case non-natural factors, are and natural systems with systemic properties exist. responsible for the different dispositions of sys­ Those, who would deny their existence would have tems that are identical in their microstructure. In to claim that all of a system’s properties are ‘he­ the case mentioned above, she would have to ad- reditary’ properties, that is to say, that they are instantiated already by some of the system’s parts. Countless examples refute such a claim. 6 In recent debate, the thesis of synchronic determina­ tion is sometimes stated in a less stronger version as the thesis of mereological supervenience, which claims that a systems’s properties (or dispositions) supervene 3 General properties are properties of a general type, on its parts' properties and their arrangement. Then, such as having a weight, or being liquid; they are not too, there is no difference in the systemic properties specific properties, such as having a weight of 154.5 without differences in the part's properties or their pounds or being liquid by a temperature of 1200 °C. arrangement (see Stephan, 1994, p. 109). The thesis of 4 The distinction between systemic and non-systemic mereological supervenience, however, is weaker than properties goes back to Broad (1933, p. 268). A more the thesis of synchronic determination, since it does recent version is from Bunge (1977, p. 501 f.). not claim the dependence of the system’s properties 5 Cf., e.g., Vollmer (1992, p. 188) and Lorenz (1988, from its micro-structure, it only claims their covari­ p. 48). ance. 642 A. Stephan • Varieties of Emergence in Artificial and Natural Systems mit, for example, that there may exist objects that closely connected: Irreducible systemic properties have the same parts in the same arrangement as are eo ipso unpredictable, in principle, before their diamonds, but which lack the diamond’s hardness, first appearance. But besides irreducible proper­ that may have hardness 2 instead of hardness 1 0 ties, there also seem to be properties that can’t be on the Mohs-scale. This seems to be totally im­ predicted before their first appearance on other plausible. Equally beyond thought is that there grounds. Therefore, the thesis of unpredictability may exist two micro-identical organisms, one is vi­ is more complex than the thesis of irreducibility of able and the other not. In the case of mental phe­ systemic properties. Thus, it is reasonable to start nomena, opinions may be more controversial; but with a discussion of the thesis of irreducibility, one thing seems to be clear: anyone who believes, which is easier to analyze, and a discussion of syn­ e.g., that two creatures identical in micro-structure chronic emergentism. could be such that one is colorblind while the Broad’s attempt to explicate a (strong) theory other can distinguish colors in the ordinary way, of emergence may count as downright classical; it does not hold a naturalistic-physicalistic position .7 reads: Weak emergentism as sketched so far comprises the minimal conditions for emergent properties. It “Put in abstract terms the emergent theory as­ is the common base for all stronger theories of serts that there are certain wholes, composed (say) emergence. Moreover - and this is a reason for of constituents A, B, and C in a relation R to each distinguishing it as a theory in its own right - it is other; that all wholes composed of constituents of held not only by some philosophers (e.g., Bunge, the same kind as A, B, and C in relations of the and Vollmer), but also by cognitive scientists (e.g., same kind as R have certain characteristic proper­ Hopfield, Rosch, Varela, and Rumelhart) in ex­ ties; that A, B, and C are capable of occuring in actly its weak form. The three features of weak other kinds of complex where the relation is not emergentism - (i) the thesis of physical monism, the same kind as R\ and that the characteristic (ii) the thesis of systemic properties, and (iii) the properties of the whole R(A,B,C) cannot, even in thesis of synchronic determination however, are theory, be deduced from the most complete knowl­ compatible with contemporary reductionist ap­ edge of the properties of A, B, and C in isolation proaches without further ado. Particularly, this is or in other wholes which are not of the form true because recent reductionist approaches - R(A,B,C)” (1925, 61). contrary to older variants - take into consider­ According to Broad’s definition a systemic ation the system’s structures too. Thus, a system’s property, which is supposed to be nomologically systemic property just being non-additive does not dependent on its system’s micro-structure (by the by itself make it irreducible. Some champions of thesis of synchronic determination), is called irre­ weak emergentism credit the compatibility of ducible and therefore emergent, if and only if it ‘emergence’ and ‘reducibility’ as one of its merits cannot be deduced from the arrangement of its compared to stronger versions of emergentism system’s parts and the properties they have ‘iso­ (e.g., Bunge, 1977). lated’ or in other (more simple) systems . 8 Although, prima facie, it looks as if Broad’s pro­ Synchronic emergentism posal gives us a clear and distinct explication of We come to the essential features of more ambi­ what it is for a systemic property to be irreducible tious theories of emergence, the theses of irreduc- (or non-deducible), a further look reveals that two ibility (or non-deducibility) and of unpredictability different kinds of irreducibility having quite dif­ of certain systemic properties. These theses are

8 Properties that might be ascribed to a system’s part 7 However, similar considerations hold for preposi­ ‘in isolation’ are, according to Broad, properties that tional attitudes only, as long as one does not subscribe depend essentially on the micro-structure of the part, to externalism, that is to say, if one does not claim while external factors, such as the part’s arrangement that, e.g., the content of a belief depends essentially and its neighboring parts, can be seen as almost irrele­ on the nature of the referents of the believer’s vant for the part’s having these properties (cf. 1919, thoughts and concepts. p. 112 f.). A. Stephan • Varieties of Emergence in Artificial and Natural Systems 643 ferent consequences are concealed. The failure to the colors or tastes of those objects) results al­ keep apart the two kinds of irreducibility has mud­ ready from a violation of the first condition, since dled the recent debate about the emergence of they were neither adequately characterizable by properties. To make things clearer, I shall first dis­ the macroscopic nor by microscopic behavior of cuss when a systemic property is reducible. For this the system’s parts, even in principle. For, when we to be the case, two conditions must be fulfilled: say that a certain object is red or a chemical sub­ The first is that from the behavior of the system’s stance has the smell of liquid ammonia, we do not parts alone it must follow that the system has some mean that the corresponding system’s parts behave property P. The second condition demands that or move in a certain way. No progress in the sci­ the behavior the system’s parts show when they ences could change this state of affairs in any are part of the system follows from the behavior way . 1 0 Broad has illustrated the fundamental dis­ they show in isolation or in simpler systems than tinction between (behavioral) analyzable and un- the system in question. If both conditions are ful­ analyzable properties by pointing to characteristic filled, the behavior of the system’s parts in other properties of organisms and secondary qualities, contexts reveals what systemic properties the ac­ respectively. tual system has. That is to say, those properties are If secondary and phenomenal qualities are not reducible. Since both conditions are independent analyzable , 11 even in principle, then there is no from each other, two totally different possibilities prospect that an increase of scientific knowledge for the occurrence of irreducible systemic proper­ will close the gap between physical processes and ties will result: (a) a systemic property P of a sys­ secondary qualities or between physiological pro­ tem S is irreducible , if it does not follow, even in cesses and phenomenal states of consciousness principle, from the behavior of the system’s parts (qualia), respectively. that S has property P; and (b) a systemic property We can now specify more exactly the feature of P of a system S is irreducible , if it does not follow, irreducibility which is central for synchronic emer­ even in principle, from the behavior of the sys­ gence. Its first variant is based on the behavioral tem’s parts in simpler constellations than S how unanalyzability of systemic properties. It reads: they will behave in S. (a°) Unanalyzability. Systemic properties which Thus, a necessary requirement for a systemic are not behaviorally analyzable - be it micro- property to be reducible is that its being ‘instanti­ or macro-scopically - are (necessarily) irre­ ated’ has to follow from the behavior of the sys­ ducible. tem’s parts. In other words: From the behavior of the system’s parts it should follow that the system However, even if secondary and phenomenal has all characteristic features that are essential for qualities belong to the class of unanalyzable prop­ having the systemic property. Broad, for example, erties, it does not follow that the specific behavior takes this condition, which is enclosed in the first of the system's parts upon which those qualities criterion for reducibility, to be always fulfilled in supervene is itself not deducible from the behavior the case of the characteristic properties of chemi­ those parts show isolated or in other (simpler) sys­ cal compounds and viable organisms. Their prop­ erties might be irreducible only by violation of the second criterion, what means that from the beha­ 10 However, whether reference to linguistic usage might vior of the system’s parts in other (simpler) sys­ answer questions concerning reducibility in a definite way is very controversial. Particularly, Churchland has tems it would not follow how they will behave in opposed arguments of the Broadian style heavily (see the actual system. In contrast, he claims that the 1988, p. 29 ff.). irreducibility of secondary qualities (e.g., the color 11 Properties that are called 'unanalyzable’ for simplicity or taste of certain objects) and phenomenal quali­ here, might be analyzable in other ways than by be­ havioral features. A certain smell, for example, might ties (e.g., the ‘how it is for us’ when experiencing be analyzed as a mixture of the smells of musk and fish-meal. This, however, would not be an analysis based on concepts of motion and behavior. The recent 9 As we will see, one type of irreducibility seems to im­ debate about qualia, particularly about the limits of ply 'downward causation’, while the other seems to functionalism can be seen as a repetition of Broad's imply epiphenomenalism. argumentation (see below). 644 A. Stephan • Varieties of Emergence in Artificial and Natural Systems tems. The irreducibility which results from a viola­ components’ behavior, does not imply, however, a tion of the first criterion of reducibility does not violation of the first criterion of reducibility. Sys­ imply, by itself, a violation of the second criterion temic properties that cannot be reduced because of reducibility. the system's parts’ behavior is irreducible might On the other side, however, even analyzable sys­ nevertheless be behaviorally analyzable. Hence, temic properties can be irreducible and therefore the two criteria of reducibility as well as those irre- emergent. This is the case when the second crite­ ducibilities that are based on the violation of these rion of reducibility will be violated, i.e., when the criteria are independent of each other. Summariz­ behavior of the system’s parts does not follow ing, we get from (a0) and (b°) the following modi­ from their behavior in other (simpler) constell­ fied version of systemic property irreducibility: ations. Broad thinks that such examples of irreduc­ (iv) Irreducibility. A systemic property is irreduc­ ible behavior might occur in chemical compounds ible if (a) it is neither micro- nor macro-scopi- and also in organisms . 1 2 His central idea is that the cally behaviorally analyzable, or if (b) the parts of a genuinely novel structure, such as, e.g., specific behavior of the system’s components, an organism in comparison to any inorganic com­ over which the systemic property supervenes, pound, might behave in a way that is not deducible does not follow from the component’s beha­ from the part’s behavior in other structures. Im­ vior in isolation or in other (simpler) constell­ plicitly, that means that the actual behavior of ations. parts that interact in wholes does not result from their behavior in pairs . 13 If the behavior of some Thus, we have to distinguish two totally dif­ system’s parts is irreducible in this respect, then ferent types of irreducibility of systemic proper­ all properties that depend nomologically on the ties. Equally different seem to be the conse­ behavior of the system’s parts (for example, repro­ quences that result from them. If a systemic duction) are irreducible too. property is irreducible because the behavior of the Thus, we can specify more precisely the second system’s parts, over which the property super­ variant of a systemic property’s irreducibility. It is venes, is itself irreducible, this seems to imply that based on the non-deducibility of the behavior of we have a case of ‘downward causation’. For, if the system’s parts: the components’ behavior is not reducible to their arrangement and the behavior they show in other (b°) Irreducibility of the components' behavior. (simpler) systems or in isolation, then there seems The specific behavior a system’s components to exist some ‘downward’ causal influence from within the system is irreducible if it does not the system itself or from its structure on the beha­ follow from the components’ behavior in iso­ vior of the system’s parts. To be sure, if there lation or in other (simpler) constellations. would exist such instances of ‘downward causa­ A violation of the second criterion of reducibil­ tion’ this would not amount to a violation of some ity, which is manifested in the irreducibility of the widely held assumptions, such as, for example, the principle of the causal closure of the physical do­ main. Within the physical domain, we would just 12 Bechtel holds a quite similar position: “although have to accept additional types of causal influ­ studying the properties of amino acids in isolation ences besides the already known basal types of may reveal their primary bonding properties, it may mutual interactions. not reveal to us those binding properties that give rise to secondary and tertiary structure when the amino In contrast, the occurrence of unanalyzable acids are incorporated into protein molecules” (1988, properties does not imply any kind of downward p. 95). In “Mechanical Explanation and its Alterna­ causation. Systems that have unanalyzable proper­ tives” Broad has examined ‘in abstracto’, under what ties that depend nomologically on their bearer’s conditions the behavior of system’s components can be irreducible (cf. 1919, p. 113f.). micro-structures need not be constituted in a way 13 Scenarios of this kind are discussed already by Mill that amounts to the irreducibility of their compo­ and Fechner (cf. Stephan, 1998, sections 6.1. and 7.3.); nents’ behavior. Nor is implied that the system’s see also McLaughlin who seems to think that emer- gentism has to allow configurational forces (1992, structure has a downward causal influence on the 52 ft). system’s parts. All the more, there is no reason A. Stephan • Varieties of Emergence in Artificial and Natural Systems 645 to assume that unanalyzable properties themselves However, bare addition of the thesis of novelty exert a causal influence on the system’s parts. does not turn a weak theory of emergence into Rather it is to ask, how unanalyzable properties a strong one, since reductive physicalism remains might have any causal role to play at all. Since compatible with such a variant of emergentism. they are not behaviorally analyzable, that is to say, Only the addition of the thesis of unpredictability , they neither seem to correspond to any ‘mecha­ in principle, of novel properties will lead to nism’ nor do they seem to result from any ‘mecha­ stronger forms of diachronic emergentism. nism’, it is hard to see how they could be causally A short consideration shows that systemic prop­ effective themselves. If, however, one can not see erties can be unpredictable in principle for two how unanalyzable properties might play a causal fundamentally different reasons: (i) they can be role, then, it seems, such properties are epipheno- unpredictable because the micro-structure of the mena. system, which exemplifies the property for the first time in evolution, is unpredictable. For, if the micro-structure of a newly emerging system is un­ Diachronic emergentism predictable, so are the properties which depend nomologically on it. (ii) However, a property can In a systematic examination of systemic proper­ be unpredictable even though the novel system’s ties that are exemplified in our world - for exam­ micro-structure is predictable. That is the case if ple, chemical, vital, or mental properties -, the the property itself is irreducible: For, if systemic question that arises is whether the properties are properties are irreducible, then they are unpredict­ reducible. In contrast, their predictability is, so to able before their first appearance. This does not speak, of no account. Within the scope of ‘emer­ preclude that further occurrences of this property gent evolution’, however, importance of the two might be predicted adequately. Since in the second questions seems the reverse: While the question case criteria for being unpredictable are identical concerning the reducibility of emerging properties with those for being irreducible, this notion of un­ seems less relevant, it is of particular interest what predictability will offer no theoretical gains be­ systems and properties might have been predicta­ yond those afforded by the notion of irreducibil­ ble, at least in principle, before they were actually ity . 1 4 exemplified. Likewise, predictability of properties Let us focus, therefore, on the first case: unpre­ that are be expected plays an important role in the dictability of structure. This version of unpredict­ development of novel artefacts. Here, however, it ability passed almost unnoticed in ‘classical’ litera­ is not predictability in principle that matters. What ture on emergentism during the 1920s, but because matters is practical predictability - it’s better to of strong interest in dynamical systems and chaotic know before you make the ‘Elch test’ whether processes this notion gains considerable signifi­ your newly constructed car will pass it or not. Un­ cance. foreseen events of this type have little to do, how­ The structure of new formed systems can itself ever, with emergence in the theoretically interest­ be unpredictable for several reasons. Thus, belief ing sense. in an indeterministic universe implies that there All diachronic theories of emergence have at bottom a thesis about the occurrence of genuine 14 A difference in extension between both notions could novelties - properties or structures - in evolution. result only in respect to those properties, which, al­ This thesis excludes at the same time all preforma- though reducible, are not predictable before their first tionist positions. appearance. A reducible property is unpredictable be­ fore its first appearance if the behavior of the system’s (v) Novelty. In the course of evolution exemplifi­ components upon which it supervenes does not follow from their behavior in those systems that exist at the cations of ‘genuine novelties’ occur again and time of prediction. The notion of unpredictability wid­ again. Already existing building blocks will ened in this way would depend, however, in a very develop new constellations; new structures contingent way on the chronological order of systems coming into being in evolution. Therefore, such a no­ will be formed that constitute new entities tion should not be important in qualifying the notion with new properties and behaviors. of emergence. 646 A. Stephan • Varieties of Emergence in Artificial and Natural Systems will be novel, unpredictable structures. However, Hereafter, predictions do not seem to be possible. from an emergentist perspective it would be of no Marginally different initial values x lead to radi­ interest, if a new structure’s appearance would be cally different trajectories of the iterated function. unpredictable only because its coming into being Values jump pell-mell, convergence and diver­ is not determined, not to mention that most emer- gence are not discernible: chaos dominates. gentists claim, anyway, that the development of Thus, it looks as if just the most exact science of new structures is governed by deterministic laws. all has led us back to one of the starting points of But still deterministic formings of new structures emergentism. Whereas - after pioneering suc­ can be unpredictable in principle , if they are gov­ cesses in chemistry and physics - we today do not erned by laws which are attributed to determinis­ count properties and dispositions of chemical com­ tic chaos. pounds any more among synchronic emergent An essential outcome of the theory of chaos is phenomena, examinations of deterministic chaos that there exist - even very simple - mathemati­ suggest the existence of systems that might de­ cal functions, whose own ‘behavior’ cannot be pre­ velop structures that are unpredictable in principle dicted. Only the rise of ‘experimental mathemat­ and thus might show structure-emergent behavior. ics’ on highly efficient computers has revealed, for Of course, one could argue that a Laplacean cal­ example, the properties of various logistic func­ culator could predict correctly even chaotic pro­ tions. Their intra-mathematical unpredictability cesses. Whether or not this could actually be the has to do with an aperiodic behavior of these func­ case, however, is not settled yet. It depends mainly tions, by which marginally different initial values on the question of what kind of information we of some variable can lead to radically distinct tra­ allow such a creature of phantasy to have. For ex­ jectories of the functions. ample, in Alexander’s considerations (cf. 1920, ii, A standard example is the logistic function pp. 72 f., 328) Laplace’s calculator knows several f(x) = |ix (l-x ) for 0 < x < 1. For a parameter earlier states of the whole world and, in addition, with 0 < [i < 4 the logistic function maps the in­ all natural laws that govern changes in the world. terval [0,1] onto itself. Of particular interest is, He seems to be able to extrapolate from his how parameter (i exercises an influence on the knowledge of all events that have occurred in the long term behavior of the function when iterated universe so far even the course of chaotic pro­ repeatedly. For 0 < < 1 the situation is obvious. cesses. But on what basis could he do that? Since All initial values of the variable x let the function chaotic processes are aperiodic, one can not deter­ f(x) approximate the value 0 after sufficiently mine definitely from those processes that have oc­ many iterations, thus, the origin is the attractor. curred up to a certain time the exact formula For 1 < (I < 3 exists exactly one attractor A of which would describe their further course. Even value A = l-l/[i: the function balances out on a if the further course of the world is governed by stable value. If equals 3, the fixed point of the deterministic laws, it does not follow from the function is ‘marginally stable’; convergence is de­ earlier events and states alone, by which laws it is cidedly slowly - an indication for fundamental governed. Entirely different continuations seem to change in its behavior. For larger values dynamic be compatible with the earlier course of the world. becomes considerably complex. In the case of 3 < Therefore, even a Laplacean calculator could fail

^ < 1 + ^ 6 values oscillate between two fixed in his predictions. If one grants, however, that he points. By increasing \x the attractors of period two knows all details of earlier world states - up to will become instable, too. We get a cycle of period infinitely many digits - , and if one grants that he four (i.e., after four iterations the values of the knows a priori which processes are governed by function approach in each case the four fixed which specific chaotic laws, then, of course, he points). At 3.56 the period doubles again and be­ would be able to predict the forming of structures comes eight, at 3.567 it becomes sixteen, and then that are governed by these laws . 1 5 I will leave it we get a quickly rising sequence of periods to 32, 64, 128, etc. - vividly one speaks of cascades. At 15 Laplace himself assumed that his calculator knows all laws governing nature, but he took those laws to be about 3.58, this sequence comes to an end. The laws of Newtonian physics. During his time, nobody period has doubled itself infinitely many times. knew anything about aperiodic chaotic processes. A. Stephan • Varieties of Emergence in Artificial and Natural Systems 647 open whether or not it is plausible to ascribe such Synopsis a knowledge to such a fabulous creature. How­ The following figure depicts the logical relation­ ever, we can preclude that foretellers of our men­ ship that holds between different versions of emer- tal capacities have these abilities, and suppose that gentism. where chaos exists, structures exist that are unpre­ Weak diachronic emergentism results from weak dictable in principle, and that is to say, that there emergentism by adding a temporal dimension in will be structure emergence in our sense. the form of the thesis of novelty. Both versions (vi) Structure-unpredictability. The rise of novel are compatible with reductive physicalism. Weak structures is unpredictable in principle, if theories of emergence are used today mainly in their formation is governed by laws of deter­ cognitive sciences, particularly for characterization ministic chaos. Likewise, any novel properties of systemic properties of connectionist nets, and in that are instantiated by those structures are theories of self-organization. Synchronic emer­ unpredictable in principle. gentism results from weak emergentism by adding the thesis of irreducibility. This version of emer­ Summing up, it may be said a systemic property gentism is important for the philosophy of mind, is unpredictable in principle before its first appear­ particularly for debating nonreductive physicalism ance, if (i) it is irreducible, or if (ii) the structure and qualia. It is not compatible with reductive which instantiates it, is unpredictable in principle physicalism any more. Strong diachronic emer­ before its first formation. Although unpredictabil­ gentism only differs from synchronic emergentism ity of structure always implies unpredictability of because of the temporal dimension in the thesis of properties instantiated by the structure, it does not novelty. In contrast, structure emergentism is en­ thereby imply the irreducibility of the properties tirely independent of synchronic emergentism. It instantiated by the structure. As far as that goes, results from weak emergentism by adding the the­ unpredictability in principle of systemic properties sis of structure-unpredictability. Although struc­ is entirely compatible with their being reducible to ture emergentism emphasizes the boundaries of the micro-structure of the system that instanti­ prediction within physicalistic approaches, it is ates them. compatible with reductive physicalism, and so it is far weaker than synchronic emergentism. Theories of deterministic chaos (in dynamical systems) can be acknowledged as a type of structure emer-

+ novelty weak emergentism ------^ weak diachronic emergentism

strong diachronic structure emergentism Fig. 1 depicts the logical relationships that hold between different versions of emergentism. 648 A. Stephan Varieties of Emergence in Artificial and Natural Systems

gentism. Likewise, its perspective is important for edge of natural laws helps us understand why the evolutionary research. In comparison to the above motion of molecules has exactly the causal role mentioned versions of emergentism the synthetic usually ascribed to heat. In doing so, it is presup­ position of strong structure emergentism has no posed - and this is of great importance - that equivalent in recent discussion. Most important the macro-physical concept of heat can be fully from a theoretical point of view are weak emer­ explicated by heat’s causal role. gentism , synchronic emergentism, and diachronic structure emergentism. “[Statement (ii)] is explanatory in the sense that our knowledge of chemistry and physics makes in­ Emergence in the Qualia Debate, in Connection- telligible how it is that something like the motion ism, and in Synergetics of molecules could play the causal role we associ­ ate with heat. Furthermore, antecedent to our dis­ Let us turn to some specific cases to consider covery of the essential nature of heat, its causal how the different concepts of emergence are ap­ role [...] exhausts our notion of it. Once we under­ plicable. We will see that in one of the central de­ stand how this causal role is carried out there is bates within philosophy of mind, namely the de­ nothing more we need to understand” (1983, bate about the nature of qualia, there is need for p. 357). a strong notion of emergence such as that of syn­ chronic emergentism. In contrast, in connection- In other words: statement (ii) is fully explana­ ism and synergetics weaker notions of emergence tory because some system’s property of heat is re­ are employed. ducible in respect to the motions of molecules, and so far it is not emergent. On the other hand, the Emergence in the debate over whether qualia are reason for statement (i) not being fully explana­ physical tory is, according to Levine, that the notion of pain (as are notions of other phenomenal states) is not In recent debate about qualia , 1 6 Nagel, Block, exhausted by the causal role of pain. The decisive Jackson, Levine, and McGinn, among others, have point is that there is no reason to believe that fir­ argued in one way or another that qualitative ing of C-fibers fits better to typical pain experi­ mental phenomena are not reducible to physical ences than to any other phenomenal experience. or functional states, respectively. If their argu­ In a series of articles, Levine and Hardin argued ments succeed, they imply emergentist or sub- about the ‘explanatory gap’-argument’s core thesis stance-dualistic positions. Most interesting and by concentrating on experiences of colors (see powerful seems to be Levine’s so-called ‘explana­ Levine, 1983; 1991; and Hardin, 1987; 1988; 1991). tory gap’-argument, which I will consider closely Levine’s last ‘move’ is the following : 1 8 in the following. Levine starts with comparing two statements, “If inverted qualia are possible, then the ques­ namely (i) ‘pain is the firing of C-fibers’, and (ii) tion why red things look reddish and not greenish ‘heat is the motion of molecules’. The decisive dif­ has no adequate answer in physical/functional ference between the two identity statements, ac­ terms. Whereas, if inverted qualia are not possible, cording to Levine, is that the second is fully ex­ then the question why red things look reddish, and

planatory, while the first is not : 1 7 “there is a ‘gap’ not some other (perhaps unimaginable) way or no in the explanatory import of these statements” way at all, still has no adequate answer in physical/ (1983, p. 357). The second identity statement is as­ functional terms. An explanatory gap persists” sumed to be ‘fully explanatory’, because knowl­ (1991, p. 39). Levine here seems to refer to objections that 16 The term ‘qualia’ denotes the experiential properties were originally directed at functionalism, namely of conscious states, i.e. of sensations, feelings, percep­ the ‘inverted qualia’-argument, and the ‘absent tions, and of propositional attitudes. 17 Talk about Tiring of C-fibers’ has its firm place in lit­ erature about the qualia-problem. It is, however, only a fill-in for an adequate neurophysiological analysis of 18 Articles Levine has published subsequently have not a pain’s material base. changed the debate considerably. A. Stephan • Varieties of Emergence in Artificial and Natural Systems 649 quaha’-argument.1 In his article “On Leaving Out seems to admit that prospects are slim to answer What It’s Like” he confirms this: “The basis of my the really central question whether perfect knowl­ argument for the existence of this explanatory gap edge of neuronal processes could allow to deduce was the conceivability of a creature’s instantiating statements about the qualitative states that are the physico-functional property in question while correlated with (or supervene upon) such pro­ not undergoing an experience with the qualitative cesses.20 character in question, or any qualitative character Let’s see why qualia are so resistent to reduc­ at all” (1993, p. 130). The ‘explanatory gap’-argu- tion. According to Levine, reduction that is ex­ ment, however, does not depend on the empirical planatory requires two stages: “Stage 1 involves possibility of absent or inverted qualia. Even if the (relatively? quasi?) a priori process of working they were not actually possible, there is no way to the concept of the property to be reduced ‘into see how qualia could be reduced to neurophysio- shape’ for reduction by identifying the causal role logical or functional states. That is because it does for which we are seeking the underlying mecha­ not even seem to be possible in principle to de­ nisms. Stage 2 involves the empirical work of dis­ duce from generally obtaining natural laws that covering just what those underlying mechanisms the micro-structures of those systems that have are” (1993, p. 132). phenomenal states have all features characteristic If one claims that a reduction that is explanatory of phenomenal states. is impossible in principle, as is claimed for qualia, In response to Levine, Hardin has tried to make that does not imply a failure of the second task. clear which explanatory gap he thinks is closable What is implied is a failure, in principle, of the first at least in principle, and which one must be left task. Apparently, phenomenal properties cannot open. If we assume that color experiences are phe­ be individuated by their causal roles: “What seems nomenally simple and unstructered, then it looks to be responsible for the explanatory gap, then, is difficult indeed to tie them explanatorily to physi­ the fact that our concepts of qualitative character ological processes. In fact, however, we should do not represent, at least in terms of their psycho­ acknowledge that at least some colors show struc­ logical contents, causal roles. Reduction is explan­ tures that fit well to neuronal processes. atory when by reducing an object or property we reveal the mechanisms by which the causal role “We would for example, expect the neural pro­ constitutive of that object or property is realized. cess associated with orange to be relatively more Moreover, this seems to be the only way that a complex than the process associated with red, and reduction could be explanatory. Thus, to the extent in fact, expect that the ‘red-making’ process would that there is an element in our concept of qualita­ be in some fashion incorporated in the ‘orange- tive character that is not captured by features of making’ process on the grounds that perceived its causal role, to that extent it will escape the ex­ redness is an ingredient in perceived orangeness” planatory net of a physicalistic reduction” (1993, (1991, p. 44). p. 134). Thus, Levines synchronic qualia-emer- Hardin concedes, however, that it is far easier gentism is based on two theses: to explain the difference between experiencing red (1) The reduction of a systemic property P is ex­ and experiencing green than to explain experienc­ planatory if and only if the realization base ing red or experiencing green themselves. We exhausts exactly the causal role which is con­ could develop an understanding of the differences stitutive of P. between both experiences without being able to explain each experience in itself. Thus, Hardin (2) Phenomenal properties (or states) are not fully graspable by the features of their causal role.21 19 Inverted qualia would exist, e.g., if things we call ‘red' look to somebody else the way things look we call ‘green’, and vice versa. Qualia are called absent, if a 20 In the meantime, others have suggested further ways system has no phenomenal experiences at all, al­ to close the ‘explanatory gap’. See, for example, Kirk though it shares with us functional descriptions of (1996) and Kurthen (1996). phenomenal states. See, for example, Block (1978; 21 Notice the analogies between Levine’s theses and the 1980). criteria for irreducibility in the section above. 650 A. Stephan • Varieties of Emergence in Artificial and Natural Systems

Several responses are possible to this analysis: tion.23 In the presence of this, the challenge for (i) one tries to avoid emergentism by showing the physicalist to manage those ‘portions’ of qualia either how properties that are not fully graspable that escape individuation by causal roles seem to via their causal roles, might yet be reducible ex­ be negligible compared to the emergentist’s prob­ planatorily, or by showing that phenomenal quali­ lems with mental causation. For, if qualitative ties are graspable adequately via causal roles. For states can not be individuated via causal roles, both variants, however, there exists nothing more then it is questionable whether they can play any than first attempts or statements of intentions.22 causal role at all. (ii) One accepts both theses and, thereby, an emer- gentist position. But then, additional questions Synergetics and emergentism arise: To what type of psycho-physical theories Synergetics - an interdisciplinary theory of co­ does synchronic qualia-emergentism belong? Can operation - addresses an original subject of emer­ it still be seen as a physicalistic position, or has it gentism: the coming into being of novel systemic to be treated as a kind of property dualism? And properties. Particularly, it is concerned with quali­ furthermore: What solution can emergentism offer tative changes in a system’s behavior that have be­ to the problem of mental causation? come instable by change of specific controlling Perspectives on whether an emergentist theory parameters (see Haken, 1996, pp. 587 and 593). of qualia can be seen as a physicalist position have Synergetics’ starting point is the observation that changed considerably during the past twenty both in animate and in inanimate nature, there ex­ years. The older arguments presented by Nagel, ist numerous systems that spontaneously form by Kripke, and Jackson had, among other things, the themselves - that is to say, in a self-organizing aim to show that physicalism is an inadequate po­ way - new structures. Classical examples are the sition. Levine changed the situation by pointing to entirely novel light of lasers in comparison with the fact that various ‘conceivability’-arguments do that of ordinary lamps,24 or the structure forma­ not necessarily have ontological consequences: tion of slime mold.25 Both examples and many “If these thought experiments show that physi­ similar ones have in common that small distur­ calism leaves something out, it can’t be in the bances of control parameters give rise to instable sense that there are facts that physicalistic descrip­ states out of which new structures emerge by pro­ tions fail to pick out, [...] perhaps, Cartesian con- cesses of self-organization. The new structures ceivability arguments can’t demonstrate that qua­ often instantiate qualitatively new properties com­ lia aren’t physical states or processes, but they at pared to those the system instantiated before - a least throw the burden of argument back onto the subject that has been in the center of emer- physicalist to show why we should think they are gentists’ considerations. physical states or processes” (1993, 125 f.). Synergetics looks for the regularities that lie be­ hind various self-organizing systems which usually Levine’s interpretation would allow, therefore, to treat synchronic qualia-emergentism as a physi­ calistic position. As things stand, however, physi- 23 The seeming dilemma, however, is not hopeless for calists have turned the tables. They demand that the emergentist, as I argued in “Armchair Arguments genuine physicalist positions may not leave out ex­ Against Emergentism” (1997). planatorily anything (cf. Beckermann, 1996b and 24 When only a small amount of energy is pumped into the device, the laser operates as a lamp. Then, its main Horgan, 1993). According to this criterion, qualia- elements, namely specific atoms, emit lightwaves of emergentism can not be seen as a physicalistic po­ about 3 m coherence length independently of each sition any more. One reason for the physicalists’ other. If the pump power is increased to a certain threshold, the atoms oscillate in phase and emit a becoming more fastidious might be that emer- giant wavetrack of about 300,000 km length. gentists seem to face a hopeless dilemma when 25 Slime mold (Dictyostelium disciodeum) normally ex­ confronted with the problem of mental causa- ists in the form of single amoebic cells. If nutriment runs short, the cells assemble at a certain place, pile up, and then differentiate into spores and stalk cells. Slime mold can move then as a whole. It wriggles on 22 Cf. to Beckermann, Kirk, Levine and Rey (all 1996). the ground like a snake (cf. Haken, 1988, p. 99-102). A. Stephan • Varieties of Emergence in Artificial and Natural Systems 651 are studied by such distinct disciplines as physics, plitude of emitted light, is sufficient to describe chemistry, or biology. Such as theories of emer­ characteristic features of lasers. Thus, transition gence synergetics has as a characteristic feature a from microscopic to macroscopic descriptions sig­ naturalistic attitude: It takes it for granted that nificantly reduces the information which is neces­ processes which lead to new structures are trig­ sary to characterize systems adequately (cf. 1987, gered and maintained by the system’s parts’ coop­ p. 140). However, from obvious compressibility of eration, and not by external ‘organizers’ or ‘super­ information Haken seems to infer compressibility natural’ entities. In contrast to stronger versions of causal factors; for he is interpreting order of emergentism, synergetics claims that formation parameters that are employed to describe a system of new structures as well as properties instantiated in a simple way as causal agents: by them can be explained and forecasted by refer­ ence to the parts’ cooperation. Hence, from a syn­ ( t c ) The behavior of the system’s parts is fixed by ergetic point of view neither should new arising just a few quantities, namely the order structures be seen as structure emergent, nor parameters. should properties instantiated by these structures be seen as synchronic emergent. Sometimes Haken puts the ‘thesis of causality’ Of central significance for synergetics are such ( t c ) in a specific jargon of synergetics, and calls it notions as ‘order parameter’ or ‘enslaving’ intro­ the ‘enslaving principle’ ( e p ) , which strictly speak­ duced by Haken himself. Quantities that are seen ing means that “slowly changing quantities (e.g., as order parameters are thought to do a double gradually increasing order parameters) enslave job: On the one hand, they are used to describe rapidly relaxing motions” (Haken and Haken- very complex processes of self-organization in a Krell, 1989, p. 27, my translation; see also Haken, simple but yet adequate way; on the other hand, 1988, p. 20 and 1996, p. 588). they are used to explain causally the ongoing pro­ cesses (cf. 1996, p. 588, and 1987, p. 139 f.). Haken’s (e p ) Order parameters enslave a system’s indivi­ ‘thesis of description’ ( t d ) is supported by numer­ dual parts, that is to say, they determine the ous examples, and by a mathematical formalism parts’ behavior. developed in (1983):26 The principle of ‘enslaving’ a system’s parts by ( t d ) By specifying the behavior of order parame­ order parameters is a specific case of downward ters of a system, the system’s behavior is ade­ causation: Order parameters are macroscopic quately characterized. Instead of describing quantities. They refer to states or properties of the a system’s behavior by specifying the beha­ whole system. As systemic quantities, they influ­ vior of all its single parts, it is sufficient just ence the system’s part’s behavior. Since Haken to specify the behavior of few relevant or­ supposes that order parameters emerge only from der parameters. cooperation of the system’s parts, he prefers to use Thus, the state of a laser can be described, in the notion ‘circular causality’ to characterize the principle, in two ways: first, on a microscopic level presumed mutual interaction between system’s by specifying the individual states of all electrons, parts and order parameters. or second, by exploiting that laser’s light has come The term ‘enslaving principle’ is not to be up to a macroscopic state. Here, a far smaller understood as a metaphor, although Haken’s dic­ amount of information, e.g., wavelength and am­ tion is full of anthropomorphisms and metaphors when he undertakes to illustrate it. Originally, it is introduced and justified within a presentation of a 26 To give some examples, order parameters are, accord­ mathematical formalism that is developed to de­ ing to Haken, prevailing light waves in a laser which ‘force’ all emitted electrons to oscillate in their way scribe adequately processes of self-organization (cf. 1988, p. 68 ), spiral fields of concentration of (see 1983, p. 2 0 8 ff.). In his own interpretation of cAMP-molecules (cyclic Adenosin Monophosphate) the mathematical formalism, Haken, however, of the slime mold (1987, p. 142) as well as language, form of government, or working climate in social sys­ jumps from a purely structural reading to a causal tems (1996, p. 590). reading of the mathematical relations - a clear 652 A. Stephan Varieties of Emergence in Artificial and Natural Systems case of the ‘post hoc, ergo propter hoc'-fallacy.27 properties or macroscopic quantities such as order For, the functionally described relation of succes­ parameters as irreducible and thus synchronically sion between order parameters and system’s parts, emergent. Instead, synergetics attempts to explain does not warrant an extrapolation to a causal rela­ the coming into being and persistence of novel tion between order parameters and system’s parts. structures and their properties by processes of self­ Haken’s causal theses appear rather odd when organization. Moreover, synergetics does not es­ applied to order parameters within the scope of tablish a scientifically based kind of downward psycho-physical theories or approaches in the so­ causation, since a reading of the mathematical for­ cial sciences. malism that jumps from a sequential relation to a downward causal relation is not warranted. “Let us choose here an extreme case, the brain. We treat neurons and their connections as sub-sys­ Em ergence in connectionism tems. Chemical and electrical activities of neurons can be described by numerous microscopical vari­ In the last decade, connectionism has received ables. However, strictly speaking it is the thoughts great attention in cognitive science. Its core idea that are efficacious as order parameters. Both is to assume a network of elementary units that components are dependent on each other” (1983, have a certain level of activation. Units are con­ p. 15 f., my translation).28 nected with each other. Units whose activation ex­ ceeds a certain threshold, can activate or inhibit Haken’s claim implies that thoughts ‘enslave’ other units according to certain weights that spec­ neurophysiological processes. Both in this case ify prevailing connections. and in social systems, Haken ascribes to order To see to what extent emergentist considera­ parameters a causal role they just do not have. tions are relevant for connectionism, I shall first There is no modus operandi according to which examine more closely the parameters that specify the working climate ‘enslaves’ some clerk (see Ha­ a connectionist net. Each net is determined essen­ ken, 1996, p. 590). The working climate does noth­ tially by three factors: (i) by the number of units ing at all. It is an expression of the atmosphere and connecting links which hold between them; within a group, it tells something - by compress­ (ii) by the function that determines the level of ing information! - about the manner of how mem­ activation for each unit; and (iii) by the rule that bers of a group deal mostly with each other, but it determines how connection weights will change. does not deal itself with the members of the group. In each case the number of units and the links The same is true for other order parameters of between them are fixed; they make up, so to social systems listed by Haken (ib.): Neither na­ speak, the ‘skeleton’ of a network which is static tional character, public opinion, nor ethics cause under ordinary circumstances: neither the number anything by themselves. Nor is anybody ‘enslaved’ of units nor the structure of their links will change. by them. A system’s actual dynamic results from the possi­ To sum up, we can say that synergetics does nei­ bility of modifying the weights of its internal con­ ther establish a strong nor a novel version of emer­ necting links. From a macroscopic point of view, gentism. It does not treat novel structures that are these continuous processes of accomodation can formed at instabilities as unpredictable and thus be seen as learning procedures. Thus, a connec- as structure emergent.29 Nor does it treat systemic tionist net ‘learns’ by locally determined changes of its connections’ weights, and not by adding 27 From a mathematical point of view Haken’s approach some further propositions to its data base. It’s a just means that solutions of differential equations de­ big challenge for connectionist researchers to con­ pend essentially upon slowly changing quantities, the order parameters, while fast-moving quantities can be strue nets that are able to optimize link weights neglected. A more detailed presentation of Haken's by themselves in a ‘training phase’ such that they mathematical considerations is given in Stephan are able to produce given outputs. When a net is (1998, section 18.1.). 28 See also Haken and Haken-Krell (1989, p. 134). 29 Thus, Haken says, “it is exactly calculable which col­ lective motion will win in the end” (1988, p. 49, my pend on very small deviations, and so are often unpre­ translation). It is only that breakings of symmetry de­ dictable for practical reasons. A. Stephan • Varieties of Emergence in Artificial and Natural Systems 653 fed with various inputs after its practice time, it in a strong sense, namely synchronically emergent. will calculate its outputs or ‘answers’ with stable Secondly, many systemic properties of nets seem weights in general. to be emergent in a ‘phenomenological’ way. By Nets are most efficient in pattern recognition this, I mean that the properties appear, or come and generalization (e.g., in acquiring rules and into being by themselves - as is the case with the forming schemata). Rumelhart, Smolensky, five rooms’ schemata - , if nets get adequate stim­ McClelland, and Hinton describe a simple net that uli. These facts of the case are referred to in Eng­ is an ideal case in point to illustrate how connec- lish by the word ‘emergent’ in its ordinary use, tionist nets instantiate schemata:30 however. No specific theory of emergence is im­ We all have ideas of what typical functional plied by this usage. Since some connectionists rooms look like. Living rooms have sofas and easy make use of the word ‘emergent’ in its ordinary chairs, but they usually do not have bath tubs and use intermingled with a more technical use, it is refrigerators; in offices and studies we might find very important to tell the notions apart. Eventu­ desks, bookshelfs, and computers, but no toilets. ally, connectionist nets develop during their train­ Rumelhart et al. represented this knowledge about ing phase - in a somewhat mini-evolutionary pro­ specific rooms’ typical furniture and equipment in cess - their ‘soft’ structures, by which I a connectionist net. The net consisted of 40 units understand the specific distribution of link each of which represented a specific room feature. weights.31 The net’s systemic properties discussed Since this example was to illustrate schemata for­ above supervene upon these structures. Now, the mation they refrained from training the net; in­ question arises whether this formation of structure stead, they set the link weights in advance. is an interesting case of structure emergence. By way of calculation there are 240 possible bi­ Let’s examine first the relationship between nary states in which the system could potentially global net properties and their realization base, settle. But actually, only five maxima which corre­ namely the net’s structure and its part’s properties. spond to five specific room types crystallized. For Considerations of connectionist net’s architectures example, by beginning a cycle with the ‘oven’-unit, and their modes of operation reveal that only the system successively adds the units for ‘ceiling’, trained nets show typical macroscopic properties ‘coffe-cup’, ‘sink’ and ‘refrigerator’. After 400 up­ such as ‘rule following’, ‘schemata formation’, or dates all features of a prototype kitchen were acti­ ‘pattern recognition’. Untrained nets do not have vated (see Rumelhart et al., 1986, p. 25). Complete those (cognitive) properties, they have only the schematas for kitchen, bathroom, office, living- disposition to acquire them. Macroscopic proper­ and bedroom could be activated by starting with ties of trained nets supervene upon both their one of these room’s characteristic features, respec­ given hard structure, and their acquired soft struc­ tively. Besides these main schemata several ‘sub­ ture. They are fully reducible to the organization schemata’ are hidden. They can be activated by of the net in consideration, the properties of its activating, e.g., ‘bed’- and ‘desk’-unit together. units (namely their activation formula), and the Then we get a schema of, say, a student’s room. properties of links consisting between its units (na- Behavior and properties of connectionist nets meley distribution of weights, and formula for give rise to emergentist considerations in many changing weights). If these quantities are known, ways. Three aspects should be discerned: first, con- the output-behavior of any net can be predicted nectionist nets obviously have systemic properties, exactly and explained. It is obvious that a net’s that is they have properties their parts do not parts, namely its units and the links between them, have. Thus, a net’s properties are at least weakly do not have any of those macroscopic (cognitive) emergent. However, it remains to be determined properties. So far, these properties a net acquires whether systemic properties of nets are emergent by training are typical systemic properties. How­ ever, since they are not irreducible systemic prop­

30 The term ‘schemata’ refers as related terms as ‘scripts’, or ‘frames’ to coherent structures of knowl­ 31 The ‘hard’ structure of a net (its ‘skeleton’) is fixed edge, which are assumed to be fundamental for rea­ by the number of units and the links between them. sonable interactions with the world. Usually, it is invariant. 654 A. Stephan • Varieties of Emergence in Artificial and Natural Systems erties, but are completely deducible from a net’s croscopic properties emerge from micro-level structure, and its parts’ and links’ properties, a interactions: “[M]any of the constructs of macro­ net’s systemic properties are merely weakly emer­ level descriptions such as schemata, prototypes, gent. They are not synchronically emergent. rules, productions, etc. can be viewed as emerging Rumelhart and McClelland discuss in great de­ out of interactions of the microstructure of distrib­ tail net properties which they call ‘emergent’, and uted models” (1986, p. 125; italics are mine). By stress that connectionist approaches do not imply this characterization, they lay stress upon the fact reductionistic, but interactionistic positions. that a net’s systemic properties come into being from the complex behavior of the net’s compo­ “We are simply trying to understand the essence nents: rules and schemata can be available without of cognition as a property emerging from the in­ being explicitly fed into the system. Clark dis­ teractions of connected units in networks. We cer­ cusses a net’s ability to develop schemata from this tainly believe in emergent phenomena in the sense point of view, too. He refers to Rumelhart, Smo­ of phenomena which could never be understood lensky, McClelland, and Hinton: “[they] detail a or predicted by a study of the lower level elements PDP model in which the properties of explicit in isolation. [...] This is the case in many fields. stored schemata emerge simply from the activity For example, we could not know about diamonds of a network of units that respond to the presence through the study of isolated atoms [...] Features or absence of microfeatures of the schemata in such as the hardness of the diamond is under­ question” (1989, p. 93; my italics).32 And he adds: standable through the interaction of the carbon “activating the unit for one prominent kitchen fea­ atoms and the way they line up. [...] Knowing ture (e.g., the cooker) would activate in a kind of about the individuals tells us little about the struc­ chain all and only the units standing for items ture of the organization, but we can’t understand commonly found in kitchens. Here, then, we have the structure of the higher level organizations an emergent schema in its grossest form” (1989, without knowing a good deal about individuals p. 94). and how they function. This is the sense of emer­ Within the quoted texts, it is not emergence in gence we are comfortable with” (1986, p. 128). a technical sense that is being discussed, but rather However, Rumelhart and McClelland’s claim a system’s abilities to acquire systemic properties that connectionism amounts to a non-reductive by self-organizational processes, which can be as­ position results from an inappropriate strong no­ cribed quite adequately by an ordinary use of the tion of reduction. Both authors, one must know, term ‘emergent’. The temporary manifestation of take a system’s properties as reduced only when schemata that were already latently ‘in’ the link’s either some of the system’s parts have these prop­ weights is interpreted, then, as an emergent prop­ erties already, or when the system’s property can erty of the net. Connectionists, thus, mainly point be reduced to some linear interactions of the parts. to ‘emergent rules’ or ‘emergent schemata’ to de­ Both possibilities of reduction, however, are given marcate their position from ‘classical’ representa- neither in the case of carbonaceous compounds, tionalism, accordingly to which all rules and sche­ nor in the case of connectionist nets. Yet, almost mata have to be fed in explicitely (see Horgan and no systemic property would be reducible accord­ Tienson, 1996). ing to these criteria. At the same time, Rumelhart There is a further feature of connectionist nets and McClelland concede that a net’s behavior is that provokes emergentist considerations: during completely intelligible, if one takes into account the phase of training or learning a net runs all interactions between its units, that is to say, if through a mini-evolutionary process - link one considers a system’s hard and soft structure. weights are adapted such that the net is enabled Hence, even from their point of view, connection­ to handle the tasks it is supposed to master. Within ism is not an instance of synchronic emergentism. In discussions of connectionist nets’ behavior ordinary use and technical use of the notion of 32 The term ‘PDP model’ refers to connectionist nets, too. It is an abbreviation for a kind of processing go­ emergence often gets intermixed. Thus, for exam­ ing on in them, namely ‘parallel distributed pro­ ple, Rumelhart and McClelland reiterate that ma­ cessing’. A. Stephan • Varieties of Emergence in Artificial and Natural Systems 655

this time the so-called ‘soft structure’ of the net macroscopic properties of nets are emergent, all develops.33 Only when the links’ weights are ad­ systemic properties of complex systems are emer­ justed adequately does a net has available desired gent. A difference to many other systems exists at macro-properties, that is, only then can it develop best in the plasticity of nets and in their capacity schemata, recognize patterns, or make use of rules. to develop in a training phase by themselves ade­ Those are not implemented explicitely as they are quate ‘attractors’ to cope with given tasks. Corre­ in symbol manipulating devices, but are extracted sponding macroscopic properties will become from given material. However, this is not a case manifest only temporarily during treatments. This of genuine structure emergence: not only does the second order property might justifiably be charac­ distribution of weights result from deterministic terized as ‘phenomenological’ emergence. Only principles, the changes of weights are even calcula­ some dynamical systems have this property. ble exactly, if we know the learning rule, the acti­ vation formula, the unit’s initial activation, the in- Conclusion tial weights, and the inputs. Were this not the case, In particular, I have distinguished three versions Rumelhart et al. could not have skipped the train­ of emergentism: weak emergentism, (strong) syn­ ing phase to calculate weights ‘on foot’ and feed chronic emergentism, and diachronic structure them into the system in advance. emergentism. Synchronic emergentism results Even though we therefore should not speak of from weak emergentism by adding the thesis of structure emergence in connectionist networks, re­ irreducibility. It turned out that this version of garding their soft structure, nets show a tremen­ emergentism is important for the philosophy of dous plasticity, when compared with other objects, mind, particularly for debating qualia. However, even when compared with other dynamical sys­ this does not establish that qualia are emergent tems. Chemical compounds, to give an example, phenomena in the strong sense. But they are good have no degrees of freedom to change their in­ candidates for being so. Synergetics, on the other ternal structure. In this respect connectionist nets hand, does not treat any phenomenon as syn­ differ clearly from seeming analogic cases such as chronically emergent, nor does it treat novel struc­ diamonds, which were referred to by Rumelhart tures as candidates for structure emergence. Phe­ and McClelland to explain ‘emergent’ system’s nomena discussed by synergetics are emergent properties (cf. 1986, p. 128). Diamonds are not dy­ only in the weak sense. The same is true for prop­ namical systems that realize only after a certain erties of connectionist nets. As we have just seen, number of reiterated steps the property of being connectionist nets do not instantiate any stronger hard. The diamond’s property of being hard is al­ type of emergence. ways manifest, it does not emerge. To sum up, connectionist nets do not instantiate Acknowledgements any stronger type of emergence. Neither are the First of all I wish to thank Judy McLaughlin who net’s properties synchronically emergent, nor is helped a lot to transform my German English into the formation of a net’s soft structure a case of American English. I am also very grateful to Wolf­ structure emergence. In the weak sense in which gang Buschlinger, Joachim Schummer, Manfred Stockier, Gerhard Vollmer and Henrik Walter for 33 The developing ‘soft structure’ depends on two factors, the net’s hard structure, and the given inputs, their comments on an earlier version of this arti­ that is external influences on it. cle. 656 A. Stephan • Varieties of Emergence in Artificial and Natural Systems

Alexander S. (1920), Space, Time, and Deity. Two Vol­ Horgan T. and Tienson J. (1996), Connectionism and the umes. Macmillan, London. Philosophy of Psychology. M IT Press, Cambridge, Bates E. (1979), The Emergence of Symbols. Academic Ma., London. Press, New York. Kirk R. (1996), How is Consciousness Possible? In: Con­ Bechtel W. (1988), Philosophy of Science. An Overview scious Experience (T. Metzinger, ed.). Schöningh, for Cognitive Science. Lawrence Erlbaum Associates, Paderborn, 391-408. Hillsdale, NJ. Kurthen M. (1996), On the Prospects for a Naturalistic Beckermann A. (1996a), Visual information-processing Theory of Phenomenal Consciousness. In: Conscious and phenomenal consciousness. In: Conscious Experi­ Experience (T. Metzinger, ed.). Schöningh, Pader­ ence (T. Metzinger, ed.). Schöningh, Paderborn, born, 107-122. 409-424. Levine J. (1983), Materialism and Qualia: The Explana­ Beckermann A. (1996b), Eigenschafts-Physikalismus. tory Gap. Pacific Philosophical Quarterly64, 354-361. Zeitschrift für Philosophische Forschung 50, 3 -25. Levine J. (1991), Cool Red. Philosophical Psychology 4, Block N. (1978), Troubles with Functionalism. In: Per­ 27-40. ception and Cognition (C. W. Savage, ed.). Minneapo­ Levine J. (1993), On Leaving Out What It’s Like. In: lis, 261-325. Consciousness (M. Davies and G. W. Humphreys, Block N. (1980), Are Absent Qualia Impossible? The eds.). Blackwell, Oxford, Cambridge, Ma, 121-136. Philosophical Review 89, 257-274. Levine J. (1996), Qualia: Intrinsic, Relational - or Broad C. D. (1919), Mechanical Explanation and its Al­ What? In: Conscious Experience (T. Metzinger, ed.). ternatives. Proceedings of the Aristotelian Society 19, Schöningh, Paderborn, 277-292. 86-124. Lorenz K. (1988), Die Rückseite des Spiegels. Neuaus­ Broad C. D. (1925), The Mind and its Place in Nature. gabe. Piper, München & Zürich. Kegan Paul, Trench, Trubner & Co., London. McLaughlin B. (1992), The Rise and Fall of British Broad C. D. (1933), Examination of McTaggart’s Philos­ Emergentism. In: Emergence or Reduction? Essays ophy, Vol. 1, Reprint. New York (1976). on the Prospects of Nonreductive Physicalism (A. Bunge M. (1977), Emergence and the Mind. Neuro­ Beckermann, H. Flohr, and J. Kim, eds.). de Gruyter, science 2, 501-509. Berlin, New York, 49-93. Churchland P. (1988), Matter and Consciousness. Re­ Rey G. (1996), Towards a Projectivist Account of Con­ vised Edition. MIT Press, Cambridge, Ma. scious Experience. In: Conscious Experience (T. Met­ Clark A. (1989), Microcognition. Philosophy, Cognitive zinger, ed.). Schöningh, Paderborn, 123-142. Science, and Parallel Distributed Processing. MIT Rumelhart D. E. and McClelland J. L. (1986), PDP Mod­ Press, Cambridge Ma., London. els and General Issues in Cognitive Science. In: Paral­ Hacking I. (1975), The Emergence of Probability. Cam­ lel Distributed Processing. Explorations in the Micro- bridge University Press, Cambridge. structure of Cognition. Vol. 1 (D. E. Rumelhart, J. L. Haken H. (1983), Synergetik. Eine Einführung. Zweite McClelland, and the PDP Research Group, eds.). MIT Aufl. Springer-Verlag, Berlin, Heidelberg, New York, Press, Cambridge, Ma., London, 110-146. Tokyo. Rumelhart D. E., Smolensky P., McClelland J. L. and Haken H. (1987), Die Selbstorganisation der Informa­ Hinton G. E. (1986), Schemata and Sequential tion in biologischen Systemen aus der Sicht der Syner­ Thought Processes in PDP Models. In: Parallel Dis­ getik. In: Ordnung aus dem Chaos (B.-O. Küppers, tributed Processing. Explorations in the Microstruc- Hg.). Piper, München, 127-156. ture of Cognition. Vol. 2 (J. L. McClelland, D. E. Ru­ Haken H. (1988), Erfolgsgeheimnisse der Natur. Syner­ melhart, and the PDP Research Group, eds.). MIT getik: Die Lehre vom Zusammenwirken. Ullstein Press, Cambridge, Ma., London, 7 -5 7 . Sachbuch, Frankfurt, Berlin. Stephan A. (1994), Theorien der Emergenz - Metaphy­ Haken H. (1996), Synergetik und Sozialwissenschaften sik oder? Grazer Philosophische Studien48, 105-115. and Replik. Ethik und Sozialwissenschaften7, 587- Stephan A. (1997), Armchair Arguments Against Emer­ 594, 658-675. gentism. Erkenntnis 46, 305-314. Haken H. and Haken-Krell M. (1989), Entstehen von Stephan A. (1998), Emergenz. Von der Unvorhersagbar­ biologischer Information und Ordnung. Wissen­ keit zur Selbstorganisation. Dresden University schaftliche Buchgesellschaft, Darmstadt. Press, Dresden. Hardin C. L. (1987), Qualia and Materialism: Closing Vollmer G. (1988), Evolutionäre Erkenntnistheorie und the Explanatory Gap. Philosophy and Phenomenolog­ Leib-Seele-Problem. In: ders., Was können wir wis­ ical Research 47, 281-298. sen? Band 2. Die Erkenntnis der Natur. Hirzel, Stutt­ Hardin C. L. (1988), Color for Philosophers: Unweaving gart, 66-99. the Rainbow. Hackett. Indianapolis. Vollmer G. (1992). Das Ganze und seine Teile - Ho- Hardin C. L. (1991), Reply to Levine. Philosophical lismus, Emergenz, Erklärung und Reduktion. In: Wis­ Psychology 4, 41-50. senschaftstheorien in der Medizin (W. Deppert, H. Horgan T. (1993), From Supervenience to Super- Kliemt, B. Lohff and J. Schaefer, eds.). de Gruyter, dupervenience: Meeting the Demands of a Material Berlin, New York. 183-223. World. Mind 102, 555-586.