Artificial Thought and Emergent Mind

1 Introduction

The field of Artificial Intelligence, and cognitive sci­ ence in general. has had many great and celebrated achievements over the years, but, at the same time, it has brought about quite a few controversies and heated debates about the adequacy of the computa­ tional metaphor and constructlvistlc methodology to our understanding of the real human mind.

It is not the aim of this paper to profess the author's support to one or another position in the dispute or to fmd one more flaw in any particular scholar's argu­ ment. Rather, in hope of alleviating the tensions be­ tween various existing camps, I intend to propose certain conceptual distinctions which may help us find some common themes among the various claims made by scholars about the nature of thought and mind.

To make this long story short: I believe that a dis­ tinction between two types of mentalistic terms will serve the AI community well, one less resistant to the adjective "artificial" than the other. I have chosen, per­ haps irresponsibly, "thought" as the former term and "mind" as the latter term. My strategy is to include into the generic concept of thought even those proc­ esses which can be externally and objectively de­ scribed, and thus they are acceptable for an intentional constructive procedure.

In this respect it appears useful to consider various scales of magnitude, or levels of analysis, on which a given entity has a meaning. The distribution over scales seems to offer a sharp distinction between complex entitles in nature, like organisms or brains, make a flower. But such an intention alone is not a sufficient condition either: even if my random playing with the cloth and wire, or my waiting for the wind to do it for me, were accompanied with my best con­ scious hope that a flower would sooner or later emerge, I would still hesitate to call the flower artifi­ cial. What would be missing then is the intention in project.

A project not only requires a prior intention, but also an explicit design specification, i.e., an external and objective description of all the relevant properties of the intended final product. External, because con­ struction (unlike, for instance, learning) is an external activity; objective, because anyone else should be able, in principle, to reproduce the project in order to achieve the same outcome.

In terms of our definition of a project and the exter­ nal and objective condition which a project demands, we can ask whether it makes sense to conceive of a project to create artificial thought, that is, to try to make something. perhaps computers, think. My claim is that any project of AI would require an external and objective description of the process of thought. This is a strong requirement, indeed, and if it were satisfied there would not remain, in our age of computers, too much more to do.

3 Thought and Mind

I have promised to propose a conception of thought that will lend validity to the project of AI. to the project of making computers think. Curiously, I achieve this goal by a simple trick: by proposing a sufficiently gen­ eral concept of thought.

Tentatively, let us define thought as the act or proc­ ess dealing with conceptual objects (concepts, thoughts) in an intelligible and meaningful way.

This requires some clarification. First. by a concep­ tual object we mean a representation of any entity, in­ dividuum, category, property or situation, either derived from direct experience in the real world or constructed in the course of past processes of thought. The seeming circularity (viz.. reference to thought) is harmless; it actually allows for a gradual enrichment of the process.

Second, intelligibility means understandability by intelligent observers (for instance, by "us"). That is, the inherent overall logic of the process should be un­ derstandable, although it is not necessary to grasp the details, or the real causes of particular actions.

And third, the word meaningfUl should exclude ran­ domly disordered, chaotic behavior on the one hand, and inert, stagnant or boringly repetitive behavior on the other hand.

These explanations betray the slight of hand of our simple trick: the definition does not mention causal microstructure, nor does it state an explicit require­ ment of a presence of an identifiable individual agent or "executor" of thought, nor does it require subjectiv­ ity or consciousness. Thus. for instance, 'conscious thought', 'artificial thought', and 'collective thought' may be three different (perhaps related) types of realization of thought. Let me illustrate the point by several examples.

(1) Assume, first, JohnS., sitting in the famous Chi­ nese room (Searle, 1980) and manipulating Chinese characters. If he understands Chinese he can carry on a genuine Chinese conversation. Such a real con­ versation involves a real conscious thought.

(2) Now assume he does not understand Chinese. He can still perform a thought process, but it is much less sophisticated and has a different intention, namely, to manipulate the Chinese characters (for him meaningless squiggles) according to the pre­ scribed instructions. Even this thought process is (normally) conscious but with consciousness playing a relatively insignificant role in the process itself (John S. could be easily substituted by a machine.)

3) In the previous situation, John S. and the Chi­ nese room (the instructions included) perform jointly still another thought process. With the help of pre­ scribed instructions for manipulating Chinese charac­ ters (symbols whose meaning is implicitly stored in the instructions) they behave as if they understood Chinese. I would classifY this process as artificial and (most likely) not conscious. (Here again John S. could be substituted by a machine).

(4) Now a somewhat different case. Consider the set of thinkers who have worried themselves over the Chi­ nese Room Problem, who talk, write, and debate about it. Together they realize a collective thought process, not artificial and not conscious (unless we believe in some sort of higher, collective con­ sciousness) 1 ·

Now, everything that was deliberately left aside or not mentioned in our "definition" of thought should be included in the concept of the mind. Relatively vague otherwise, our concept of mind should include, in particular, all important properties ascribed to it by philosophers: intentionality, rationality, free will, men­ tal causation, subjectivity, and, above all. con­ sciousness.

4 The Difference between Conscious and Artificial Thought

In recent years there has been a noteworthy shift of interest in cognitive sciences and analytical philoso­ phy towards the issue of consciousness. Although in a recent monograph on neuropsychology ((Kolb and Whishaw, 1990)) the term "consciousness" does not appear in the index, an increasing number of scien­ tific and philosophical books are concerned with this topic (e.g. [Dennett, 1991, Marcel and Bisiach, 1988, McGinn, 1990, Searle 19921) . Whether an author grants a privileged ontological status to consciousness or whether he intends to eliminate reference to con­ sciousness altogether, it is quite clear that AI cannot completely ignore this recalcitrant concept.

1 I concede a certain degree of vagueness in using the adjectives 'conscious', 'artificial', 'collective', etc., to describe a thought process: it is not dear, for example, whether the adjectives are to be understood as different brands of process or as different brands of realization of thought processes. Both interpretations are sound provided we do not take the latter alternative as meaning that aU thought processes admit aU brands of realization.

Consciousness is a part of our inner subjective ex­ perience and as such it is not, in principle, accessible to the outside observer. Therefore those functions of the mind, like conscious thought, that are intimately and inseparably connected to consciousness cannot be fully described and presented in an external and objective language without sacrificing some essential component. As we have argued above, without an ex­ plicit design specification there is no project, and without a project there is no sense in talking about something being artificial.

However, if we restrict ourselves to those thought processes which can be objectively described (recall our sufficiently general concept of thought}, there should be no objection to anyone venturing into the project of realizing them artificially. In fact, when the description is sufficiently precise, complete and un­ ambiguous, as it is in the case of formal algorithms, there are well-known standard tools for executing such a project (namely, programming systems and computers) . There are limitations due to undecidabil­ ity and complexity of certain tasks, but these limita­ tions are not what we here consider.

I have argued that, in principle, conscious thought cannot be, without a substantial loss, converted into artificial thought (which claim may be taken as a re­ jection of a thesis analogous to the strong AI thesis [Searle, 19801) . What about the converse: is artificial thought convertible to conscious thought? In a certain sense and in certain cases the answer is, obviously, yes. The sense is the following: processes of thought which admit artificial instantiations admit, in princi­ ple, also instantiations with conscious control. Recall case (2) above (manipulating meaningless squiggles in accord with a prescribed set of instructions) .

One may argue as follows . If there are instances of artificial thought processes realizable as conscious thought processes, does it not follow, contrary to the above claim, that at least some conscious processes, namely those just mentioned, can be realized artificially?

This argument is valid, but not philosophically in­ teresting. Our claim does not deny the possibility of extracting certain parts or certain components of con­ scious mental activity and converting it into, say, a computer program, with the behavior or functional structure equivalent to the behavior or functional structure of that mental activity. There are two possi­ bilities. Either the presence of consciousness is not essential for the mental activity in question (as in ex­ ample (2) above) in which case the equivalence, being trivial, does not say anything at all about artificial re­ alizability of consciousness or consciousness is essen­ tial for the mental activity (which may be the case of example (I)) in which case extracting the programma­ ble component will yield an entirely different thought process.

The issue of inner (first-person) conscious mental life as opposed to an outer (third-person) point of view has been thoroughly discussed by several authors, most recently by Searle [1992) . My aim in the rest of this article is to concentrate on another aspect of mental processes or on their (natural or artificial) re­ alizations. This other aspect is related to the variety of scales of magnitude relevant, or essential, to these processes. For this purpose (and for those who prefer the visual metaphor) let me first introduce a new imaginary "dimension" corresponding to varying scales of space and time.

5 Scale Dimensions in Nature

Perhaps we can best start with an example: con­ sider ordinary, geographical, 2-D maps of the same region but of different scales, superimposed one on top of another, with larger-scale maps on top of smaller-scale maps.

We can imagine scales (expressed, say, by real numbers) without limits both "downward", to the small, microscopic and submicroscopic scales, as well as "upward", to the large, astronomical scales and be­ yond. Assuming, moreover, a dense sequence of scales we obtain a continuum represented by a spe­ cial coordinate axis. Let us call it the scale axis and the corresponding dimension the scale dimension.

Somewhere in the "middle" of the scale axis exists our habitat, the scale-local world of human magni­ tudes, our "scale-here". Unlike the ordinary spatial "here", the natural "scale-here" has the same position on the scale axis for all people (and perhaps for ani­ mals of about our size). Both spatial "here" and scale "here" are smeared and cannot be exactly localized to a point.

In the same way as we introduced the scale dimen­ sion for space, we can introduce the scale dimension for time representing various magnitudes or "speeds" of time. Again there exists a natural "scale-here" for time scales, corresponding to the rhythms of human life and thus, under normal circumstances, common to all people. In particular, our thought processes span the interval roughly between milliseconds and hours. 2

6 Thin~s, Events, and Artifacts: Their Distribution over Scales

The scale-local objects (entitles of size/or duration basically accessible to humans) are a special case of objects that I shall call scale-thin, i.e. objects "occur­ ing" in a limited range of spatial or temporal scales. We are used to organizing real-world entities accord­ ing to their relatively specific position on the spatial and/or temporal scale axis and according to the mu­ tual relation of their positions into several categories. The most common of our organizing concepts are things and events.

2 In fact, the 100 millisecond scale has been proposed as the demarcation line between the classical symbolic-algorithmic paradigm in Artificial Intelligence (above 100 ms) and the subsymbolic-connectionistic approach (below 100 ms) (Hofstadter, 1982( . (My opinion is that what plays really the main role is the interaction between distant levels; cf. Section 13) .

By things, we intuitively and typically mean those entities wWch are separable, with identifiable shape and size, and which persist over time. Events, on the other hand, typically have a relatively short duration and are composed of the interaction of several, per­ haps many things (of various sizes). However, in the world of all scales there is no essential difference: things are just long-lasting events and events are just short-lived things (where "long" and "short" are rela­ tive with respect to our temporal scale perspective). Many other entities (vortices, clouds, flames, rivers, networks, sounds, bubbles, winds, ceremonies, meet­ ings, wars) have an intermediate character.

It is customary, when describing concrete phenom­ ena in the world, to treat their spatial structures (shape and internal composition) and their temporal structures (internal dynamics and behavior) sepa­ rately. TWs separation, together with our scale-thin language and our scale-local perspective, helps con­ ceptually, but. at the same time considerably narrows our perception of reality (in a way reminiscent of the Baconian "idola tribus" -idols of the tribe) .

For illustrative purposes, let us restrict ourselves to the case of spatial scale dimension. The reader is in­ vited to make Ws own generalization to the temporal scale dimension.

Various, typical objects that we observe in nature can be categorized according to their "distribution" over scales. Let me explain what, here, distribution is. First, for each such object we identity various scales relevant to its spatial features, like, for instance, sizes of its components and relative distances between in­ teracting components. Obviously, the scales cannot be identified sharply so that a continuous function over the scale axis with salient peaks or elevations around certain scales would result. Let us call this function the relevance function or the (spatial) scale spectrum of the object.

Now, according to the distribution of peaks in the scale spectrum we can, in a first approximation, iden­ tity four basic (even if not sharply separated) catego­ ries of objects. First, there are single bodies (stars, stones, dust particles) with only one salient peak in . the scale spectrum (if we ignore the lower, molecular and atomic structure). Second, there are clusters (gal­ axies, clouds, heaps) with two or more sparse peaks. Third, there are complex systems or organisms, typi­ fied by a large number of relatively dense but still dis­ tinguishable peaks spread over a certain interval of scales (I shall return to this category later). Fourth, there are scale-homogeneous structures, i.e. objects with continuous spectra (in a certain interval) of rele­ vant scales. These last structures are relatively rare in nature; examples are fractal shapes and also t1uids near their critical points.

Up to now, we have classified natural objects. To ex­ tend the concept of scale spectra to human artifacts (tools, engines, computers. houses, cities). we have to reinterpret the notion of (scale) relevance. Obviously, what matters here is much less the relevance (of vari­ ous scales) for those objects themselves than the rele­ vance for, or from the point of view of, those who conceive, construct and use them. There are, often only a few, relevant (in this sense) scales for such ob­ jects, occasionally separated by large gaps.

Consider, for instance, the computer as a physical object. The most important spatial scale is the local scale of users (it is the scale on which the computer is designed and meaningful). Then there are several well-known relevant scales (of hardware architecture, processing units, logic circuits, semiconductors. down to the scale of quantum phenomena). each associated with a special engineering discipline and with a spe­ cial design and specification language. 3

Whatever the scale spectrum in the designers' per­ spective is, there is always one and only one relevant scale (peak in the spectrum) for most artificial objects, including computers. It is the local scale "here" of us. the users, where the meaning of the object is located.

Basically the same holds for any natural or artificial language and, in general, for any symbolic represen­ tation (provided humans can read the language and can interpret the symbolism). Symbols, symbolic pat­ terns, and syntactic objects are suitable for conveying meaning only within a narrow range of scales, beyond which narrow range they are incomprehensible. We should bear this in mind when discussing differences between artificial and natural thought.

7 Levels and Their Hierarchies

In the case of complex objects, there is a close rela­ tionship between their distribution over scales and a hierarchy of their structural, functional, or descrip­ tional levels. In many situations in which we find it convenient to talk about various levels, we can also distinguish corresponding scales or ranges of scales, spatial as well as temporal. Accordingly, Salthe (1991) uses the generic term scalar hierarchy whenever the levels are characterized by different scale, as opposed to the specijication hierarchy with levels based on de­ gree of specification or generality. While a scalar hier­ archy may be based on the part-whole distinction, a specification hierarchy is typically based on the token-type distinction.

If a certain structure or object has distinguishable salient peaks or elevations in its scale spectrum, it is natural for us to associate with such peaks and eleva­ tions appropriate levels of a scalar hierarchy. Moreo­ ver, because our languages are not suitable for large scale span, we develop specific descriptional lan­ guages for particular levels.

3 In actual computing, the temporal scale spectrum may be more important.

To illustrate this point, let us consider a cluster, for instance a cloud of dust. Such an object may be stud­ ied on the global level (for example, in terms of its overall shape and size and in terms of its global prop­ erties, like opacity) or on the level of its components (the dust particles, their individual properties, such as distribution, density, and interaction) .

One may, of course, ask about the actual reality of levels as such. Do they exist independently of our analysis and description of objects and events? I be­ lieve that sometimes they do, at least partly, but that sometimes they are, again partly, our mental con­ structs. The scale-thin world of our ordinary percep­ tions and thoughts makes it difficult for us to grasp more than a certain limited range of scales at once. For this reason, and other reasons, we tend to decom­ pose objects of our concern into structural levels and events (and processes) into functional levels. Obvious differences of individual levels yield different descrip­ tions, different languages and, eventually, different disciplines. If all is done properly, the decomposition may match something which approximates the real differentiation of nature.

There is one problem which may be crucial for our understanding of complex systems: whether and how can distinct (possibly distant) levels of a system di­ rectly interact. In the folloWing section, I shall make a few comments on this problem.

8 Interaction across Levels

The scalar hierarchy is commonly treated in sys­ tems science with the tacit assumption that "constitu­ tive dynamics at different scalar levels are largely screened off from each other (non-transitivity of ef­ fects across levels)" ([Salthe, 1991). p.252). and with the resulting belief that "three contiguous levels should be sufficient to understand most of the behav­ ior of any real system" (p.253). This assumption and belief are, in fact, included already in the term 'hierar­ chy' (in contrast to 'heterarchy'). According to system scientists, occasional influences from distant levels are generally considered as "perturbing fluctuations" that need not be included in a dynamical description of the system in question. I think this view is inher­ ently connected with the explanatory role of causality in science.

Scientists base their understanding of the processes of nature mostly on causal interaction; it is the princi­ pal explanatory apparatus. Yet the scale-thin concep­ tual field restricts our experience and intuition more or less to an in.fra-level "left-right" causation: we refer to an earlier event to explain a later event. Therefore, the ease of treating the inter-level, micro-macro cau­ sation as unproblematic, indeed the only acceptable, type of interaction between levels (Searle, 1992) is somewhat surprising.

Of course, there are cases in which it is quite legiti­ mate to employ causal explanations between levels, usually from one level to an adjacent higher level. For instance: the properties of molecules cause the growth of a crystal to a specific global shape, the dis­ organized movement of molecules causes Brownian motion of larger particles, etc. But one has to be care­ ful about generalizing this way of thinking to every­ thing. For example, to say that "mental phenomena are caused by neurophysiological processes in the brain and are themselves features of the brain" (Searle, 1992) suggest too liberal an interpretation of the term "are caused", even if we agree, for the sake of understanding, on a strong assumption, namely, that there is a natural hierarchy of levels above the neuro­ physiological one with some higher level attributable to the mental phenomena.

One of the few theories in science that deal with inter-level interactions is Gibbs-Boltzmann's ·statisti­ cal physics (thermodynamics and the study of collec­ tive phenomena) . It succeeds rather by eliminating the lower (microscopic) level from the macroscopic laws through decomposition of the phase space to what is considered macroscopically relevant subsets and by introducing new concepts, such as entropy (which is, of course, a wonderful trick). As an indirect result we obtain, for example, an "explanation" of macroscopic asymmetry of (physical) time. Can we, however, really say that the time asymmetry is caused by behavior of particles?

There is a well-known technique of renormallzation (cf.,e.g., (Wilson, 1979)) which deals with problems that have multiple scales of length. It is particularly suitable for phenomena near critical points and has applications in various branches of physics. But it is not a descriptive theory of nature and, therefore, has little ontological relevance.

Another relevant area is the study of deterministic chaos. Here people become more and more used to situations in which extremely tiny fluctuations are al­ most instantaneously amplified to a macroscopic scale. What seems to be a purely random event on one level appears to be deterministically lawful behav­ ior on some lower level. This is (mathematically) a re­ current situation. We can, therefore, take the deterministic description as something permanently hidden behind the scale horizon, albeit always avail­ able formally for explanatory purposes.

Particularly interesting, and surprisingly much ne­ glected, is the compelling question of the asymmetry of interactions with respect to the scale axis. Why is the arrow of putative causality, or of other natural in­ fluences, usually assumed to have a direction from lower levels to upper levels? Is it the heritage of the clock-work mechanistic conception of the nature or one of its inherent asymmetries? Ut should be noted that not everybody excludes the idea of "downward" causation (Popper and Eccles, 1977)).

In this respect, we can find some inspiration in the notion of information (as an ontological category; cf. Bohm's concept of active information (Bohm, 1990)) and, even perhaps, in the notion of the mind (cf Sec­ tion 10).

9 Organisms and Brains: Multilevel Interactional Structures

We have used (in Sec. 6) the term organism for natural objects with many relevant scales in a large interval; alternatively, we can talk about a relatively dense Werarchy of structural and/or functional levels. Moreover, in living organisms, there is a strong mu­ tual interaction between particular levels. Both den­ sity and interaction are crucial features here. In the computer, for example, there are several prominent levels and there is (some) interaction; but there is no density. In the fractal, there may be density but there is no (physical) interaction.

Let us state our key question: is it not the very exis­ tence of such a hierarchy of interacting levels that makes living organisms (and brains) different from computers, clouds, and fractals? Is it not perhaps just tWs property that makes it so difficult to submit their behavior to a mechanistic explanation?

I doubt that the function, meaning, and actual be­ ing of living organisms can be associated with some particular level (or scale). Each of these aspects takes place equally well on the level of molecules as on the levels of cells, organs, individuals, social groups or ecosystems. It comes about at many different scales of space as well as time.

Consider the following opinion of the physicist Bar­ row ([1991). p.97): "There exists a form of WerarcWcal structure in Nature wWch permits us to understand the way in wWch aggregates of matter behave without the need to know the ultimate microstructure of mat­ ter down to the tiniest dimensions." I am afraid that in order to accept tWs statement we would have first to restrict considerably the meaning of 'understand' and 'behave'. Otherwise, if we wanted to apply Barrow's statement to so complex an organism as the brain we would run into problems of identltYing the depth be­ low wWch further levels cease to be significant.

In fact. there is a growing number of works suggest­ ing a certain relevance to the brain activity of all levels down to the quantum scale. For instance Beck and Eccles (1992) propose a mechanism whereby the probability of exocytosis of synaptic transmitters is, by means of a quantum-mechanical tunneling proc­ ess, influenced by mental events. Incidentally, if the lower level quantum effects happen to have a certain important role in the conscious mind, tWs may yield an argument for the unique scale location of the over­ all size of the brain.

In general, our considerations do not favor reduc­ tionism in biology and psychology. Indeed, to under­ stand life and the mind does not mean to reduce it to some basic components, but rather to appreciate vari­ ous influences, bounds, and interactions between all structural and functional levels. close as well as re­ mote with respect to the the scale dimension.

1 0 The Location of the Mind

Perhaps a similar expansion of the scope of view might help us to understand better the nature of the mind and consciousness, or at least to avoid certain persistent fallacies. One such fallacy is the belief, held by some cognitive scientists, that the mind is nothing but a collection of processes occuring on a certain privileged level above the level of neurophysiological processes in the brain. This fallacy is probably caused by at least two misconceptions.

The first miconception may be the oveijudgement of the computer metaphor. If someone maintains that the brain is a (sort of) computer equipped with pro­ grams and that the mind is a collection of such pro­ grams (or processes controlled by them). then he is immediately drawn to the language metaphor (that the programs may be "written" in some language). And, since languages happen to be scale-thin (cf. Sec. 6). it is natural to take mind to be scale-thin too, i.e. restricted to a specific level, namely the same as the level of (human) communication.

The second misconception may be the confusion of the intentional content of mental states with those states themselves. TWs content, i.e. topics of our be­ liefs, objects of our perceptions, images of our fantasy, and goals of our plans, are primarily tWngs of ordi­ nary size - "things which a baby can handle and (pref- · erably) put into his mouth" [Popper and Eccles, 1977). Mental states are believed to be neurophysiological states; because it is absurd to think that neurons can handle the same tWngs as babies, it is taken for granted that the mental level is sufficiently above the neuronal level. A similar attitude is held even by some of those philosophers of the mind who are bravely op­ posing eliminative and reductionistic materialism. For instance, Searle [1992) claims that "conscious states are simply higher-level features of the brain" (p. 14).

But is it proper, in the context of mental phenom­ ena, to talk about "levels" at all? Even if we did notre­ strict ourselves to the scalar hierarchy (cf. Sec. 7). it would be a mistake, I believe, to treat mentalistic terms as sometWng that should be, by their use, carefully confined to a certain "level'', "domain" or "subject area".

Hofstadter )1979) was one of the first authors who discussed the connection between mental phenomena and a Werarchy of levels. He rightly pointed to the im­ portance of inter-level interaction on distance (includ­ ing loops) but he wrongly embedded mental levels into the scalar Werarchy of functional levels of the brain and he confused the latter with the specification Wer­ archy of indirect reference (p. 709). Moreover, his functional "holism" is not quite consistent with his own warning against the use of nondifferentiated lan­ guage for different levels of description, to wWch use he ascribes the many confusions in psychology.

It may well be the other way round. Perhaps all these confusions come from precisely such strict frag­ mentation of concepts into levels. If, for instance, con­ sciousness is to be understood as a property of a body, it certainly should not concern certain some or another single level but the whole living organism "penetrating" through many mutually interacting and cooperating levels, including, perhaps, even the level of quantum physics.

11 The Emergence of Emergence

I suggested that we can limit the use of the term "artificial" to those cases in which there is a clear dis­ tinction between the project and the designer, and where an external design specification is possible and both a prior intention and an intention in project are present (Sec. 2). This applies very well to the tradi­ tional logical-symbolical-computational AI (nicknamed GOFAI - Good Old-Fashioned AI) .

Recent advances in connectionist architectures sug­ gest an alternative strategy which blurs the above dis­ tinction and in which there is no presumption of an a priori external description of the structure of the task [Havel, 1992). One may construct a complex dynami­ cal system in the form. typically, of a large number of mutually communicating units, and then patiently wait for an appearance of some complex emergent phenomena that might support certain processes of thought. Of course. this type of thought would be somewhat alien and incomprehensible to us, and it would not fall. in our terms, into the category of the natural nor of the artificial. Indeed, it would not be (entirely) natural because the requisite dynamical sys­ tem (connectionistic or other) is artificial and it would not be (entirely) artificial because there is no intention in project (cf. Sec. 2) .

This alternative strategy brings us to the idea of emergent mind. The concept of emergence, especially in the context of mental phenomena. deserves a sepa­ rate study. There are actually three meanings of this word, not always easy to distinguish. The traditional meaning in evolutionary theory (G. H. Lewes in mid-19th century, C. Lloyd Morgan in early 20th cen­ tury) emphasizes the temporal aspect: the rise of a system that cannot be predicted or explained from an­ tecedent conditions, e.g. the emergence of life, of man. etc.

In its second meaning, the word emergence has re­ cently been increasingly used for phenomena that ap­ pear to be natural on a certain level of analysis, but somewhat resist (though not completely exclude) their reduction to an appropriate lower level. Typical exam­ ples of this are collective or mass properties: the li­ quidity of water (a molecule of H 2 0 is not liquid) or democracy in society (one person cannot form a democratic system). Often this type of emergence is only apparent when caused by the intractable com­ plexity of the lower-level structure.

The third meaning is inspired by the second and is used often as an argument against reductionism. A property (specified by a certain theory T 1 ) is said to be (properly) emergent if it has real instances. if it is co­ occurent with some property recognized in a reducing theory T 2 , but which cannot be reduced to any prop­ erty definable in T 2 (cf. [Churchland, 1986). p. 324). Property dualism is characterized by the conviction that "even if the mind is the brain, the qualities of subjective experience are nevertheless emergent with respect to the brain and its properties" (ibid, p. 323).

Let us consider the thesis that thought occurs as an emergent phenomenon on some higher level of a hierarchical system, with low levels being purely mechanistic. This thesis would help materialistic mo­ nism avoid the concept of the soul as an unknown ghostly substance, which is regarded as flowing or fly­ ing in another world. However, if the motivation for this avoidance is the mere resistance to accepting an unknown or unknowable entity, then not even the concept of (proper) emergence will help, at least until something more is known about the matter. For in­ stance, as I pointed out earlier, there is little or no un­ derstanding of interaction between different levels.

On the other hand, the emergentist thesis cannot be easily refuted and we cari tentatively accept it for the sake of discussing the chances of connectionist AI.

12 Collective phenomena

Global behavior of a connectionist system can be viewed as a specific case of a much more general con­ cept of a collective phenomenon. Collective phenom­ ena typically occur in large collections of individual units, the behavior of each unit being partly depend­ ent on the behavior of some other "neighboring" units. The remaining, autonomous part of this behavior may be based on a random and/or rational individual decision.

The connectionist model is an example of a one-level system (all units at the same level of description) . In contrast. the concept of a hierarchical collective sys­ tem incorporates the idea of an iterated division of larger tasks to smaller subtasks. This idea is natural for the top-down AI strategy ([Minsky. 1985)), but at the same time it may support emergent collective phenomena.

According to the weight of the autonomous part of behavior of individual units, we can distinguish two opposite modes of global behavior or, using the lan­ guage of statistical physics, two phases:

(1) the rigid, bureaucratic system of primitive obedi­ ent agents (low autonomy). and

(2) the chaotic, anarchic system where everyone does whatever he likes.

(Cf. [Dennett, 1991). Chapter 9) . Various disciplines. from physics to the social sciences, offer many exam­ ples of mixed or intermediate cases. For instance, we may have a system of initiative agents, competing for recognition. each with his own idea, while at the same time all are attentive to the ideas of their colleagues. In a rigid system, a good, new idea can occur only with great diftlculty; in the chaotic system, it is always lost; but, in the intermediate system, it may propagate easily through large areas of the network. In physical systems we encounter similar situations near phase transitions [Little, 1990). A great deal of attention has been recently paid to the intermediate case, called the edge of chaos, featuring interesting properties, among them a sort of evolutionary stability.

Collective systems with highly parallel activity are interesting alternatives to classical. serial­ computational models. Dennett (1991) uses the idea of multiplicity of competing agents (called homunculi) in his theory of consciousness. Collective systems have extremely large combinatorial complexity (the number of global states grows exponentially with the number of units). Such a complexity is not, however, a disadvantage. It yields redundancy and redundancy supports self-organization and self-improvement.

Another interesting collective phenomenon is the emergence of islands of cooperative behavior or altru­ ism in a large set of egoistic individuals )Axelrod, 1984).

13 Towards a Scale Holism

Let me conclude with a few more or less speculative remarks.

I carinot resist a suspecting that the term "emer­ gence" is currently used mostly to conceal our igno­ rance of links between entitles on a certain, relatively well-uderstood level. and entitles observed or as­ sumed on the next higher level (for reasons to be seen soon let us call it the first-order emergence). There is nothing wrong with such using the term provided it is just a first step in future reserach, at least in two directions.

One direction could be to study more thoroughly the nature of those links. We may ask, for instance, to what extent is the first-order emergence (from lower to upper level) reducible to phenomenona studied in nonlinear dynamical systems theory, like, let us say, greatly amplified fluctuations.

A second direction could be, I believe, of more criti­ cal importance, especially if we want to study such evasive entitles as mental phenomena. I have argued above (in Sec. 9) that certain complex systems (organ­ isms, brains) should not be understood as scale-thin objects, but. rather, as structures penetrating through many scales and featuring interactions through many levels. This argument suggests good reasons for introducing a new type of emergence that I shall call the second-order emergence. An entity is second-order emergent if it is not associated with some particular level but arises from global interac­ tion of many levels of some scale-extended complex system.

My thesis is that if the mind admits of physicalistic or naturalistic understanding at all, then such under­ standing should be in terms of the second-order emergence rather than the first-order emergence. This applies also to its essential components, conscious­ ness for example. 4

One particular example might illustrate the point. One of the most important components of the mind is memory. This, originally mentalistic, concept is now more and more used as a feature of a physical system - either the brain or the computer. In fact, the knowledge-memory distinction is an interesting varia­ tion on the mind-body distinction. If knowledge were emergent on a certain specific level of the scalar hier­ archy, the next lower level would play the role of a substrate for syntax: it would house meaningless structures with combinatorial features allowing suffi­ cient differentiation and assignment of atomic units of meaning. Not only this assignment but even the com­ binatorial features can be specified only by means of an external activity (of a designer or observer) which cannot be an outcome of spontaneous emergence. (In this respect cf. )Searle 1992).)

· I am not concerned here with the question of subjective nature of the mind.

Therefore a truly emergent knowledge has to have an extended presence over many levels or over many scales with smooth downward degradation of a local semantic content. This can take place in a rather ru­ dimentary form already in distributed connectlonistlc systems, where very low-level objects (units and con­ nections) are bearers of a minimal semantic content, undetectable but "ampliflable", typically through the cooperation of many objects [Havel, 1990).

Some researchers hope that by incorporating the connectlonistlc or cooperative approach into AI, one may achieve substantial progress in the general project of the artificial mind. Perhaps the introduction of these approaches can be considered as progress. not, however. towards anything artlflcial, and proba­ bly not significantly towards what is essential about mind.

As I have tried to clarifY, the artificial involves both intention and project. and the project involves both specification and solution. Now, overcoming the speci­ fication issue (lack of descriptive means) by construct­ ing only a lower-level substrate (for instance an artlflcial neural net) and, then, by letting emergence work is a great idea. However, this strategy tends to eliminate the designer's intentional component. Moreover, such "hybrid" methodology works only in the case of first-order emergence. If it turns out that the second-order (scale-holistic) emergence is a neces­ sary component of the mind, the hybrid methodology will fail, unless, of course, we discover the means for "creating" scale-extended objects - real organisms.