If " ·r 1 From S. F. Frowen (ed.), Hayek: Economist and Social Philosopher: A Critical Retrospect, London: Macmillan, 1997, 9-29. 2 The Connectionist Mind: A Study of Hayekian Psychology 1 Barry Smith INTR ODUCTION I shall begi n with a disc ussio n of recent work in cognitive science , and it may be useful to note that the title of this chap ter might equally well have been something like 'Artificial Intelligence and the Free Mar ket Order' . Readers might also care to note that I am, as far as the achie veme nts and goa ls of research in artificial intelligence are concern ed, some thing of a sceptic . My appeal to cognitive scien ce in what follows is designed to serve clarifica tory ends, and to raise new questions of a sort that will become clea r as the chapter progre sses . Art ificial intelligenc e (AI) researc h has two goals: that of sim ulating hum an intelligence via the construction of machines whose operations will in cru cial respects be analogous to intelligent human performances.i and that of contributing thereby to our theoretical understanding of what human intelligence is. For the practitioners of Al the human mind is itself a sort of computer, and the structures and functions characteristic of the mind are in consequence seen as being capable of being realised in a range of different sorts of materi al, both organic and inorganic. AI research has for some time been do minated by two' competing meth odologies, resting on two distinct paradi gms as to what intelligence might be and what sort of computer might best be employed to simulate or reinstantiate it. On the one hand is the olde r and more orthodo x 's ymbolic' or 'symbol-process ing ' paradigm, which sees intelligence as a matter of the sequential ma nipulation of meaningful units (terms, co ncepts, ideas) of roughly the sort with which we are fami liar in reasoned intro spection . The upholders of this paradigm seek to construct artificially intelligent systems out of entities that are symbols (have both semantic and syntactic properties), operated upon in ways that mod el the rational processes studied by logic. On the other hand is the more modem, and in some respects more challengi ng, 'c onnectionist' or 'sub-symbolic' paradi gm. This sees intelli9 II The Connectionist Mind Barry Smith gence as a matt er of the processing of units much more finely grained in ch ara cter than tho se with wh ich we norm all y suppose ourselves to be familiar in co nsc ious experience. Such processing is to be conce ived by analogy not with proce sses of reasoning as commo nly understood, but with the massively parallel processing of electrica l impulses by the billions of nerve s distri buted thro ugh the human brain, nerve s bound togethe r in net works that are subject to a con stant and subtle forming and reform ing of co nnec tions. In the corresponding simu lations sub-symbolic entities particip ate in numerical - not symb olic - com putatio n, and internerve connect ions are modell ed by applyi ng varia ble weightings to the numerical values (for example signal intensities) that pass through the system from node to node. The mod els here are derived from neurosc ie nce, but also from statis tica l thermo dynamics and other disci plines dealing with processes of an holistic sort. embrace what we might call the Helm holtz hypothe sis, to the effect that the intuitive pro cessing of the expert is merely a case of fast and unconscious proce ssi ng of the lingui sticall y formali sable sor t. Nat ive spea kers are see n from this perspective as unconsc iously interpreting ling uistic rules that coul d in principle be given explicit formulation in grammatica l theories . Systems construc ted on the basis of the Helmho ltz hypothesis ha ve, how eve r, proved too brittle and too inflexi ble to model hum an expertise, and it has turn ed out that the task of establishing the needed artic ulatio ns o f expert knowl ed ge in term s of con text-free rules (for examp le for the dom ain of common sense) faces intractable diffic ulties of a hitherto unforeseen sort. As Smolensky would ha ve it, the Helmholtz hypo thes is ' has contributed essentially no insight into how knowle dge is represented in the brain' (Smolensky, 1987, p. 7). Here, therefo re, we shall conce ntrate primarily on the work of the connectionists. PROPOSITIONAL VERSUS NON -PROPOSITIONAL KNOWLEDGE A CONNECTI ON IST MAC HINE For the defend ers of the symbolic paradigm it is language, built up out of discrete and repeatable units, that provides the domin ant theoretical model of cognition, and in his essay 'On the Proper Treatment of Connectionism ' of 1987, still the definitive formul ation of the connectionist paradigm, Paul Smolensky admi ts that the lingu istic formalisatio n of knowledge brings with it important advantages . It make s the knowledge in question accessible to different peopl e at different times and places , and it makes it possible for us reliabl y to check whether co nclusio ns have been validly reached. It also helps to guarantee learn ab ilit y: the beginner lea rns by following simple co ntext-free rules that can be understood with little or no experie nce of the relevant dom ain . Scie nce , acco rdingly, seeks in eve ry case the sort of sys tema tic clarity that linguistic formula tion provides. Connectionists, however , are interest ed neithe r in the rule-bound, deliberative efforts of the novi ce, nor in the sort of static kno wledge that comes packaged in the form of scientific theories. Rather they are interested in the spontaneo us, intuit ive, tacit , dy namic, practical knowledge of the expert , and they would insist thereby that there are some spheres in which we might all claim to possess huge amounts of knowl edge of this sort. Thus we are all experts, for example, in the business of everyday motion and perception , and this is also true of common sense and our native langu age. Cert ainly the de fenders of the sym bolic paradigm can dev elop views of their own with regard to intuitive knowl edge. Thus, for exa mple, they can It is usef ul to describe very briefly what a simpl e connectionist machi ne might loo k like (Figure 2.1). The machine in question is a sim ple classificatory engine. Imagine a series of dis tinct numerical inputs (signa ls), derived, for example , from a battery of photosensiti ve devices trained in succession upon objects that it is the mach ine' s job to recogn ise (classify). These signals, with give n initial signal strengths , become transfo rmed in their passage through the system according to the 'w eig hts' of the links between successive nodes, 10 - - > -----1 - - >-----1 Inputs Output - - >- ----1 - - >-----1 Figur e 2. 1 Simple connection ist mach ine The Connectionist Mind Barry Smith weig hts tha t may be either positive or negative. The values passi ng through the machi ne are finall y integrated into a si ngle output, wh ich is then associ ated with some value from a predet ermi ned rang e of classificatory co ncepts (say: cat, rabb it, elephant and so on). The behavio ur of the classical ('symbolic' ) machine is fixed in ad vance by its program, so that the react io n of the machine to all admissible inp uts is predic table from the sta rt. The idea behind the connec tionist machine, in co ntrast, is that it need s to be ' trained up ' by means of a large numb er of initial trials, the res ults of which are co ntro lled (in the simplest case ) by a human operator who adjusts the rel evan t wei ghtings by han d - and to some degree at ra ndom - in orde r to brin g abo ut impro vem en ts in the classificatory success that is achieved at each successive stage . If, on the basis o f a given distributio n of we ights, the classificatory output is inco rrec t, alg or ithm s can be used to adjust the weights of node linkages in such a way that the machi ne ach ieves a greater success rate in future trails. It is po ssi ble to prove that such adju stment will eve ntually, und er ce rtain rea sonable sorts of initial condition s, reach one of a perhaps infinite range of stable states and that the machine in question will thereafter be able correc tly to classify new o bj ec ts from the given se t without the need of further human intervention. The idea o f lingu istic or proposi tio nal kno wledge (o f kn owing that ), whic h defe nders o f the sym bolic paradigm take as their starting po int, go es hand in hand with the idea of know led ge as a memory sto re, with which is asso c iated a set of fixed , explicit ru les for cog nitive processing. T he designers of connec tio nist sys tems , in co ntras t, set out to model the fact that knowing how is an evolving ca pacity of the know ing subject and is such that the content of knowledge and the proc ess of gaining kno wledge are not clearly separable from each other. There is no sto reho use of memories from the connectionists' poi nt of view ; rather the connec tionist sys tem ' re mem bers ' on ly in the se nse that its proce ssing pattern s are subjec t to cha nge, being con stantly and cumulativel y affected by what has been experienced in the way of pro cessing in the past. Th e ' knowledge ' in suc h a sys tem is greater to a degree tha n under the sym bolic parad igm , suc h as to influence the course of processing. Hence fro m the co nnec tionist perspective it is necessary to reje ct the opposi tion betw een competence as a matter of knowing expl ic it rules and pe rformance as a matter of the application of suc h rules. In the world o f standard symbolic AI (as here somewhat simp listically conce ived), ther e is an initial plan (the prog ram ) co ntro lling all the successiv e stages of proce ssin g in way s wor ked out in advance . Each of the se successive stages is then intelligible on the sy mbo lic level in virtue of the fact tha t each unit (say 'cat' or 'h orse ' ) ca n be see n as enj oying an intelli gible mean ing o f its own that rem ains fixed through time. In the p ara~ l e l world of the con nectioni st machine , in co ntrast, knowledge and meanm g co me to be di stributed amo ng the nod es of the machine , so tha t, as Sm olensky po ints o ut, what correspond to the concept s of wh ich we are con sciou s are here 'complex pattern s of acti vity distributed ove r ma ny units . Each un it pa rtic ipa tes in man y such pattern s' (Smolensky, 1987, p. 8). Th e interact ion s be tween un its are , it is. true , ent ire ly. simple in nature: they are essen tia lly a matter of num encal val ues be ing passed throueh the system and subjec ted to simple numerical adjust men ts at suecessive sta ges . It is our kno wledge of certain numer ica l pro perties of the se interactio ns th at allows us, by adjusting we ights in succession, stu mblingly to bring the sys tem to one or other stabl e state. But there is no way in which we can deri ve laws or principles that would make these process es inte lliz ible at the le vel o f concepts and meanings .:' Considered from this "' per spective the ind ividual nodes of the machine have no absolute values at, all; their values are entirely a func tion of their po sition in the sys tem of rela tions de termined across the whole network. Work in sym bo lic A I (again fro m ou r somewhat simpli fied perspective) has been marked by a tendency to produce simulat ions of intellige nce that wou ld apply only to rigi d ly well-organi sed wo rlds with ~ rel a.t i ~el y small numbe r of inde pende nt and we ll-behaved dim ensions of vanation . Such si mula tions have in add ition pro ved ' fragile ' in the sense that they have rev ea led them sel ves to be subject to almost ins ta ntaneo us bre akd own when an attempt is made to extend them to new and more comp lica ted sort s of cases. Connect ionist m ach ines, in co ntrast, have been cons tructed that show a certain degree of el asticity. This, how ever, as we ha ve seen, is at the price of inte lligibility: the behaviour of th.e mach.ine, v.:hile well under stood mathem aticall y, becomes from the po int of view of our conce ptua l und erstand ing a black box . From this, however. it foll o.w~ too that the elasticity of even co nnec tionist sys tems must be of a very limited sor t. For if we do not under stand , conceptually, what is go ing on inside the machine, then we have no me ans of reconfiguri ng it in systemat ic ways to take acc ount of new circumstances or new sorts of proble ms. Suppose,. for exam ple, that we have a sys tem we have taught to translate s~oken in to wr itten English, a mac hine we have trained lip ~ n terms of ~tandard En gli sh pro nunciation. If, now , we want our machine to transcr.lbe , s~y , English pron oun ced with a We lsh accent, then we have n ~ way In whi ch we ca n sys tema tically adju st wha t has been learn ed s.o far in ord~r to take accou nt of th is new proble m, and we have to retrain the machine from scra tch." 12 13 . ..., IS The Connectionist Mind Barry Smith Co nnectionism is in the eyes of so me a radical posi no n prec ise ly becau se it is held to embod y a view according to which intelli gent, cogni tive processes are themselves non-intelligible; that is, are such as to resist our attempts at theoretical understanding. This view has been most clearly expressed by Dreyfus and Dreyfus, two long-standing critics of symbolic artificial intell igence research, who affirm that the app roach of classical AI 'seems to be failing because it is simply false to assume that there must be a theory of every domain. Neural net wor k modell ing, however , is not comm itted to this or any other ph ilosophical assumptio n' . (Drey fus and Dreyfus, 1988, p. 37) They conclude from this that any success of the connec tio nist approach will imply that one of the two primary goa ls of artificial intelligence research - namely that of cont ributing in positi ve ways to a theoret ical understanding of intell igence as such - will have to be abando ned. We shall return later to the sort of know ledge we can have of mind and other 'complex phenom ena' . most impor tant examples of such elements are the 'sensatio ns' of standard empir icist ph ilosophies. For Mach, however, sens ations are classified differently from elements of other sorts, not because of what they are intrinsically, but entirely in reflection of their relationship with other elements: ' the eleme nts of all complexes prove on closer inspection to be homogeneo us' (Mach, 1959, p. 14) . 'The antithes is between ego and wor ld, between se nsa tio n (appearance) and thing,.. . vanishes, and we have simply to deal with the connexion of the elements.. ., of which this antithesis was only a partially appropriate and imperfect expression' (ibid.) Ma c h's idea s will prove of interes t in what follows also because he embraces a doctrine of what he calls the 'eco nomy of thought' : we are to understa nd the constantly shifting world of elements-in-connection as subjec t thro ugh and thro ugh to the law of least action; the world is a dy namic sys tem within which elemen ts, in forming connections, seek always the path of least resistance in such a way as to establish stable minima; and the behaviour of e lements, including psychological elements, is for Mach to be understood en tirely from this perspective (which Mach deri ves from evolutionary biology). One thinker influenced tremendous ly by Mach is Hayek, and the doctrines at the heart of Hayek's psychology are in fact taken over directly from the older thinker. Hayek tells us that the problem of The Sensory Orde r is that of establishing the relationship between the 't wo orders ' of the subjec tive, sensory , perceptual and phenomenal on the one hand, and of the objective, scientific and physical on the other (Hayek, 1952, 110).5 T he ma in thesi s o f the work is indee d an ambitious extrapolation of Mach' s own thesis concerning the nature and status of sensations. Hayek wants to show that all attributes of men tal experience can be explained in terms of the syste m of connections of corresponding groups or patterns of nerve-excitations. He wants to show , in other words , that those mental properti es with which we suppose our selve s to be acqu ainted through introspectio n, throu gh the observatio n of other people' s beh aviour, thro ugh history, poetry and so on are - lock, stock and barrel - a matter of certai n structural or relational pro pert ies of the nerv ous sys tem. Where Mach , therefore, remai ns only on the threshold of con nectionism in holding on to a view of mind as an order of sensations (and thus of proper ' mental' entities on the ' symbolic ' level, to use Smole nsky' s terms ), Hayek crosses this threshold by em bracing a conception of mental properties as relational features of events occurring in the sub-symbolic doma in of nerve-excitations. On the other hand, however, Hayek is less radical than Mach . For where Hayek places his faith in physics and the physicist' s understanding of the nervous system, Mach sees the phys ical order 14 HAYEKIAN PSYCHO LOGY The significance of the opposition between symbolic and co nnec tionist approac hes is not restricted to the field of artificial intelligence. It is employed by Dreyfus and Dreyfus also as the basis for a new sort of history of ph ilosophy . According to the Dreyfusian scheme, philosophers are to be di vided into two groups: precursors of symbolic Al on the one hand - abo ve all Plato, Descartes, Brentano, Husserl , the early Wittgenstein - and precursors of connectionism on the other, amo ng whom are to be included not least Heidegger , Merleau-Ponty, Michael Polanyi and the later Wittgenstein. From the point of view of philosophers in the first group (according to the Dreyfus accou nt), human reasoning is a line ar , propositi onal matter; it is constrained by cont ext-free rules and built up out of sensory and co nceptual 'e lements' that are in the business of mirrorin g or picturing corresponding (predetermined) elements of external reality. From the poin t of view of the latter, in contrast, human reason is an intuitive , creative, co ntextual, hol istic affair - a matter, again, of knowing how rather than of knowing that. The philosoph ical story is of course a mite more complicated than Drey fus woul d have us believe. Co nsider the philoso pher-scienti st Mach , a think er who straddles the boundary between elementarism and holism in ways not read ily allowe d for in the Dre yfusian histor iograp hy. Mach' s philosophy is centred around the idea that the world is built up of 'element s' subjected to con stantly changing patterns o f combination. The 16 The Connectionist Mind and the associate d notions of space, time and ca usality as themsel ves mere ly the spurio us prod ucts of the 'economy of thought' ." To see how Hayek' s co ncep tion works we must note first of all that, while the external sti muli that cause activity in our nerve fibres can be classified in physical terms, stimuli that are phys ically the same will in many cas es not appear to us as the same from a qua litative point of view. This is so, Hayek holds, beca use it is not what is transmitted along a nerve fibre that is responsible for its characteristic effects , but that the fibre is exc ited at all ( 127 , 13 1, 142). More precisel y, Hayek reje cts view s, such as those defended by Johann es Muller and Ewald Hering, according to which it must be some property of the individual impul ses proceed ing in the different fibres tha t in some sense 'corresponds' to the differences of the sensory qualities. Hayek maintains, rather , that the specific character of the effect of a particular impulse is due ' neither to the attributes of the stimulus which caus ed it, nor to the att ributes of the impulse, but may be determined by the position in the struc ture of the nervous system of the fibre which carries the impulse ' (135). What we ca ll ' mind' is for Hayek nothing else than 'a part icul ar order of a set of eve nts taking place in some org ani sm and in some manner related to, but not identical with the physical order of eve nts in the en vironment' (149). But how, the n, can certain relation sh ips between nonme ntal eve nts turn the m into mental ev ents? Th is que stion ca n be appr oac hed first of all from an adaptive poin t of view: how is it possib le that in a certain pa rt of the physical order (namely in the organism) a subsystem is formed tha t has the property of intentional ity in the sense that it can be said to reflect some features of the physical order as a whole , in ways that have ada ptive significance? How does such reflection enable the orga nism to behave more appr opri ately in rela tion to its surroundings? Hayek , in contract with his predecessor, is not concerned with questions of 'atomism ' or 'elementarism ' : Like Mach , however, he is a struc turali st both with respec t to sensory qua lities and with respect to the world as a whole. Thu s for example he affirms that 'the whole orde r of sensory qualities ca n be ex ha ustive ly described in terms of (or "consists of nothing but") all relationships ex isting between them' (155). Or again: 'the orde r of sensory qualities no less than the order of physical events is a relatio nal order - eve n though to us, whose mind is the total ity of the relations constituting that order, it may not appear as such' ( 156).7 There are no abso lute q ualities of se nsations from Hayek' s point of view . Indeed, belief in absolute quali ties ' is probably one of the main roots of the belief in a peculia r mental substance' (191 ). Another such root is the belief in a store house of memo ries, which Hayek also rejects. Hayek also Barry Smith 17 anticipates the ce ntral thes is of AI research to the effe ct that mind (intelligence, cognition) can in principle be realised in a wide range of different material stru ctures. As the very same pattern of movement s can be instan tiated in a swarm of flies and in a flock of birds, so the very same abs trac t pattern can be instantiated in a musical score and a gramophone record . And then , as Hayek poi nts out: It is at least conceiv able that the part icular kind of order whic h we call mind might be built up from any one of seve ral kinds of different elements - elec trica l, chem ical or what not; all that is requi red is that by the simple relationship of being able to invoke eac h other in a cer tain order they correspond to the structure we ca ll mind (229 ). MENTAL TOPOLOGY We are to imagi ne the mind as a receptor orga n subjected to a constant barra ge of physical stimuli. Some patte rns of stimu li will tend to occur frequently toge ther. Thi s, Hayek hypothesises, will tend in turn to bring about esp ecially close connections betwee n the corresponding groups of fibres and it will tend to lead also to the formation of espec ially dense co nnections to corresponding central neuro ns. (Here Hayek definds a view very similar to tha t form ula ted by Donald Hebb in 1949 , a view that even today serves to define one of the sta nda rd types of model used in co nnectionist research.)! It is in this way that impulses occ urring in different fibres may come to be experie nced as qualitatively equ ivalent. The nervous system there by acqu ires a new sort of topological structure: a distance-function comes to be defined upon it, reflecting the deg ree to which different quality grou ps ' belong together' as a result of the bond s (or 'cell assemblies' ) betwee n thern.? The ' position' of a neuron in the resu ltant sys tem is the n a matter of co nnections, of the paths laid do wn to other neurons: ' groups of neurons may have a larger or smaller part of their connexio ns in common . We can thus speak of greater or smaller degrees of similarity of the pos ition of the different neuro ns in the who le system of such co nnexions ' (326). And then : ' A very high deg ree of similarity in the position of the dif ferent neurons in the sys tem of connex ions is likely to exis t wherever the neurons are served by receptors sensitive to stimuli which alw ays or almost alwa ys occ ur together' (328). Furth er, because we can expect receptors that are sens itive to the same kind of physical stimuli to be frequently excited at the same time.'? it will follo w that correspond- 18 The Connectionist Mind Ba rry Smith ingly dense networks of connections will also be established laterall y between the correspondi ng central neuro ns. In add ition, the forma tion of cell assemblies will be similarly affected by ch anges in the inner states of the body (with the occ urrence of feelings of pain, hunger and so on ) as well as by correspondi ng actions. Thu s we can take a ste p towards explai ning the adaptive properties of the ' mind' as a sys tem of relat ionships: classificato ry co mpete nce and classificatory performance are (as for the connec tionists) one and the same. Moreover: in man y instances it is likely that certain kinds of stimuli will usually act toge the r on the org anis m when the organ ism itself is in a particular state of balance or of activity, ei ther beca use the stimu lus regularly occ urs under conditions produci ng that state, or because it occurs pe riodicall y so as to coi ncide with some rhythm of the body . The impul ses which reg ister such externa l stimuli will the n beco me connected wit h impulses rece ived from the propriocep tors which reg ister the different states of the orga nism itself (333). It is this internal nesting structure, Hayek holds, that expl ains how what is after all a me re system of impulses 'can produce "models" of extreme ly complex relationships between stimuli, and indeed can reproduce the order of any co ncei vab le struc ture ' (336). Thi s in turn would explain not merely outer- and inner-d irected intentionality (or in other words conscio usness of world and of self), but also the unlimited range of ment al compre hension (our capacity to direc t ourselves inte ntionally to objec ts in all conceiva ble categori es ). The sophistication of the sys tem is increased still further via the distinction between ef fective and potential co nnec tions between neu rons (two kinds o f following). Hayek recognises also - with his connectionist successors - tha t there are influences transmitt ing inhibition (or in other words negative weightings), the existence of which 'ex tends the range of possible differen ces in the position which any one impulse may occup y in the whole system of con nexions ' by introducing the possibility of impu lses having effects tha t are directly opp osed to eac h other (346) . The whole resultant process is then not so much a matter of 'classification' as of 'evaluation' , since the proce ssing engine is capable of making distinctions of degree (both positive and negative) as well as distinctions of kind (357). Th e system of multiple clas sificatio n ca n cope not only with isolat ed se nsory input s but also with com plex Gest alt stru ctures (higher-o rder categorial formations) of a range of different sorts: A sys tem of co nnections hereb y grow s up in which is recorded the relat ive freque ncy with which, in the history of the orga nism, the differen t groups of intern al and external stimuli have acte d together. Th e psychological classifica tion of physic al stimu li is then effe cted , in the s imp les t case, as follo ws. Each sep arate primary impulse has, as a result o f these acquired conn ections, a certain / ollowi ng (a train or wake ) of sec ondary imp ulses (recall the pass age of numerical values through the nodes of the co nnec tionist system). It is the total or partial identity of this followi ng that makes given primary impu lses mem bers of the same class (334 )." Note that the followi ngs are a matter of purely phys iological co nnectio ns betwee n imp ulses - they are not a matter of association s between (meaningful) mental events . Here again, therefore, Hayek is at one with the co nnectionist tenden cy with in contem porary cognitive science. MULTIPLE CLASS IFICATION The classificato ry syste m hereby co nsti tuted is of great sophistication. Each impulse becomes a member ' not merely of one class but of as many distin ct classes as will correspond, not only to the numb er of other impulses which cons titute its followi ng, but in add ition also to the number of possibl e combinations (pairs, triples . . .) of such other impulses' (335). Thi s ge nerates a system of mult iple cla ss ificat ion of a sort in which 19 the classificatory responses are not diffe rent in kind from, but are eve nts of the same sort as, those which are the obje ct of class ification. In consequence it is possi ble that one and the same even t may appear both as an object of classifi cation and as an act of classification (336). The fact that chai ns o f further processes ('associations' ) can be evoked not only by the 'eleme ntary ' sensory qualities . .. but also by certai n 'abstract' attributes of diffe rent gro ups of sensations (such as figures, tunes, rhythms, or abstract concepts) , has usually been regarded as an insurm ou ntab le obstacle to any physiological explanation of mental processes.F For the approach follow ed here no such difficulty arises: the problem of the equivalence of 's imilar' complexes of stimul i is not different in principle from the prob lem why the same assoc iations should beco me attached to different impu lses which corres pond to the same 'elementary' qualities (369).13 21 The Connectionist Mind Barry Smith This se nsitivity to Gestalt struc tures in the environ ment is built up only slow ly, as the brain becomes tuned to those abstrac t pattern s that are of adaptive significance. There is forme d thereby a ' map' reproduc ing relations between cla sses of events in the env ironment in a process that is affec ted both by the sector of the world where the individ ual lives and by the individ ual' s own body. This map (or ' mental model') is then of course not o nly very imp erfec t but also subj ect to contin uo us although very grad ual change (526) . Its representations arc structural (co mpare the imperfect and abstract way in which cartographical maps represent topographica l fea tures) and they are a matter of accumulated knowledge , so that the map represents only the kind of world the org anism has experience d in the past. Th e map ' s job, from Hayek ' s perspec tive, is to facilitate a ce rtain sort of cognitiv e laz iness (or what Mach ca lled 'thought economy' ). It allows the org anism to avoid the need to classi fy individual event s and com bination of events from scratch in each suc cessive occ asion of confrontation. The map' s structural character may also help to explain a central feature of animal intelligence: plas ticity - the capacity to reuse rules learne d in one context in other struc turally similar but perh aps otherwise qu ite alien contex ts (a ca paci ty that may also be illustrated , on the leve l of high intelle ct, in the use of analogy in science). Th e map thu s plays the rol e of memo ry in Haye k' s theo ry - not however that of a passive memory storehouse, but of an active memory competence. Mo reover this active competence in a certain sense precedes our articu late mental expe rience. For as Hayek poi nts out: of that object at all , but a set of relations by which our nervous syste m 20 we do not first have sensations whic h are then preserved by memory , but it is as a res ult of ph ys iological mem ory tha t the physiological impulses are converted into se nsa tions. The co nnexions bet ween the physiological eleme nts are thus the primary phe nome non whic h creates the me nta l pheno mena (250). But no w, because this memory competence cum classificatory order is subject to constant evolution, it follows that so too is the sensory world in which we live. The rich ness of this world, 'is not the starting point from which the mind der ives abstract ion, but the produ ct of a great range of abstractions which the mi nd must possess in order to be ca pable of enc ing that richness of the partic ular' (Hayek, 1978 , p. 431). It is as if we crea te the wo rld as we go along by imposi ng classificatory schem es upon the mate ria prima wh ich is the phys ica l substrate. The particul ar qualit ies we attri bute to an object in our ex perience are in this sense 'n ot properties classifies them ' (637) . As was pointed out by Joh n Gr ay ( 1986), there is a cert ain grai n of Kantianism here .!" thou gh it is questionable whether this Kantian elem~nt (and the sol ipsism that thr eatens in its wake ) can be see~ as bel~g extended thro ugh the whole of Hayek' s thinking if the latter IS to retai n any degree of coherence . Note, too, that while the thesi~ that we c~eate the worl d as we go along by imposing neurally determme? c1as~lficato~y demarcations upon physical stimuli may be cap able of being defended 10 relatio n to the wor ld of se nsory qualities , it faces obvious difficulties when an attempt is made to extend it to the 'objective' worlds of physics and neur ology. Fort una tely there is nothin g in Hayek' s work to suggest that these dom ain s, too , might be mer ely the reflec tio ns of evolved classifica tory competences of physicists and neurologists. ~heir ~Ievelop­ Hayek 's contention is that human beings in the cou~'se ment build up a system of differentiations between stim uli 10 which each stim ulus is aiven a de finite place in a slowly changing 'o bjec tive ' order of increasing ~omplexity and sophistication. It is this place in .the ob~ective order that ' represents the sig nificance which the occurrence of that stimulus in different co mbinations with other stimuli has for the organism ' (2 17). The central nervous sys tem is an ada ptive engine for the constant reclassification - on many le vels (including co nceptual and emotional levels) - of the legi on o f imp ulses proceedin g in it at any momen~. We crea te the world in which we live in the sense that there are, on the Side of nerve excitations , no fixed conceptual units able to mirror or picture corresponding (predeternlined) ele men ts of external reality in a o~e-to-on~ way . Onl y insofar as the nervous sys tem has learnt to treat a particular st~mu lus eve nt as a mem ber of a certain class of eve nts, can this event be perceived at all, for only thus ca n it obtai n a position in the system of sensory qua lities . IS 0: ARGUMENTS AGA INST HAYE KIAN- HEBBIAN CONN ECTI ON ISM First, one import ant argument agai nst a view of the sort outlined above !s of a type that ha s been mar shalled against neurologicall y or.ie~ted ~ork 10 co gnitive science quite generally and relates to a characteristic ~lsuse of language that con sis ts in imputing to parts of the hum~n orgam~m what ca n properl y - if words are to retain their standard meanings - be Imputed on ly to the who le. This occ urs , for ex ample, in phrases such as 'the nervou s sys tem perceives ' , ' the system of acq uired con nec tions remem - .,.. ! 22 The Connectionist Mind bel's', 't he processing engi ne learns ' , and so o n. It seems not possibl e to spec ify the me anings of terms such as 'perce ives ', ' remem bers ' and ' learns ' in these phr ases in ways that will be co nsistent with our normal use and understandi ng of the term s (so that there is a sense in which we do not really know wha t these phrases mean ). Second , Hayek' s view, like that of Ryle in The Concept ofMind (1949 ), has no way of dealing with delib erate, conscious thinking (with reaso ning as a logica l process) - this, too, is a sta nda rd failing of co nnect ion ist approaches - and its account of consciousne ss is correspo ndi ngly weak . Hayek himself treat s 'conscio us experience as merely a special instance of a more general phenomenon ' , holding that the sphe re of mental eve nts _ tha t is, the sphere of even ts that are orde red o n princ iples analogo us to those re veal ed in introspectio n - 'evidently transcends the sphere of conscious even ts and there is no j ustification for the attit ude frequently me t that e ither ident ifies the two or even ma intai ns that to speak of unco nsci ous menta l even ts is a co ntra diction in term s ' ( In ]). Thi s implies , ho we ver, that Hayek has, like Mac h and his Gestal tist successo rs, no mea ns of draw ing a clea r distinc tio n betwee n intentionality as a matter of reflection (or ' isomorphism') and intentionality as a matter of 'conscio usness of' or 'aboutness' in the sense of Brentano and his foll owers. Third , Hayek has no way of dea ling with what we might ca ll ' me ntal causality ', or in other words with the connec tion (so central to the work of Lud wig von Mises) between reaso n, choice and actio n. Th e system of The Sensory Order leaves no apparent room for plan ning, for se lf-contro l and for the deliberate se lf-shapin g of the conscious subj ect (no room, indeed, for any self or ego or fo r the unity of consciousness). J6 A concl usion of this sort might well be acceptab le to the followers of Hume and Ma ch , for whom consciousness is in any case dispersed and reactive.!? But what, the n, is to serve as a basis for the methodological indivi dualis m that so pervades Hayek ' s work in the soc ial scie nces ? Fourth, Ha yek ' s view has no way of explai ning the relat ive stability over time of our cognitive faculties and our qualitati ve contents, and it has no way of expl aining the massive similari ties in cog nitive capac ities be tween o ne indi vidual and ano ther. Kluver inadv erte ntly re veals the weakness of Hayek ' s position in this respect when , in his 'In trod uctio n' to The Sensory Order, he praises Hayek ' s theory becau se it suggests certain definite line s of experimentation. ' For instance, it should be poss ible not only to change sensory qu alities exp erimen tally, but to crea te altogether new sensory qualities which have never been experienced before ' . 18 Th e prob lem for Klu ver and Hayek (as also for Hebb and his successors) is that it seem s not to be possible to change or inve nt sensory qua lities in this Ban }' Smith 23 way: a certain fixity of species seems to pertain to the world of qualitative experience. The hypo thes is of cognitive universals has indeed found empiri cal suppo rt ac ross a wide rang e of sensory and intelle ctual phenomena in recent years, and it seems that most linguists and anthropologists wou ld now adays assume as a matter of course that the truth of this hypothesi s is pr esupposed not merely by the intertran slatabil ity of all known languages but already by the very fact of lingui stic communication itself. Hayek' s acco unt of the acq uisition of perceptual and conceptual cla ssificatio n skills see ms moreo ver to amount in the end to a sort of tabula rasa view of these matters that seems difficult to square with what is now known abo ut the impressive mental compe tences of new -born babies . Fifth, it is not clear ho w an approach along Hayekian-Hebbian lines can do j ustice to the crea tive open-endedness that seems to be involved eve n in our day- to-day act ivities of mental classification (a feature of our mental life that Hayek is otherwise at pains to insist upon) . For in the real world, as Wittgenstei n, among others, has emphasised , cla ssification is often far from being a neat and tidy affair, yet it is a limitati on of standard co nnect io nist systems that they can be made to ope rate (the corresponding algorithms of back-propagation ca n be applied ) only if the relevant cl assi ficatory space is fixed in advance. 19 TH E THEORY OF COMPLEX PHENOMENA Wh at, the n, ca n be learn ed in a positive se nse from Hayek 's work in psychology? Hayek shows, most importantly, that the central idea behind the connectio nist paradigm is at horne not only in psychology and neurology but also in the sphere of economics. For the mind, from the perspective of The Sensory Order, turn s out to be a dynamic, relational amiir that is in man y respects analogous to a market process . The mind is a 'continuous strea m of imp ulses, the signi ficance o f each and every contributio n of which is determined by the place in the pattern of channels through which they fl o w' , in such a way that the flow of representative neural impulses can be compared 't o a stock of capital be ing nourished by inputs and giving a co ntinuous stream of outputs ' (Hayek, 1982, p. 291 ).20 In 'The Use of Knowledge in Soc iety' (first publ ished in 1945), Hayek describes the pri ce sys te m as a mechanism fo r communicating information; and then , in the mind as in the market sys tem, it is remarkable how little exp licit (conscious) knowl edge is required by the agent in order for him to be able to react in appro priate ways to changes in his circumstances. In the 24 25 The Connectionist Mind Barry Smith mind as in the market the mo st ess ential info rmation is passed on in the form o f abbreviated 's ignals' (as contextually situated nerve impulses or prices, respec tively). The pri ce sys tem is complexity of the se relatio nships is so grea t, and they are subjec t to such con tinual variation. that in an y case we could never reach the po int wher e we co uld exhaustively describe them . On the other hand the theo ry of co nnectionist devi ces - of Hebb Model s and Bo ltzm an n Machines, of Hopfield Nets and Mu ltilayer Perc eptrons is ma the mati call y we ll de veloped, and the que sti on thu s ar ises as to whethe r this new bod y of theor y might not be used to throw new light on those 'c omplex phe nome na' that are at the cen tre of Hayek' s interes ts not only in psyc ho logy but also in ec onomics and the social sc iences in general. Does co nnectio nist the ory yie ld a new solution to the old problem at the hear t of all economic theori sing in the Austri an tradition, a problem (crudely pu t) that turn s on the fac t that the phenomena with which the Austria ns w ish to deal in their theor y are seen by the Austri ans themselves as bei ng in a sense too 'comp lex ' to be theori sed about? Hayek him self, fam iliarly , holds that in relation to the complex phenom ena of the soc ial and psycholog ical sciences we ca n have on ly 'qualitative ' understanding, and not ex act pre d ictio n. T he co mplex phen om enon that is the eco nomic system ca n thus not be made ra tional or subject to 'control' .21 Indeed predi ctio n and co ntrol are already impossible in biolo gy (co nsider the problem of predic ting how a given spec ies will ev o lve), and social, psycho logical and biologi cal phen om ena are such that the notion of a law of nature as a re la tio nship obtaining between a few phenomen a, linked together by a simple relationsh ip such as cause and e ffec t - a notion most at hom e in the field of astro no my - ca nnot be applied. As We imer puts it, 't he prejudice tha t in order to be scie ntific one must produce laws may yet pro ve to be o ne of the most harmful of methodological conc eptions ' (Weimer, 1982, p. 244). And as we are now onl y too aware, the prejudice that scientific knowledge in this sense has alre ady been attained in relat ion to the soc ial wo rld has proved to be one of the most harmful concepts in the sphere of politi cal economy. . a kind of machinery for registerin g change, or a sys tem of telecommun ica tions which enables indi vidu al produ cts to watch me rely the movement of a few po inters, as an en gineer might watch the hand s of a few dials, in order to adj ust their activities to changes of which they may never know more than is reflected in the price mo veme nt. . .. The marvel is tha t in a case like th at of a sc arci ty of on e raw material , without an order bei ng issued, without more than perhap s a handful of peo ple knowing the ca use, tens of thousands of people whose identi ty co uld not be ascerta ined by months of investigation , are made to use the material or its products mo re sparing ly; that is, they move in the righ t d irection (Haye k, 1949, p. 87). Bo th the m ind and the sys tem of the market order are products of tho usands of years of evo lutio n, and both evolv ed through a massive number of tria ls and errors and throu gh associated processes of 'training up ' , as each succ essive generation of indivi duals learn ed on the one hand to be co nsc ious, and on the ot her hand to play the role of market part ic ipan t. How ever the analo gy bet ween min d and market breaks down at least in this respe ct: that the former cannot on pain of con tradiction be seen as hav ing co me int o bein g ab initio as the unin ten ded conseq uence of intended act ions. Th e struc tura l simi larit y betwee n the psychological and the eco nomic sphe re is und ers tandable fur the r in term s of the fact that both hav e the same dynami c root , for eco nomic cha nge rests in no sma ll part no t on abso lute values but on our (al ways relati ve) valuations, and the latter are a ps yc hological matter, a matter of co nstantly sh iftin g and cha ngi ng networks of relations determined throu gh and thro ugh by con text and per spective and by the 'economy of thought' . Minds and markets are also co mparable in that the understanding we ca n have of each is of a quite different nature from that wh ich we can have o f phy sical sys tems . This is indeed, for our pre sent purposes, the most important respect in which Hayek antic ipates co nte mporary co nnectio nism . As Hay ek points ou t, the oper ations of the mind are non persp ic uo us bec au se we are not explicitly aware of the re lat ion sh ips between the different qualities that constitute the mental order: we merely mani fest these relationsh ips in the discrim ination s we perform (a case of knowing how rather than knowin g that ). Furthermore the number and Notes I. 2. This chapter is a result of work carried out on the project 'FormalOntological Foundations of Artificial Intelligence Research' , sponsored by the Swiss National Foundation. I am grateful to Graham White, Wojciech Zetaniec, Gloria Zuniga and participants in the Hayek Memorial Symposium for helpful comments. Current chess-playing computers are not the result of research in artificial intelligence in this sense, since they operate on the basis of exhaustive search strategies quite different from those employed by human chess players. 26 3. 4. 5. 6. 7. 8. 9. 10. II . out, The sub-symbolic programs of connectionism can, as is often be translated in such a way that can also be on classical (von Neumann) machines; but the translated programs are then not the kind of 'symbolic' program that the Helmholtz hypothesis requires. Partridge, 1990, 71. References in form are to the numbered sections in The Order 1952), with the decimal point removed. See Mulligan and Smith, 1988, pp. 1501'. As notes, a structuralist approach of this sort does, however, very well explain the existence of certain patterns of which we are aware. Thus, for example, it can explain the existence of those 'intermodal ' attributes that such as strong, weak, mild, are revealed in our use of mellow, tingling thick, rough, hot, corse, hollow, luscious, smooth and so on in relation to sensory ties of every sort (I On the role of 'Hebb models' in the development Cowan and Sharp (1988, pp. In a note to that that there is a link between Hayek's work and the of himself more recent connectionist research (ibid., p, 119, note 57). Order (sec p. vii, and also adverts to Hebb at various points in The sections 249, 3 I5,334,538). The idea of mental topology can be conceived as a of the marginal approach adopted by Hayek and other of the Austrian school in economic theory. As Hayek writes: 'we shall expect fairly close connexions to be formed between the neurons served by neighbouring receptors which are sensitive to stimuli which occur frequently together because they emanate from the same physical objects, such as pressure and temperature, certain chemical agents acting simultaneously on mouth and nose, etc., etc.' (331) . , Hayek draws attention to the disIn his 'The Sensory Order after 25 positional connotations of this terminology of following, and also to the of the neural redundancy and randomness involved in the classificatory system: The 'followings' of all the impulses in the central nervous or system at anyone time are thus assumed to determine the readiness of the system to do new things internally or Which of these potential neural events (toward which the system is inclined at any particular moment) eventuates would be decided by the partial overlapping of these followings through which, by summation, the potential effects of those linkages would be made actual. Only where a sufficient number of impulses coverged on anyone neuron would it be made to 'fire' and to send out impulses to hundreds or thousands of other neurons (Hayek, 1992, p. 290). 12. 13. Barry Smith The Connectionist Mind For example by Stout, 19I5. Mach, too, allowed his sensory elements to picture certain qualities of higher order. Cf. Schulzki, 1979, pp. I58ff. 14. IS. 16. 17. 8. 19. 27 See also Petitot and Smith, 1990, but contrast Smith, 1990, p. 265. At one point goes so far as to assert that an event that sets up, via peripheral stimulation, impulses in the brain of an entirely new kind 'could not at all' (636). As Brentano stresses (1973, Book Il, Ch. IV), mere similarity and association, however and many-levelled, do not as such add up to For Mach, the self 'is not a monad isolated from the world, but a part of the world and caught up in its flow ... so that we will no longer be tempted to see the world as some unknowable something. We are then close enough to ourselves and sufficiently closely related to the other parts of the world as (Mach, 1917, p. 462) . to be able to hope for genuine Kluver, 1952, pp. xxi I'. Compare the remark of Dreyfus and Dreyfus to the effect that 'All neural net modelers agree that for a net to be intelligent it must be able to gene.ralize; that is, given sufficient of inputs associated with one particular output, it should associate further inputs of the same type with that sa~: output. The question arises, however: What counts as the s am~ type. (Dreyfus and Dreyfus, 1988, pp. 38f). The oppos ition at issue h~re.ls neatly captured in the following remark by G. L S, Shackle on what IS, in effect, the classification problem faced each day in the business world: Kaleidic effects , typically the response of asset values to 'the news', the abrupt and necessarily unforeseeable reaction of expectation-formers to announcements which in their nature are unhe ralded and of a purport quite unknown in advance, pose important problems of notation,. The basic and essential character of the kaleidic phenomenon renders mappropriate the accepted methods of ana lysis of fully-known probl~ms. It is the fact, astonishing and yet natural, that m econormcs we habitually and, it seems, unthinkingly assume that the problem facing an economic subject, in especial a business man, is of the same kind as those set in examinations in mathematics, where the candidate unhesitatingly (and justly) takes it for granted that he has been given enough information to construe a satisfactory solution. Where, in real life, are we justified in assuming that we posse ss 'enough' information? (Shackle, 1972, p. 184). 20. 21. Cf. also Lavoie et al., 1990. 'The rationality of complex systems is not localisable in a single locus of control, and it is therefore never "conscious'" (Weimer 1982, p. 245), References Brentano, Franz (1973) Psychology from an Empirical Standpoint, translated from the lst German edition of 1874 (London: Routledge & Kegan Paul). Cowan, Jack D. and Da vid H. Sharp (1988) 'Neural Nets and Artificial Intelligence' , Daedalus, vol. I 17, no. I, pp. 85-122. Th e Connectionist Mind Barry Smith Dreyfus, H. L. and S. E. Dreyfus (1988) 'M aking a Mind versus Mode ling the Brain. Artificial Intelligence Back at a Branchpoint ' , Daedalus, vol. 117, no. I, pp.15-43. Gray, John ( l986 ) Hayek on Liberty, 2nd edn (Oxford: Blackwell). Hayek, F. A. ( 1945) T he use of Knowledge in Society,' American Economic Review, vo l. 35, pp. 5 I9- 530. Reprinted in Hayek, 1949. Hayek, F. A. (1949) Indi vidualism and Economic Order (London: Routledge & Kegan Paul). Haye k, F. A. ( 1952) The Sensory Order. An Inquiry into the Foundati ons of Theoretical Psychology (London: Routledge & Kegan Paul). Hayek , F. A. (1978) The Primacy of the Abstr act' , in New Studies ill Philosophy, Politics, Economics and the History of Ideas (London and Henley: Routledge & Kegan Paul), pp. 35-49. Hayek, F. A. ( 1982) ' The Sensory Order After 25 years' , in Weimer and Palermo, 1982, pp. 287-93. Hebb, D. O. ( 1949) The Organization of Behavior. A Neuropsychological Theory (New York and London: John Wiley). Kluver, H. (1952) ' Introduction' , in Hayek, 1952, pp. xv- xxii. Lavo ie, D., H. Baetjer and W. Tulloh (1990) ' High-Tech Hayek ians: Some Possib le Research Topi cs in the Economics of Computation ' , Market Process, vol. 8, pp. 120-4 1. Mach , Ernst ( 19 17) Erkenntnis und Irrtum: Skizzen zur Psychologic der Forschun g; 3rd edn (Leipzig: Barth). Mach, Erns t (1959) The Analysis of Sensations , translated from the 1st German edition of 1886 (New York: Dover). Mulligan, Kevin and Barry Smith ( 1988) ' Mach and Ehrenfels: The Foundations of Gestalt Theory', in B. Smith, (ed.), Foundations of Gestalt Theory (Munich and Vienna: Philosophi a), pp. 124-57. Partridge, D. ( 1990) 'Connectionism is Better for Engineers than for Scientists' , in Tiles, et al., pp. 55- 84. Petitot, Jean and Barry Smith (1990) ' New Found ations for Qualitative Physics' (with commentary by Adam Morton ), in Tiles, et al., pp. 231--61 . Ryle, Gilbert ( 1949) The Concept of Mind (London: Hutchinson). Schulzk i, Ewald ( 1979) Der Mensch als Elementenkomplex und als denkokonomische Einhe it. Zur Anthropologie Ernst Machs , dissertation (Univers ity of MOnster). Shackle, G. L. S. ( 1972) Epistemics & Economics. A Critique of Economic Doctrines (Cambridge: Cam bridge University Press). Smith, Barry (1990) ' Aristotle, Menger , Mises: A Categorial Ontology for Economics', History of Politi cal Econom y , ann ual supp lement to vol. 22 , pp. 263-88. Smo len sky, Paul ( 1987) 'On the Prope r Treatment of Co nnectionism ' (University of Colorado at Bould er , Departm ent of Compute r Science , vari ously reprinted) . Stou t, G. F. (19 15) Manual of Psychology, 3rd edn (London: University Tutorial Press). Ti les, J. E., G. T . McKee and C. G. Dean (eds) (1990) Evolving Knowledge in Natural Science and Artificial Intelligence (London: Pitman). Weimer, Walter B. (19 82) ' Hayek's Approac h to the Problems of Complex Pheno mena: An Introduction to the Theoretic al Psycholo gy of The Sensory Order ', in Weimer and Palermo, pp. 241-85. Weimer , W. B. and D. S. Paler mo (eds) (1982 ) Cogni tion and the Symbolic Processes , vol. II (Hillsdale, NJ: Lawrence Erlbaum). 28 29