Collaborative information environments to support knowledge construction by communities

AI & Soc (2000) 14:314-330 �9 2000 Springer-Verlag London Limited ~tl ~t S O C I I ~ ' ~

Early-Connectionism Machines

Roberto Cordeschi Department of Communication Sciences, University of Saleno, Fisciano, Italy

Abstract: In this paper I put forward a reconstruction of the evolution of certain explanatory hypotheses on the neural basis of association and learning that are the premises of connectionism in the cybernetic age and of present-day connectionism. The main point of my reconstruction is based on two little-known case studies. The first is the project, published in 1913, of a hydraulic machine through which its author believed it was possible to simulate certain "essential elements" of the plasticity of nervous connections. The author, S. Bent Russell, was an engineer deeply influenced by the neurological hypotheses on nervous conduction of Herbert Spencer, Max Meyer and Edward L. Thorndike. The second is the project, published in 1929, of an electromechanical machine in which the author, the psychologist J.M. Stephens, believed it was possible to embody Thorndike"s law of effect. Thus both Bent Russell and Stephens referred to the principles of learning that Thorndike defined as "connectionist". Their attempt was that of simulating by machines at least certain simple aspects of inhibition, association and habit formation that are typical of living organisms. I propose to situate their projects within the frame of the discovery of a simulative (modelling) methodology which I believe might be considered an important topic of the "Culture of the Artificial". Certain more recent steps toward such a methodology made by both connectionism of the 1950s and present-day connectionism are briefly pointed out in the paper.

Keywords: Artificial intelligence (history of); Connectionism; Cybernetics; Simulation of behaviour; Learning theory

1. Introduction

In a former paper in AI & Society I put forward certain points by which it might be poss ib le to give a def ini t ion of what I be l ieve to be an impor tant topic of the Culture of the Artif icial : the machine simulat ion methodology of l iving organism behaviour. 1 In put t ing forward these points, my aim was that of isolat ing recurr ing themes which help unders tanding the deve lopment of such a machine s imulat ion methodo logy from its, so to speak, discovery during the first hal f of the twentieth century up to the present time. These points could also serve the purpose of identifying convergences and d ivergences be tween both the different research p rogrammes of AI and "new"

Early-Connectionism Machines 315

connec t ion i sm and the s imulat ive projects of the age preceding the deve lopment

of AI, indeed preceding cybernet ics itself: the age ! there proposed to define as

"proto-cybernetics". It was during this age that the simulative or behaviour-modell ing

approach of the Culture of the Artif icial was discovered. I tried to show how Clark L. Hull"s "robot approach" was an interest ing case study of the discovery of such a s imulat ive methodology during the late 1920s and the mid-1930s.

Hull h imse l f ment ioned two former machine s imulat ion projects that in fact seem to share the core of at least some of his methodological assumptions: the two machines designed by the engineer S. Bent Russell (1913) and by the psychologis t J.M. Stephens (1929). 2 Both machines were designed so as to embody certain

hypotheses on the plast ici ty of nervous connect ions pointed out at the t ime by

psychologis ts in order to explain the physical bases of learning. In both cases Edward L. Thorndike 's laws of learning were considered, but far more explici t ly

by Stephens than by Bent Russell . As we shall see, if Bent Russell was more interested in the s imula t ion of certain aspects of the law of exercise, Stephens was

directly interested in the s imulat ion of the law of effect, i.e. the very core of

Thorndike ' s theory of learning. It is stressed here that it was Thorndike who first used the term "connect ionism"

in order to define his own research on the laws of learning. Referring to Thorndike, and to Hull too, Donald Hebb called "connect ionis t" his own approach, as did Frank Rosenblat t , this t ime referr ing to Hull and Hebb. In this paper I shall use the

term "early connec t ion i sm" in regard to Bent Russe l l ' s and Stephens ' machines in order to keep clear the dis t inct ion between these machines and those embodying Hebb ' s rule (or its variat ions) during the 1950s, which I shall touch upon in the

last section of the paper. 3 In fact the main aim of this paper is to point out the s ta tement of certain hypotheses that at tempted to account for the neurological

~See Cordeschi (1991). In that paper I pointed out five points which might be so summed up. (i) Functionalism. Adaptive, learning and intelligent processes are independent of the features of organic structure: there is a functional equivalence between their instantiation in an organic structure and their instantiation in an inorganic or artificial structure. (ii) Simulation. The hypothesis that adaptive, learning and intelligent processes are mechanical is tested by constructing simulation models. This modelling ("synthetic") method is a proof of the sufficiency of these mechanical processes and the redundancy of vitalistic explanatory notions. (iii) Representationalism. The knowledge of the external world occurs, at least in higher organisms, through representations. The same could be said of certain artefacts that simulate their behaviour. (iv) Identity of explanatory principles. The explanatory principles of more "simple" processes of adaptation, learning and intelligence are not qualitatively different from those at work in the "higher" processes, such as insight, creativity or consciousness. (v) Mentalism. By using a mental or teleological language, an external observer can legitimately describe the behaviour of an artefact simulating an organism behaviour as an "intentional system" (to use Daniel Dennett's well-known term). For a first justification of these points see my above-mentioned paper on the discovery of the artificial; for further details see Cordeschi (1998). A statement of the general framework of the Culture of the Artificial is given by Negrotti (1999).

2Bent Russell's and Stephens' papers were mentioned by Krueger and Hull (1931). These papers fell into oblivion for a long time (at least Stephens' paper was mentioned incidentally by Boring, 1946, and Miller et al., 1960). S. Bent Russell, an engineer from St Louis, was later the author of several papers (mostly published in the Psychological Review during the last tents) in which he supported a mechanistic theory of learning influenced by John Watson' s behaviourism. His hydraulic device was discussed by the psychologist Max Meyer (see below) and by the biologist Judson Herrick in his popular book The thinking machine (Herrick, 1929). The psychologist J.M. Stephens tried a mechanistic interpretation of the law of effect (see Postman, 1947, for Stephens' role in the history of the law of effect). As Stephens himself pointed out later on, he had tried to reduce other laws of learning, including conditioning, to such a law (Stephens, 1967).

316 R. Cordeschi

bases of learning, and thus paved the way to Bent Russe l l ' s and Stephens ' ear ly connec t ion ism machines . It is sugges ted that these machines might be v iewed as case studies of the discovery of the above-ment ioned simulative methodology. Certain present-day important steps toward such a methodology are discussed in the concluding sect ion of the paper.

2. Meyer's Drainage Theory and Thorndike's Connectionism

Let us therefore begin by consider ing the theories on nervous conduct ion that Bent Russe l l took into account for the project of his machine which he publ i shed in 1913. The authors he ment ions in his paper are Herber t Spencer, Max Meyer and Edward L. Thorndike - a somewhat mixed bunch.

Meye r had formula ted his vers ion of what was to become known as the "drainage theory" of nervous conduct ion. 4 Such a theory rested on the old analogy of the nervous flow running like a l iquid through a system of pipes of varied (and modif iable) capaci ty , connected by one-way valves corresponding to the synapses (al though the synapse hypothes is was not universa l ly accepted when introduced). 5 In contrast, Thorudike had proposed a theory of the formation and reinforcement of S - R (s t imulus- response) connect ions, soon to f ind its own analogy in the te lephone switchboard, i.e. a ne twork of countless units, cor responding to the neurons, which enable the nerve message to be sor ted th rough a vas t range o f m o d i f i a b l e connec t ions , cor responding to the synapses. 6

Whi l e drawing largely on Spencer (1890) and, more direct ly, on Meyer (1911) and Thorndike (1911), Bent Russel l conjectured that:

1. cont inuous or f requent s t imulat ion of neurons at short t ime intervals could result in a s t rengthening of the connect ions be tween neurons and in an increase of their conduct iv i ty (or in a decrease of their resis tance);

2. d iscont inuous or not f requent s t imulat ion of neurons could result in a weakening of the connect ions and in a decrease of their conduct ivi ty (or an increase of their resis tance) .

These two statements give us the nervous sys tem image of the t ime as conce ived by Bent Russel l . W e shall refer to it as the standard theory of nervous conduct ion.

3See Hebb (1949) and Rosenblatt (1958). On some issues of the history of old and (more or less) "new" connectionism, see below and Cordeschi (1991, 1998). Insightful analyses of the history of connectionism are given by Walker (1990) and Savage and Cowie (1992), which distinguish themselves from recurrent generic as well as anachronistic investigations, as for example Valentine (1989).

4The psychologist Max Mayer came to the United States from Germany. Meyer's work and long life are reported by Esper ( 1966, 1967). According to the psychologist Walter Pillsbury, Meyer, in his 1911 book The Fundamental Laws of Human Behavior, "propounded all the essentials of the doctrine [of behaviourism] a year or more before Watson adopted his position" (quoted by O'Donnell, 1985: 215).

5Meyer himself did not accept the synapse hypothesis (see Meyer, 1912). Before Meyer, drainage theory was stated particularly by Spencer (see the first 1855 edition of The Principles of Psychology) and fully developed by William James and William McDougall (see James, 1890; McDougall, 1905). Drainage theory is a little- studied chapter in the history of neurology between the nineteenth and the twentieth centuries. An account of this is given by Cordeschi (1998: Ch. 2).

6On Thorndike's connectionism see Hilgard (1956: Ch. 2). See also the monograph by Jon~ich (1968).

Early-Connectionism Machines 317

Basically, the - by no means new - idea was that repetition of stimulation played an important role in the process of habit formation or of learning: the so-called law of practice. Confining ourselves solely to the authors mentioned by Bent Russell, we may trace the most generally accepted version of this law back to the above-mentioned first edition of Spencer (1890). While accepted as necessary equally by Meyer (1911), for his drainage theory, and Thorndike (1911), for his theory of S-R connections, the law of practice was seen by both as an insufficient or, at any rate, secondary law in any account of the nervous correlate of habit formation and learning.

Thorndike stressed its insufficiency when he included it in his law of exercise, pointing out that the S-R connections do not form because they are repeatedly activated, but because they are followed by a reward or, at any rate, by some "satisfaction" for the organism, according to a law he saw as fundamental and more far-reaching than the law of practice, i.e. the law of effect. 7 As for Meyer, he dismissed the law of practice as trivial, preferring what he defined the law of double stimulation. This law, he believed, held the key postulate for the theory of habit formation which he formulated during the period we are concerned with here. Briefly, this theory laid the emphasis on the simultaneous, or almost simultaneous, occurrence of stimulation rather than on their frequency. The cause of modified resistance in the nerve connections, and thus of a change in the nervous flow (the cause of learning, in a word), lies in the fact that two stimuli were present simultaneously or in close succession, and that the nervous flow triggered by one was able to attract the nervous flow triggered by the other. The final result was that one stimulus substituted another, producing the same response. 8 To account for this phenomenon Meyer (1911) offered the analogy of the drainage pump, which illustrated the "attraction" that a stronger nervous stream exerts on a weaker one.

Both Thorndike and Meyer recognised that repeated, frequent stimulation, or "use", obviously played a positive part in reinforcing the connections once they had been formed (according to the different postulates of the two theories), but both rejected the idea that it could account for the formation of new connections. It would seem that this distinction between the formation of new connections and their reinforcement through use was not considered by Bent Russell in his statement of the standard theory. Perhaps it is not by chance that Bent Russell 's references to Meyer ' s Fundamental laws concern the issues regarding the frequency of the

7Thorndike's statement of these two laws in Animal Intelligence is the following: "'The law of effect: Of several responses made to the same situation, those which are accompanied or closely followed by satisfaction to the animal will, other things being equal, be more firmly connected with the situation, so that, when it recurs, they will be more likely to recur; those which are accompanied or ciosely followed by discomfort to the animal will, other things being equal, have their connections with that situation weakened, so that, when it recurs, they will be less likely to occur. The greater the satisfaction or discomfort, the greater the strengthening or weakening of the bond. The law of exercise: Any response to a situation will, other things being equal, be more strongly connected with the situation in proportion to the number of times it has been connected with that situation and to the average vigour and duration of the connection" (Thorndike, 1911: 244). See Postman (1947) for the development of Thorndike's original treatment of the laws of exercise and effect.

Sin fact, this is a possible statement of Meyer 's law of double stimulation, which he distinguishes from the law of practice, or as he puts i t - perhaps with a critical reference to Watson 's position (see footnote 9 below) - the laws of duration, frequency and recency (Meyer, 1934). At the time in which Meyer gave early statements of the law of double stimulation (Meyer 1908, 1911), he was not aware of Pavlov 's conditioned reflex theory, but his own theory amounted to much the same thing (Meyer, 1934: 181).

318 R. Cordeschi

stimulation and the references to Thorndike's Animal Intelligence concern the law of exercise more than the taw of effect (Russell, 1913:16-17). 9

3. Artificial Nervous Systems: Bent Russell's Machine

Let us see how Bent Russell conceived his machine and how he justified his simulative methodology. His general approach is a seminal one, which was later to find favour among psychologists and behavioural scientists.

Bent Russell began with a concise exposition of the postulates of what I called above the standard theory of nervous conduction. He then described the design of a hydraulic machine which "embodied", as he said, the neurological hypotheses of the standard theory, before going on to "compare the results obtained with the machine with those given by live nervous connections" (Russell, 1913: 15). This comparison allowed him to conclude that it was actually possible to use a mechanical device to simulate the "essential elements" of the neurological phenomena which, according to the standard theory, accompanied inhibition, habit formation and some kinds of associative learning. Russell's paper was concluded by a brief analysis of some differences between the behaviour of the machine and that of the organic nervous system. It was believed that this machine analogy "with some modification" (p. 34) could simulate other, presumable more complex, kinds of learning responses.

It is interesting to examine in further detail how Bent Russell intended to use his machine to simulate those features (or "essential elements") of the nervous system stated by the standard theory, in order to discover the original, pioneering aspects of his mechanical analogies vis-it-vis older, more conventional proposals. Bent Russell's machine worked by compressed air or hydraulic pressure. What could be called the fundamental unit of the machine is the transmitter, the original diagram of which is reproduced in Fig. 1.

In fact the transmitter is a valve with an inlet "pressure pipe", in which the pressure of a flux of air or water is assumed to remain constant, and an outlet "meter pipe" which discharges such a flux. The maximum opening of the transmitter, owing to the sequence of strokes that activate it, is variable with time, and accordingly the intensity of the flux also is variable with time. It seems that we already have an important novelty with respect to the old hydraulic analogies of the nervous system (e.g. those pointed out by drainage theory itself): the simulation of the effect of events ranging over time intervals by an actually working mechanical device. The transmitter has at least a kind of "memory": it modifies its own behaviour according to its previous functioning or "history". The time interval between one stroke and the next of the transmitter (in fact, of his spur valve 6 in Fig. 1) is of decisive importance for the analogy: it embodies the function of the time lag in determining

9In the papers mentioned that Bent Russell published subsequently in the Psychological Review, he took no further reference from Meyer and Thorndike, but took reference mainly from John Watson' s Behavior, where the latter had stated the law offrequency and recency. This was the one law of learning Watson placed at the basis of both the formation and the fixation of nervous connections, as also of all forms of learning. Essentially, such a law was the law of exercise - once the latter "is stripped of its unnecessary implications of bonds and connections", as Watson concluded (Watson, 1914: 256n). In his more recent papers Bent Russell summed up explicitly the standard theory with this one law (see Russell, 1916, 1917a, 1917b).

Early-Connectionism Machines 319

?

Fig. 1. Bent Russell's hydraulic transmitter. The transmitter consists of three valves so arranged that the spur valve 6 slides over the other two valves: 11, the ratchet valve, and 12, the lag valve. The combined sliding movement of the three valves closes and opens to a variable degree the pressure ports 4 and 5 and the issue ports 14 and 13 to the flux, which is assumed constant, coming from the pressure pipe 2. The flux is thus transmitted through the ports to the outlet meter pipe 15. In the figure the transmitter is shown with the spur valve 6 fully rightward, which is the fully activated position. The input consists of inward (i.e. rightward) strokes when the transmitter is quiescent (i.e. with the spur valve 6 fully or partially leftward). Without going into more details, owing to the way the valves are connected, we have that the greater the frequency of the successive strokes on the spur valve 6, the further the ratchet valve 11 is pushed leftward by the reciprocating motion of the spur valve 6 (due to the rapid succession of input strokes) and the action of the pawl 8, and the wider the issue port 13 is opened to the flow coming from pressure port 5. Moreover, the further the ratchet valve 11 is pushed forward, the greater the leftward shift of the lag valve 12, and the wider the issue port 14 is opened to the flow coming from pressure port 4. On the other and, owing to the way the valves are connected, if the time interval between one stroke and another of valve 6 increases, issue port 13 will automatically and gradually tend to close, so reducing the flow. In conclusion, a rapid succession of strokes of the spur valve 6 will cause an increased opening of the transmitter and thus a cumulative increase in the flux discharged through meter pipe 15. On the other hand, if the strokes are spaced out there will be an overall reduction in the area of the opening and thus a reduction in theflux discharged through meter pipe 15 (after Russell, 1913).

the va r i a t ion in the ne rvous r e sponse p red i c t ed by the s tandard theory. A c c o r d i n g to

the lat ter , we h a v e seen that ne rvous c o n d u c t i v i t y is i nc reased by m e a n s o f r epea ted

s t imulat ions occur r ing c lose toge ther in t ime and is decreased wi th longer rest per iods.

The ana logy b e t w e e n the func t ion ing of the nervous sys tem and that o f the mach ine

goes fu r the r than this. As a c o n s e q u e n c e o f the s tandard theory , cer ta in in i t ia l ly

l o w - c o n d u c t i v i t y ne rvous c o n n e c t i o n s could , by means o f r epea ted s t imula t ion , be

" f o r c e d " (as both M e y e r and T h o r n d i k e said) to t r ansmi t the impulse , so b e c o m i n g

o f h igh conduc t iv i t y . In the mach ine , this type o f e f fec t cou ld be s imula ted th rough

the ac t ion o f seve ra l t ransmi t te r s w o r k i n g in conce r t in a va r ie ty o f combina t ions .

T h e m e c h a n i c a l d e v i c e con t ro l l i ng the func t i on ing o f severa l c o n n e c t e d t ransmi t te rs

was ca l l ed by Ben t Russe l l the c o u p l i n g gang: bas ica l ly , this func t iona l ly r ep roduces

in the m a c h i n e the n e t w o r k o f n e r v o u s c o n n e c t i o n s w i t h bo th h i g h and l o w

conduc t iv i t y . Ins t ead o f f o l l o w i n g Ben t Rus se l l in his desc r ip t ion o f the dev ice , I

shal l p r o v i d e a func t iona l d i a g r a m that shows h o w it works in the s imu la t i on o f a

320 R. Cordeschi

very simple kind of learning he considers: the inhibition of a response. In Fig. 2 we have assumed that three transmitters, shown in the diagram as T1,

T2, T3, have been connected up in a net, with their respective meter pipes which ultimately discharge the flux into two separate collecting pipes - respectively originated from the meter pipe of T1, on one side, and from the meter pipes of T2, T3, on the other - in such a way that the responses R1 and R2 will result in being antagonistic (for the sake of simplicity the inlet pressure pipes of the transmitters are not shown in the figure). Initially high-conductivity meter pipes are indicated by heavy lines in the figure. We assumed that the flux gives rise to two different (in fact antagonistic) movements or responses R1 and R2. The transmitters are variously connected up by means of two "sensory terminals" (as Bent Russell calls them) S 1, $2. By acting on each of the latter certain transmitters are activated (by the inward strokes of their respective spur valves: see valve 6 in Fig. 1). For instance, by acting on S1, T1 and T2 are then activated simultaneously (but T3 is not). This means that acting on S1 gives rise to a flux that is stronger towards R1 than towards R2, and we have R1 as a response. On the other hand, acting on $2 gives rise to R2.

It will be recalled that a transmitter discharges a greater flux the greater the frequency of activation. Now suppose that we frequently act, in rapid succession, on S 1 and $2. As indicated in the net of connections in Fig. 2, both S 1 and $2 act on the transmitter T2, which will therefore be activated at closely spaced intervals and discharge an increasing flux towards R2. This will produce the response R2. After a "training" period consisting of repeated activation of S1 and $2, it will be sufficient to act on S1 alone in order to produce R2. As Bent Russell concludes, "the result (of the working in concert of the transmitters depends) largely upon what might be called the "experience" of the transmitters in the combination" (p.26).

S2 S1

i [ - ~ - . - - ~ R1

w

R2

Fig. 2. A network of Bent Russell's transmitters simulating the inhibition of a response (see text).

By the net of transmitters of Fig. 1 Bent Russell believed to simulate the neurological correlate of the phenomenon of the inhibition of a response (in this case R1). He gives the description of an experience commonly exemplified through

Early-Connectionism Machines 321

the burned child: the sight of the fire (S 1) excites a forward movement (R 1); the resulting

effect of the burn ($2) excites a backward movement (R2). After repeated experiences

the sight of the fire no longer excites a forward movement (Russell, 1913: 18-19). l~

Bent Russell introduced his project with some general remarks on the possibility that engineering might be combined with physiology and psychology in research on the nervous system. Looking back over the age of cybernetics, we detect a prophetic ring in

his words, written thirty years before its advent.

It is thought that the engineering profession has not contributed greatly to the study of the nervous system, at least since Herbert Spencer, an engineer, wrote his book on psychology. As the cooperation of workers in different fields of knowledge is necessary in these days of specialists it may be argued that engineers can consistently join in the consideration of a subject of such importance to man. (Russell, 1913: 21)

Bent Russell 's proposal of such a multidisciplinary study of the nervous system, "in

these days of specialists", was bound not to be accepted, and he himself soon abandoned

his project of the simulative machine. 11 It was Meyer who discussed Bent Russell 's

machine in the context of various mechanical analogies of the nervous system current at

the time, but mostly as an argument in his own controversy with McDougall. The latter,

in his Body and Mind, had censured as "automaton theory" any mechanistic interpretation

of organism behaviour (McDougall, 1911). Meyer 's reaction was that of describing McDougall as the champion of the vitalistic or "ghost theory" of organism behaviour

(Meyer, 1912: 367), then offering the existence of Bent Russell 's "mechanical organism"

as an anti-vitalistic argument:

We have here a demonstration of the possibility of an "organism", capable of learning and forgetting, which obeys no ghost whatsoever, but only the laws of mechanics ... If it is proved that a mechanical organism can learn and forget without the interaction of a ghost, we have no right to assert that a biological organism can not. (Meyer, 1913: 559)

In fact Meyer 's interpretation of Bent Russell 's machine is one of the issues of the simulative methodology I pointed out above: if a machine is able to modify its own

behaviour as an organism does - the argument runs - then there is no need to invoke

non-physical (in fact non-mechanical) principles to account for the ability, typical of the organism, to adapt itself to the environment and to learn.

4. Stephens' Machine: A Simple Artefact Embodying Thorndike's Law of Effect

This very argument was the starting point of another machine simulat ion project, this time more consistent with Thorndike 's theory of learning. As we saw, Thorndike,

1~ fact the functioning of the machine suggests that the initially weak S1-R2 "connection" gradually strengthens through repeated activation until S 1 itself suffices in bringing about R2. Thus in the machine a backward movement prevails on a forward one. Certain limitations of Bent Russell's device are pointed out in Cordeschi (1998). These might be at the core of author's next critical reflection on learning machines (see Russell, 1917b: 421).

HBefore cybernetics, such a proposal was taken seriously within the above-mentioned robot approach of HuI1. This was perhaps the most lively pre-cybemetic project of a multidisciplinary study of the behaviour of organisms. Laurence Smith mentions Hull's intention, in 1930, of drowning the attention of physicists, chemists, physiologists and engineers in his own project ofbchaviour simulation machines (Smith, 1986: 162).

322 R. Cordeschi

in his law of effect, seemed to stress the introspective aspects of the effect itself: it was the pleasure or satisfaction of the response that causes the strengthening or stamping in of the connections, and it was the displeasure or discomfort that causes their weakening or stamping out. This motivational or introspective terminology was criticised by behaviourist and mechanistic psychologists. Notwithstanding Thorndike's insistence on the contrary, the role of satisfaction in the selection of the successful response and of discomfort in the elimination of the unsuccessful one might be perceived as a paradox: it seemed as if there were apsychic cause leading to a physical effect.

It was in this context that Stephens, in 1929, stated his proposal of a "strictly mechanistic interpretation" of Thorndike's position. Such an interpretation should eliminate motivational terms such as "satisfaction" and "discomfort" from the statement of the law of effect, accepting only the neurological terms proposed by Thorndike himself: those of the "strengthening" or "weakening" of bonds or connections between neurons. Still, according to Stephens, Thorndike, when using these neurological terms, did not seem "to regard the neural condition in itself as the determiner" of the learning process (Stephens, 1929: 429-430).

Now Stephens' strictly mechanistic interpretation of the law of effect was based on more or less the same facts Thorndike observed, for example, in his famous experiments on cats which learn by trial and error how to escape from a puzzle box: to successive presentations of a constant stimulus situation, the animal reacts with a variety of random responses, until by chance one is selected which produces the required effect (the opening of the puzzle box and the escaping outside it). This successful response is repeated on further presentations of the same stimulus situation. 12 Stephens' hypothesis is that the selective agency of the successful response is not a "mental" state of pleasure or satisfaction, but solely the effect of the response itself. Such an effect leads by "retroaction", as he puts it, to a specific neural mechanism, which is responsible of the resistance variation of the connection that gave rise to the successful response. 13 Anyway, it is the exact nature of such a neural mechanism (the objective correlation of Thorndike' s motivational terminology) that should be clarified: which is the best hypothesis, among those pointed out by psychologists in order to explain how nervous connection resistance does change?

Stephens' way of answering this question is an interesting case study of simulative methodology at a germinal stage. He starts from the statement that, in many of the current explanations, the phenomenon of learning was analysed into "the elements of conduction bodies, connections and variable resistances", i.e. "independently of protoplasm". He then undertakes to implement a mechanical synthesis of those elements, so as to produce a kind of "non-protoplasmatic learning", in fact a "synthetic test" of the (mechanistic) neurological hypotheses, that might explain the resistance variation of nervous connections when learning occurs. A "learning machine", as he

12See B oakes (1984) for an introduction to animal psychology of Thorndike's time.

~3Tbe term "retroaction" was used by Stephens in a further paper (Stephens, 1931: t43), in which a second learning machine is also sketched. He believed this notion to be similar to Leonard Troland's "retroflex action" (Troland, 1928). This latter notion was mentioned by Miller et al. (1960: 44) among those involving a kind of (cybernetic) feedback loop.

Early-Connectionism Machines 323

puts it, might thus represent, if not a verification, at least an argument in favour of the "learning theory" the machine embodies: since the theory was used as the basis for the construction of the machine, the latter can be taken as a test of the theory itself.

It is worthwhile re-reading the passage below. As a pioneer in his work, Stephens seems surprised to discover that the mechanist ic metaphor suggested a possible new conception of the relationship between psychology and neurology, and a novel possibility of testing a psychological theory - a kind of testing that will become popular among psychologis ts through the development of machine simulation of behaviour.

The above conception occmTed to me as a means of evaluating those expositions of learning couched in neurological terms. This test of possible synthesis had, at first, only theoretical interest. To my surprise, however, I found that some of the analyses presented startling possibilities of mechanical synthesis . . . . I have given the description of the machine first and then tried to point out the principles of its operation. ... I have then tried to indicate the theories from which the present conception was generated, through the medium of synthetic verification, and to point out the modifications .... I have tried to use no explanation of animate learning which could not be considered to work in a machine. I have considered protoplasmatic organisms as very complicated machines and nothing further. (Stephens, 1929: 422, 423)

"Very compl ica ted machines" - but what kind of machines ? Two mechanisms should be present in order for learning to occur according to Stephens ' explanat ion o f the law of effect. The first must be capable o f d iscr iminat ing be tween (at least) two condit ions; the second must keep on changing S - R connect ions until one rather than the other of those condi t ions appears, and to cease the changing when the successful connect ion is selected. Might it be poss ib le to construct a working machine that would real ise, so to speak, in the metal this k ind of select ive learning mechanism? In Stephens ' words, the "essent ia l e lements ' o f this concept ion of learning might be "synthet ized into a machine which modif ies its own responses to achieve a given end"? (Stephens, 1931: 152).

The machine Stephen des igned in order to give a response to this quest ion is the s imple e lec t romechanica l device in Fig. 3. It can " learn" to react to a dis turbance f rom the outs ide environment , which consists of press ing a f inger on the machine. The design o f the machine is such that any of its - in fact two - movements which does not hit the f inger will be changed; on the other hand, the machine wil l a lways make that response which proved successful the last time it was stimulated by pressing the f inger on it (Stephens, 1929: 426).

Accord ing to Stephens, his machine is a quite s imple inorganic real isa t ion of the two above-men t ioned neural mechanisms: the one d iscr iminat ing among different s t imulus si tuations, the other changing S - R connect ions until a response is achieved that, as he puts it in his further paper , has a "survival value" for the organism (Stephens, 1931: 143), and f rom that t ime on not changing the re la t ive connection. Fur thermore , the mach ine embodies or synthes ises the neuro logica l hypothes is Stephens be l ieved to expla in the variation of resistance of nervous connections: there is an increase in the re la t ive res is tance of all connect ions not achieving a certain effect (Stephens, 1929:429).14

Therefore the machine has a certain abi l i ty to modi fy its own internal organisat ion on the basis of its own previous "his tory". Nevertheless , its per formance remains

~4In his opinion this view, which differed from that emphasising frequency and recency, was put forward at the time in order to explain the phenomenon of nervous plasticity (Stephens, 1929: 429).

324 R. Cordeschi

1

Fig.3. Stephens' simple electromechanical model. Themachineiscompoundedofacagewhichconsistsof a quadrilateral of insulated wire, in which insulation is removed only at points A and C. There is a clear electrical circuit from any part of the cage to any other part. Briefly, when the cage is pressed with a finger at either of the two points A and C, circuits are closed so that the cube B, with its system of bars, will move towards A or C at random, until the finger pressing on A or C is hit by B2 or B 1 respectively. From that time on, the device is so designed that B will continue to move towards the pressing finger, until the finger itself is eventually placed at the other of the two points. As Stephens puts it, the movement of B might recall the action of a blindfold man trying to touch a hand that is placed near him. Note that the stimulus (the pressing by the finger) has potential connection with two responses, the resistance of the two connections being variable. The machine is so designed that if a response produces one effect (hits the finger) the resistance of the connection- the connection which caused the successful response - will not be increased; if, on the other hand, that response produces another effect (does not hit the finger, thus hitting the cage) the resistance of that connection - the connection which caused the failing response - will be increased (after Stephens, 1929).

qu i te c rude : the m a c h i n e is ab le to r e spond wi th on ly two responses . O f cou r se

S tephens was aware o f the l imi ta t ions o f his mach ine : i f we o b s e r v e its behav iou r ,

" i t s eems as i f we h a v e a typ ica l e x a m p l e o f an ima te l e a rn ing" ; but it mus t be

admi t t ed that " i t is on ly the p r inc ip l e o f l ea rn ing that has been here i l lus t ra ted" by

the mach ine . N o t w i t h s t a n d i n g , he b e l i e v e d that a m o r e c o m p l e x mach ine , w h i c h

w o u l d r e s pond to any one o f any g i v e n n u m b e r o f s t imul i by any one o f any n u m b e r

o f responses , " w o u l d i n v o l v e no n e w p r i n c i p l e and is qu i te poss ib le for any one wi th

the funds at his d i sposa l " (p.425, m y I ta l ics) .

Thus , at the r ise o f c o n n e c t i o n i s m , B e n t Russe l l and far m o r e exp l i c i t ly S tephens

p o in t ed out s o m e of the po in t s I pu t fo rward as the c o r e o f the b e h a v i o u r s imu la t i on

m e t h o d o l o g y o f the Cul tu re o f the Ar t i f ic ia l . It was stated that a so -ca l l ed " i n o r g a n i c "

or " n o n - p r o t o p l a s m a t i c " mach ine , w h e n it b e h a v e s as p red ic t ed by the b e h a v i o u r a l

theory it e m b o d i e s , m i g h t be cons ide r ed as a success fu l t es t for such a theory . T h e

"es sen t i a l e l e m e n t s " (as bo th B e n t Russe l l and S tephens ca l l them) that the l ea rn ing

m a c h i n e shares wi th the l ea rn ing o r g a n i s m revea l a c o m m o n f u n c t i o n a l o r g a n i s a t i o n

b e t w e e n them, and this jus t i f i e s the m e c h a n i s t i c exp l ana t i on o f learn ing . In fact on ly

Early-Connectionism Machines 325

simple kinds of learning can be simulated by then existing machines; it is hoped anyway that more complex kinds of learning might be simulated through the progress in the construction of simulative machines. In any case, the very existence of those machines did give at least an early argument in favour of the sufficiency of the principles involved ("no new principle", in Stephens' words, must be called upon), so that learning might not be a distinctive (non-physical) feature of living organisms.

These simple machines seemed therefore to suggest for the first time that learning has not to be viewed as the opposite of automatic or machine behaviour: on the contrary, learning is a particular kind of automatism, and choice can be mechanised. A "new" kind of machine - in fact, a machine capable of modifying its own internal organisation - provides new ideas on how to fill the gap between the organic and the inorganic worlds.

5. Many Years Later: Old and New Connectionism

In editing their collection of key texts in the history of connectionism James Anderson and Edward Rosenfeld included in it the introductory chapter from Hebb's Organization of Behavior, pointing out that the author presented his own theory as a "form of connectionism", before stating the law of learning that, in a quantitative statement, is known as "Hebb's rule" (Anderson and Rosenfeld, 1988). 15

One should remember that Hebb, in so doing, refers to the tradition going back to Thorndike, who, as we saw, had called his own approach "connectionistic" (and thus would have deserved a place in Anderson and Rosenfeld's book). In fact Hebb makes frequent reference to both Thorndike and Hull, pointing out both affinities and differences. Particularly as regards Thorndike, Hebb considers his own approach as an attempt "to fill in the gap" left open by Thorndike's approach to the physiological processes underlying his early statement of the law of effect, thus to the problem of motivation (Hebb, 1949: 177). This is the very question Stephens raised in 1929. But now Hebb rides on the back of twenty years of debate on the law of effect and the learning principles of connectionism, and he is therefore in a position of putting forward a proposal of his own to that question.

The general point in which his ideas differ from Thorndike's is summed up by him in the statement of "a theory of association ... in which an association may be between autonomous central process instead of between afferent and efferent processes" (Hebb, 1949: 177). Thus it was the emphasis on autonomous, central processes and S-S connections between different cortical sensory centres rather than simple, linear S-R connections that Hebb saw as the essential feature distinguishing his theory from the "older idea of neural transmission", e.g. Thorndike's or Hull 's (p. 11). In fact, when Hebb referred to his theory as a "form of connectionism" he qualified it as a variant of the telephone exchange kind

Is"When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased" (Hebb, 1949: 62).

326 R. Cordeschi

(Thorndike's famous metaphor), i.e. one in which central (reverberating) processes were taken into account, and not only S-R (linear) connections. Once Thorndike's early treatment of the law of effect was refused by Hebb, the learning process appeared to him as "a lasting change of facilitation between the activities of specific neural structures" (p. 180), a change of facilitation which was explained by his own association law.

Apparently, the association law or "rule" that made Hebb so well known among present-time connectionists was considered by him as the less novel part of his theory. On the contrary, it was the entire framework of neurophysiological hypotheses used by him ("cell assemblies", "phase sequences" and so on) that marks him out from early connectionists & la Thorndike (p. xix). 16 So much so that Hebb saw his position as intermediate ("halfway") between the "connectionists", the old supporters of the S-R connections, and the "configurationists", basically Lashley and the Gestalt psychologists, adamant supporters of the central processes (p. 58).

Some years later, Rosenblatt himself gave an evaluation of Hebb's theory of learning in which the novel hebbian neurophysiological hypotheses were pointed out (Rosenblatt, 1958: 385-386). He recognised that Hebb's work was "a source of inspiration" for his own approach. On the other hand, he pointed out what his own theory accomplished beyond what has already been done by Hebb. Basically, Hebb's neurological hypotheses had been put forward in a qualitative way: "Hebb ... has never actually achieved a model by which behavior ... can be predicted from the physiological system. His physiology is more a suggestion of the sort of organic substrate which might underlie behavior, and an attempt to show the plausibility of a bridge between biophysics and psychology" (p. 407). As a designer of a behaviour-based model, the well-known Perceptron, Rosenblatt believed that the latter actually was an attempt in completing such a bridge, thus revealing the common functional organisation that machines share with organisms and pointing out the possibility of a physical explanation of behaviour: a crucial issue of the simulative methodology that I am pointing out in this paper.

By the study of systems such as the Perceptron, it is hoped that those fundamental laws of organization which are common to all information handling systems, machines and men included, may eventually be understood. (Rosenblatt, 1958: 408)

Therefore, it was from the point of view of behaviour model theorist that Rosenblatt put forward the theoretical requisites of his model, i.e. "simplicity", "verifiability" and "generality or explanatory power" (p.405-407).

In fact Hebb's rule, in a quantitative statement concerning the neural plasticity or change of facilitation between neurons, was soon embodied in several neural nets it la Roesenblatt during the 1950s. 17 In more recent years, Hebb's rule was considered just one among several postulates explaining the change of facilitation between specific neural structures. For example, Hawkins et al. (1983), as a consequence of their well-known researches on a simple invertebrate such as Aplysia,

~See Amit (1988) on the relevance of these hebbian theoretical constructs for a theory of internal representations.

~See J.K. Hawkins' review, in which a recasting of (simple) Perceptron is given in accordance with Hebb0s rule (Hawkins, 1961: 38).

Early-Connectionism Machines 327

have classified different types of facilitation from sensory to motor neurons with regard to different cellular mechanisms for strengthening connections during the learning process, such as simple conditioning. Basically, one might distinguish between a non-hebbian or presynaptic facilitation and a hebbian or pre-postsynaptic facilitation. 18

Comparison between the working of a Bent Russell 's network of transmitters, embodying the standard theory, with a (much simplified) one-layer neural net c) la Rosenblatt embodying Hebb 's rule, might be enlightening on this point. In the former case (see Fig. 2) the initially low-conductivity S l - R 2 "connection" becomes a high- conductivity one through frequently repeated activations of both S 1 and $2, until S 1 alone suffices to bring about R2. In the latter case, suppose one has a one-layer neural net consisting of two association units A1 and A2. These receive stimuli from a sensory system, and connect them with two different response units R1 and R2 through connections whose weights can be modified according to Hebb 's rule. Now the story goes: initially the weight w2 of the connection A2-R2 is high while the weight wl of the connection A1-R2 is low; subsequently wl increases, because R2 is already active whenever A1 is activated. In other words, since A1 and R2 are active in close succession, the strength of the relative connection increases according to Hebb 's rule until the activation of A1 alone (i.e. without immediately successive activation of A2) gives rise to response R2.

Therefore, different hypotheses on synaptic facilitation of nervous connections are embodied in these two machines. Bent Russell 's early connectionist machine embodies, at least in a crude form, the hypothesis of the facilitation by summation of paired activity of two presynaptic sensory neurons: in the model of Fig. 2, paired activation of two sensory terminals S 1 and $2 produces greater flux through the transmitter T2, and therefore a "facilitation" which gives rise to the movement or response R2. On the other hand, the hypothesis embodied in our one-layer connectionist net is that ofpre-postsynaptic facilitation: in such a simple net, if the "presynaptic" sensory activation of A1 is immediately followed by that of A2, which causes the "postsynaptic" activation of R2, then the A1-R2 "synapse" will be strengthened. Note that a comparison between these two models is possible, despite the different contexts in which they were stated, because of the novelty I pointed out above in Bent Russell 's model: the presence of a time-sensitive mechanism, which embodies a kind of "plasticity" rule.

6. Conclusions

The statement of certain hypotheses on the neurological bases of learning thus paved the way to different connectionist models or machines. The electromechanical technology of early (pre-cybernetic) models was too crude to realise the hope of a relevant progress in modelling learning through machines, and those projects were

~SIn fact, according to the authors, there is another non-hebbian presynaptic facilitation mechanism beyond that discussed in the text below. This is the very cellular mechanism they discovered inAplysia. For another taxonomy of hebbian and non-hebbian rules see Churchland and Sejnowski (1992: 250-253).

328 R. Cordeschi

soon abandoned. In fact early success of connectionist neural nets in the mid-1950s was mainly a consequence of the development of computer science and early computer technology: for example, Rosenblatt 's Perceptrons were simulated on an IBM 704 computer (Rosenblatt, 1958: 403). Several limitations of these one-layer Perceptrons were pointed out afterwards by Marvin Minsky and Seymour Papert in their famous book (Minsky and Papert, 1969). What happened is now well known: further progress in computer technology gave rise, during the 1980s, to the computational resources needed to simulate on computers multi-layer nets with more powerful learning algorithms, such as the backpropagation rule (see Rumelhart et al., 1986), thus leading to advanced "new" connectionist models of learning. 19

To conclude, I shall point out briefly an example of this more recent story: the experiments on the neural net models of learning in Limax, another invertebrate. In my view, it is an enlightening example, at least as regards certain types of simple learning like those I was referring to in this paper. In fact the example shows how in this case model building converged toward better and better approximations of the behavioural phenomenon, and how these approximations were made possible by progress both in learning algorithms and in computational resources.

Tesauro (1990) reviews the experiments on Limax and shows how certain basic phenomena of classical conditioning (i.e. general features of first- and second-order conditioning) could be simulated as a simple one-layer net - one of the sort I described above. This can be viewed as a model that utilises local representations and Hebb 's rule as plasticity rule or learning algorithm for modifying the weights. 2~ A first improvement of this simple model led to the simulation of higher forms of conditioning and to good predictions, as regards the behaviour of organisms in general during this kind of learning. Anyway the model could not say in detail how conditioning exactly occurs in a particular organism such as Limax.

A more profitable improvement of the model became possible in the general framework of"new" connectionism of the mid-1980s, using distributed representations and a Hopfield net with a Hebb 's rule as learning algorithm (Gelperin et al., 1985). A further improvement of the model was then given by introducing multiple layers and hidden units, with the backpropagation rule as learning algorithm, in the "back- end" of the model (that dealing, in particular, with the generation of the appropriate motor responses). This latter model was proposed by Tesanro himself (Tesanro, 1986).

Comparison of the behaviour of these "new" connectionist models with that of the real animal during learning was actually profitable: the models suggest several behavioural experiments and provide new insights into the bases of conditioning in Limax. Tesauro points out the common theoretical requirements of these models, i.e. their "simplicity" and "generality", that lead to a grasping of the "essential ingredients",

19The role of Minsky and Papert's book in the demise of neural nets during the 1960s has been commonly overestimated, as much as the role of computational resources undervalued. As James McClelland puts it: "I don't believe it was [Minsky and Papert's book] per se which discouraged Perceptron research in the sixties .... A certain scale of computation is necessary before simulations show that neural networks can do some things better than conventional computers. The computing power available in the early sixties was totally insufficient for this" (mentioned by Crevier, 1993: 309).

2~ a localrepresentation each sensory stimulus is represented by a single neuron. In a distributed representation (see below) each sensory stimulus is represented by a pattern of activity in a population of neurons.

Early-Connectionism Machines 329

as he puts it, of the behavioural phenomenon (Tesauro, 1990: 75). Here we are again at the core of the methodology put forward by any simulation theorist. Anyway there is something more then this here. As Tesauro concludes, behaviour-based models are always a simplification of the phenomenon, but the interaction of modelling studies with cellular and behavioural studies of real animals leads this time both to the design of new experiments and to refinements and improvement in the design of the model itself, based on experimental data that cannot be explained by existing models. This is perhaps a major lesson for the simulation methodology.

A c k n o w l e d g e m e n t s

A partial draft of this paper was presented at the Third International Meeting on "The culture of the Artificial", Santa Sofia, Italy, 14-15 October 1994. I am grateful to Giuseppe Trautteur and to the late Marino Giannini for their suggestions and criticisms. I thank Massimo Negrotti for helpful discussions on the notion of the Culture of the Artificial. This work was partially supported by MURST 60% research funds and CNR (Progretto Finalizzato Robotica).

R e f e r e n c e s

Amit, D. (1995). The Hebbian Paradigm Reintegrated: Local Reverberations as Internal Representations, Behavioral and Brain Sciences. 18. 617-657.

Anderson, J.A. and Rosenfeld, E. (eds) (1988). Neurocomputing. MIT Press, Cambridge, MA. Boakes, R. (1984). From Darwin to Behaviorism. Cambridge University Press, Cambridge, UK. Boring, E.G. (1946). Mind and Mechanism, American Journal of Psychology. 59. 173-192. Churchland, P.S. and Sejnowski, TJ. (1992). The Computational Brain. MIT Press, Cambridge, MA. Cordeschi, R. (1991). The Discovery of the Artificial: Some Protocybernetic Developments 1930-1940, AI

& Society. 5. 218-238. Cordeschi, R. (1998). La scoperta dell' artificiale. Psicologia, neurologia e macchine intorno alla cibernetica.

Dunod-France, Milan. An enlarged edition is in preparation, Kluwer Academic Publishers, Dordrecht. Crevier, D. (1993). AI. The Tumultuous History of the Search for Artificial Intelligence. Basic Books, New York. Esper, E.A. (1966). Max Meyer: The Making of a Scientific Isolate, Journal for the History of the Behavioral

Sciences. 2. 107-131. Esper, E.A. (1967). Max Meyer in America, Journal for the History of the Behavioral Sciences. 3. 107-13 t. Gelperin, A., Hopfield, J.J. and Tank D.W. (1985). The Logic of Limax Learning. In A. Selverston (ed.)

Model Neural Networks and Behavior. Plenum, New York, 237-261. Hawkins, J.K. (1961). Self-Organizing systems: A Review and Commentary, Proceedings of the IRE. 49. 31-48. Hawkins, R.D., Abrams, T.W., Carew, T.J. and Kandel, E.R. (1983). A Cellnlar Mechanism of Classical

Conditioning in Aplysia: Activity-Dependent Amplification of Pre-synaptic Facilitation, Science. 219. 400-405.

Hebb, D.O. (1949). The Organization of Behavior. Wiley and Chapman, New York and London. Herrick, C.J. (1929). The Thinking Machine. University of Chicago Press, Chicago. Hilgard, E.R. (1956). Theories of Learning. Appleton, New York. Hull, C.L. (1943). The Principles of Behavior. Appleton-Century, New York. James, W. (1890). The Principles of Psychology. Holt, New York. (Reprinted 1950, Dover, New York.) Jonqich, J. (1968). The Sane Positivist: A Biography of Edward L. Thorndike. Wesleyan University Press,

Middletown, CT. Krueger, R.G. and Hull, C.L. (1931). An Electro-chemical Parallel to the Conditioned Reflex, Journal of

General Psychology. 5. 262-269. McDougall, W. (1905). Primer of Physiological Psychology. Dent, London. McDougall, W. (1911). Body and Mind. Methuen, London. Meyer, M. (1908). The nervous Correlate of Pleasantness and Unpleasantness, Psychological Review. 15.

292-322.

330 R. Cordeschi

Meyer, M. (1911 ). The Fundamental Laws of Human Behavior. Badger, Boston, MA. Meyer, M. (1912). The Present Status of the Problem of the Relation Between Mind and Body, Journal of

Philosophy. 9. 365-371. Meyer, M. (1913). The Comparative Value of Various Conceptions of Nervous Function Based on

Mechanical Analogies, American Journal of Psychology. 24. 555-563. Meyer, M. (1934). Frequency, Duration and Recency vs. Double Stimulation, Psychological Review. 41.

177-183. Miller, G.A., Galanter, E. and Pribram, K.H. (1960). Plans and the Structure of Behavior. Holt, Rinehart &

Winston, New York. Minsky, M.L. and Papert, S. (1969). Perceptrons. MIT Press, Cambridge, MA. (Reprinted 1988 with authors'

new Prologue and Epilogue.) Negrotti, M. (1999). The Theory of the Artificial. Intellect Books, Exeter, UK. O'Donnell, J.M. (1985). The Origins of Behavorism. New York University Press, New York. Postman, L. (1947). The History and Present Status of the Law of Effect, Psychological Bulletin. 44. 489-563. Rosenblatt, F. (1958). The Perceptron: A Probabilistic Model for Information Storage and Organization in

the Brain, Psychological Review. 65. 386-408. (Reprinted in Anderson and Rosenfeld, 1988.) Rumelhart, D.E., McClelland, J.L. and the PDP Research Group (1986). Parallel Distributed Processing:

Explorations in the Microstructure of Cognition, 2 vols. MIT Press, Cambridge, MA. Russell, S.B. (1913). A Practical Device to Simulate the Working of Nervous Discharges, Journal of Animal

Behavior. 3. 15-35. Russell, S.B. (1916). The Effect of High Resistance in Common Nerve Paths, Psychological Review. 23.

231-236. Russell, S.B. (1917a). Compound Substitution in Behavior, Psychological Review. 24. 62-73. Russell, S.B. (1917b). Advance Adaptation in Behavior, Psychological Review. 24. 413-425. Savage, T. and Cowie, R. (1992). Are Artificial Neural Nets as Smart as a Rat?, Network. 3. 47-59. Smith, L.D. (1986). Behaviorism and Logical Positivism: A Reassessment of the Alliance. Stanford

University Press, Stanford, CA. Spencer, H. (1890). The Principles of Psychology (3rd edn). Longman, London. Stephens, J.M. (1929). A Mechanical Explanation of the Law of Effect, American Journal of Psychology.

41.422-431. Stephens, J.M. (1931 ). Some Weaknesses in the Explanation of Habit Fixation as Conditioning, Psychological

Review. 38. 137-152. Stephens, J.M. (1967). The Process of Schooling: A Psychological Examination. Holt, Rinehart & Winston,

New York. Tesauro, G. (1986). Simple Neural Models of Classical Conditioning, Biological Cybernetics. 55. 187-200. Tesauro, G. (1990). Neural Models of Classical Conditioning: A Theoretical Viewpoint. In Hanson, S.J. and

Olson, C.R. (eds) Connectionist Modeling and Brain Function. MIT Press, Cambridge, MA, 74-104. Thorndike, E.L. (1911). Animal Intelligence. Macmillan, New York. Troland, L.T. (1928). The Fundamentals of Human Motivation. Van Nostrand, New York. Valentine, E.R. ( 1989). Neural Nets: From Hartley to Hebb and Hinton, Journal ofMa thematical Psychology.

33. 348-357. Walker, S.F. (1990). A Brief History of Connectionism and its Psychological Implications, AI & Society. 4.

17-38. Watson, J.B. (1914). Behavior. Holt, New York.

Correspondence and offprint requests to: Roberto Cordeschi, Department of Science and Communication, University of Salerno, 1-84084 Fisciano (SA), Italy. Email: [email protected]