Explicit memory measures, mainly shown in recall or recognition tasks, reflect the conscious retrieval of past experiences, while implicit memory refers to memory that does not require a conscious retrieving of episodic or conceptual traces (as occurs in priming tasks, categorization, habits and skills learning, classical conditioning, etc.; see Richardson-Klavhen and Bjork, 1988; Schacter, 1987, 1992; Schacter, Chiu and Ochner, 1993, for reviews).
A large number of studies have shown experimental dissociations between explicit and implicit memory tasks, and so giving consistency to that differentiation. Thus, a very conclusive result comes from the fact that the subjects who suffer from anterograde amnesia obtain very poor results in recall or recognition tasks but they get the same results as normal subjects in implicit tasks (Warrington and Weiskrantz, 1970; for a review, see Shimamura, 1986). Similar results are found in the case of schizophrenic patients (Schwartz, Rosse and Deutsch, 1993) or anaesthetized patients (Roorda-Hrdlickovà, Wolters, Bonke and Phaf, 1989). In thesame way, Graf and Mandler (1984) have discovered that the kind of processing carried out on the stimulus (perceptual vs semantic) affects the observed execution in explicit memory tasks, but it does not affect the performance observed in implicit memory tasks (see also, e.g., Bowers and Schacter, 1990; Graf and Ryan, 1990).
There are two general theories which try to explain these dichotomies. The first one, based mainly on a neurophysiological tradition, states that different systems or stores would explain such dissociations. Thus, Squire, Knowlton and Musen (1993) defend the idea that declarative memory is based on limbic/encephalic structures, more recent in a philogenetic sense than non-declarative memory, while other types of memory not related to limbic/encephalic structures (more ancient, philogenetically speaking) would confirm the execution in implicit memory tasks, tasks that reflect the way in which living beings unconsciously respond to the environment. We will call this approach the «multiple systems theory» (see also Schacter et al, 1993).
On the other hand, and from a more cognitive tradition, various authors support the idea that dissociations between both types of memory reflect the fact that different processes of codification and/or retrieval are involved in the access to the same memory trace (see, e.g., Graf and Mandler, 1984; Graf and Ryan, 1990; Roediger, 1990; Roediger, Srinivas and Weldon, 1989). That is why data-driven tasks are more affected by perceptual manipulation than conceptually-driven tasks, while the latter ones are more sensitive to semantic or conceptual elaborations. We will call this approach the «processing view theory» (see Roediger, 1990).
In order to bring forward evidence on such two basic theoretical approaches we conducted several simulations about two well-established experimental dissociations.
SIMULATION 1. Simulation of the effects of anterograde amnesia
in implicit and explicit memory tasks.
We shall refer to the fact that amnesic subjects obtain very poor results in recall or recognition tasks, but they perform at the same level as normal subjects in implicit tasks (Warrington and Weiskrantz, 1970).
In order to simulate such effect we shall begin by creating four different neural networks, whose basic structure appears in figure 1. All of these four networks share a series of characteristics: they consist of 3 layers back-propagation models (called, from bottom to top, input layer, hidden layer and output layer), interconnected through one-way, bottom-up connections with random initial values in the range ±0.10, which work with a sigmoidal activation function and a back-propagation learning rule. The back-propagation algorithm uses a learning rate of 0.30 (in connections that linked input units to hidden units) and 0.15 (in connections that linked hidden units to output units). The value of the momemtum term was 0.40 (see Ruiz, Pitarque, Dasí., and Algarabel, 1994, for more details about the back propagation networks).
The input layer was formed by 30 units grouped in 3
modules of 10 units each: the first one, or «contextual module», collects the
contextual information asociated with the learning or the retrieval of a certain
item. On the second, the «graphemic module» represents the graphemic information
corresponding to each item. And the third represents the semantic information
corresponding to each word, it is the «semantic module» (see figure 1). This
idea of subdividing the input level in different modules, already appears in
other works (see, e.g., Murre, 1992; Ratcliff, 1990; Schreiber, Rosset and Tiberghien,
1991). As it is known, codification and retrieval of an stimulus in recall and
recognition tasks, are not only determined by the physical or semantic features
of the item, but also the context in which it was acquired plays a key role
(see, e.g., Feustel, Shiffrin and Salasoo, 1983, Jacoby, 1983; Schreiber et
al, 1991). This phenomenon, mitially revealed in the «encoding specific principle»
(Tulving and Thompson, 1973), implies the need to include a module of input
units in our model that allows us to represent contextual information associated
to each item in a different way from other type of information (for example,
physical or semantic features associated with the stimulus). On the contrary,
in datadriven tasks (such as priming tasks, completion or identification tasks)
perceptual or graphemic information seems to be more relevant in the retrieval
of memory items than contextual information. This is the reason why our model
will also include a module of input units that represents this graphemic information.
The hidden layer was always formed by 10 units, forming
only one module only in the one-store or processing view proposal, or grouped
in 2 modules of 5 units in the two-store proposal (see figure 1).
The output layer was formed by four units connected
to all the units in the hidden laver, regardless of whether they are divided
into one or two modules.
In figure 1, the differences between the four networks are shown. To make it more simple, not all the connections have been represented in that figure, but arrows show a total interconnection among all the units that form the related modules. The ONE/NORM network represents the one-store proposal (processing view) for a normal subject. It can be seen that all the input units are associated with all the hidden units, so that the stored information could be represented in just a trace, as proposed by authors like Roediger (1990) or Graf and Mandler (1984). On the other hand, the TWO/NOR network represents the multiple-system memory approach in normal subjects, in which the explicit module receives inputs from the three inferior modules, while the implicit module is not connected to the contextual input module, since implicit memory, does not seem to be related with the context in which items are learned (see Squire et al, 1993).
As regards the simulation of amnesia in the context of a neural network, we have opted to identify the concept of amnesia with the removal or «freezing» of certain connections, in such a way that from the moment when the network is «damaged», those connections will stop modifying themselves as learning advances (see also, e.g., Murre, 1992; Wolters and Phaf.,1990). So, ONE/AMN and TWO/AMN networks represent the one-store and two-store proposals, respectively, for amnesic subjects. As shown by the ONE/AMN network, the connections that associate the units in the contextual module with units of the hidden layer have been deleted, while comparing the TWO/AMN network with the TWO/NORM network, the connections that associate the units of the contextual module with the units of the hidden layer of the explicit memory module have been also removed.
Given that the results obtained in a back-propagation model
can differ from one simulation to other, and since the weights of connections
are randomized at the beginning of each simulation, so being different (see
Kolen and Goel, 1991; Ratcliff, 1990), we carried out 10 different simulations
(with the same data) with each of the aforementioned networks. In this way,
each of there replications would simulate the answers given by 10 different
«artificial» subjects, being treated as such in the statistical analysis carried
out afterwards (see Murre, 1992).
Once the four types of network were created, we then carried
out the training process of the networks (or training phase). The task
to simulate consisted of a category learning task in which each network had
to learn to classify correctly 40 exemplars, each corresponding to one of four
different semantic categories (belonging 10 exemplars per category). In order
to do so, a file formed by 40 stimuli (exemplars) was created, each one formed
(a) a sequence of thirty 1 or 0 (input pattern), the
first 10 representing the contextual information associated to each stimulus,
the next 10 representing the graphemic information, and the rest representing
the semantic information associated to each item. The modules that represented
contextual and graphemic information were random sequences of 1 and 0, given
that graphemic and contextual features can vary from one stimulus to other.
Contrary to this, the semantic features that defined each exemplar were common
90% of the time within the 10 exemplar that formed each semantic category, while
there was no relationship among the four different categories.
(b) a sequence of four orthogonal vectors (formed by one 1
and three 0: 0001, 0010, 0100, 1000) that indicated the category each examplar
belonged to (target pattern).
During the training phase the network readapted the weights of its connections during 5000 processing cycles (limit at which we observed that the learning of networks was satisfactory).
After the learning phase, we carried out the test phase,
each model being measured in a recall as well as in an identification task.
The recall task was operationalized presenting the contextual module of the
input pattern previously learned to the network, and observing if it was able
to classify the examplar into the correct category. On the other hand, in the
identification task, the graphemic module of the input pattern previously learned
was presented to the network, also observing if the network categorized each
The execution of the network was measured (dependent variable)
calculating the square root of the quadratic mean error (from now onwards, RMS
error) made during the classification of the 40 examples. This value shows both
the root of the quadratic mean difference between the activation value in each
of the output units and the activation value that was expected in each case,
taking the average of the 40 patterns. That is to say,
where oip represents the activation of the output
unit i for the pattern p and tip represents the expected
activation for that unit for the same pattern (target pattern), s is
the number of output units (4 in our simulations) and q is the number
of input units (40 in our simulations).
In each network, and before the learning phase, the RMS error
was also measured, in both recall and identification tasks. The aim of this
was to measure the execution of networks before they learned anything, in order
to obtain a baseline (or control line) from which the results could be compared.
The average RMS error of all untrained networks was 0.50, with hardly anv variability
among them. In this way, every execution of a network which gives rise to a
RMS error below 0.50 would be an indication that the network recalled or identified
in some way the presented examplars.
Results and discussion
We analyzed separately the data corresponding to the one-store proposal vs corresponding to the two-store proposal. We began by making an ANOVA 2*2 subjects (normal vs amnesic; between subjects variable) by task (recall vs identification; within subjects variable) for the one-store networks. This analysis showed as significant the main effects of the variables task (F(1,18)=15.939, MSe=0.000125, p<0.0001) and subjects (F(1,18)= 76.480, MSe=0.000125, p<0.0001), as well as the interaction of both variables (F(1,18)=67.053, MSe=0.000125, p<0.0001). The means of this interaction for one-store networks appear in figure 2.
A simple effects test for the analysis of this significant interaction was carried out, and the result was that the differences in the means between amnesic and normal subjects in the recall task reached statistical significance (p<0.01), while in the identification task that difference was not statistically significant. This result coincides completely with the data found by authors such as Warrington and Weiskrantz (1970) in experimentation with humans.
Regarding the analysis carried out on the data coming from the two-store networks, the 2 *2 ANOVA (subjects: normal vs amnesic; task: recall vs identification) also showed significant the main effects of the variables task (F(1,18)=463.560, MSe=0.0000056, p<0.0001) and subjects (F(1.18)=64.147, MSe=0.0000056. p<0.0001), as well as the interaction between both variables (F(1,18)=20.927, MSe=0.0000056, p=0.0002). The means of this interaction for two-store networks appear in figure 3.
A simple effects test for the analysis of this interaction
showed that the differences in the means between amnesic and normal subjects
in the recall task and in the identification task reached the statistical significance
(p<0.01). The latter result, indicating that the performance of amnesic and
normal subjects differs significantly in the identification task, contrasts,
however, with the data found Warrington and Weiskrantz (1970).
Taking these results into account, the one-store proposal seems to be more suitable than the two-store proposal simulating the dissociative effects showed by normal vs amnesic patients in explicit and implicit memory tasks. The crucial point of this simulation is that it demonstrates that the damage in an specific structure of the network affects only to a certain kind of memory (the explicit one), allowing a normal execution in the other. Overall, the results found support the predictive capacity of a distributed connectionist model, in agreement with the results found by other connectionist models of human memory (see, e.g., Masson, 1991; McClelland and Rumelhart, 1985: Murre, 1992).
SIMULATION 2. Simulation of the effects of the type of processing
of stimuli on implicit and explicit memory tasks.
Graf and Mandler (1984) subjected two samples of subjects to two types of tasks either a semantic processing of words (in particular, each subject was asked to rate them on a 5 points scale depending on the level of concretion/abstraction, number of meanings, and pleasure or displeasure that they connoted), or a graphemic or perceptual processing (where subjects had to decide whether two particular words had letters in common, count the number of letters in each word that orthographically formed a closed space, for example as in Q, P, or D, or were formed by intersecting lines, as occurs with T, L, H, etc.). On confronting the performance of both samples in a word completion task (HOUS _ type) vs a recognition task, they saw that the semantic processing of information only affected the execution observed in the explicit task (improving the percentage of correct recognitions), but it did not affect the performance observed in the implicit task. These results are consistent, and have also been achieved in other types of tasks (see, e.g., Bowers and Schacter, 1990; Graf and Ryan, 1990; Pitarque, Algarabel and Meseguer, 1992).
The procedure of simulation took on three steps. First, we subjected eight ONE/NORM networks and eight TWO/NORM networks, similar to these used in the previous simulation, to the learning of the 40 previous examples corresponding to 10 different semantic categories, during 3000 cycles, so that these trained networks showed the knowledge that subjects have when they undertake experiments such as the one carried out by Graf and Mandler (1984). Secondly, half of the networks (each network representing an artificial subject) were made to process information semantically, while the other half processed information perceptually. Finally, their performance in a completion task (implicit) and in a recognition task (explicit) was measured.
With no doubt, the most complex part of this simulation was to operationalize what a semantic or perceptual processing of an stimulus means in the context of a neural network. We decided to operationalize the perceptual processing by presenting the graphemic module of each stimulus, not allowing the connections that associate the input units with hidden units to be modified, while the connections that associate hidden units with output units were able to benefit from that graphemic processing. This is because the perceptual processing of an stimulus seems to be an automatic bottom-up process not related with the creation of new associations between the context and the item (see e.g. Graf and Mandler, 1984). By other hand, to operationalize the semantic processing we decided that networks processed the input patterns of the graphemic and semantic learned modules, giving a noise of 50% to the contextual module (because context can vary every time that a stimulus is processed), letting the connections modify by this semantic processing. Each stimulus was processed, perceptually or semantically, eight times (that is to say, the group of 40 examples was processed for 320 cycles).
The completion task was operationalized by presenting to the networks the first seven units of the graphemic module previously learned (see figure 1), in a similar way to a typical completion task. In these tasks, the first letters of a stimulus are presented (e.g., HOUS), and the subjects are required to complete the chain with the first letter they can think of. The recognition task was operationalized by presenting to the networks the input pattern learned corresponding to the three input modules (contextual, graphemic and semantic; figure 1).
As in the previous simulation, the RMS error was the dependent variable, reflecting the level of accuracy in the categorization of the previously learned examplars. The rest of the characteristics and parameters of this simulation remained similar to the simulation 1.
Results and discussion
As before, we analyzed separately the data corresponding to the one-store proposal vs corresponding to the two-store proposal. With regard to the one-store networks, an ANOVA 2*2, type of processing of stimuli (semantic vs perceptual; between subjects variable) by type of task (recognition vs completion; within subjects variable) showed as significant both the main effects of the type of processing (F(1,14)= 33.421, MSe= 1.795*10-4, p<0.0001) and type of task (F(1,14)= 5327.125, MSe= 1.676*10-4, p<0.0001), as well as their interaction (F(1,14)= 34.236, MSe=1.676*10-4, p<0.0001). The means of this interaction are represented in figure 4.
A simple effects test for the analysis of this significant
interaction showed that the type of processing of stimuli did not influence
significantly the results obtained in the completion task (F(1.28)= 0.014, MSe=
0.0001, p = 0.91), while it did on the explicit task (F(1,28)= 67.609, MSe=0.0001,
p < 0.001), in accordance with the results obtained by Graf and Mandler (1984)
and other laboratories.
By other hand, with regard to the twostore networks, the ANOVA 2*2 showed as significant both the main effects of the type of processing (F(1,14)= 19.249, MSe= 2.078*10-4, p=0.0006) and type of task (F(1,14)= 3993.438, MSe= 2.031*10-4, p<0.0001), but not the interaction (F(1,14)= 2.159, MSe=2.031*10-4, p=0.1638). The means of this interaction are represented in figure 5.
So our results, together with those found by other authors (see, e.g., Graf and Ryan, 1990; Schwartz et al., 1993), seem to confirm then the validity of the one-store processing theory to explain such kind of explicit/implicit dissociations.
The results of the simulations carried out regarding the theoretical controversy between one-store models vs two-store models on explicit/implicit memory support the proposal that just one store can explain the results found. The models that simulated normal and amnesic subjects assuming the existence of only one memory store, have reproduced the data in a large number of experimental investigations (see, e.g., Challis and Sidhu, 1993; Graf and Ryan, 1990; Schwartz et al, 1993). In considering the one-store model simulating the execution of amnesic subjects, such performance in explicit tasks does not seem to depend so much on the damage in a certain structure, but on the damage in certain connections. To be more specific, those who would not permit the conveying of information concerning the context or episodic characteristics related to the stimulus, at the time of the study. This is what seems to be confirmed by the first group of simulations. Besides, from a theoretical point of view the results are coherent with the fact that a one-store proposal is always more parsimonious than a theory which treats different sub-systems as the reason for the different experimental effects found.
On the other hand, the results of the second group of simulations
(one-store networks) that attempted to simulate the dissociative effects of
the type of processing to what the stimuli was subjected to, coincide with those
experimentally obtained by Graf and Mandler (1984) and Graf and Ryan (1990).
The data given by simulations show how the performance in the explicit task
is better compared to the implicit one when the stimuli have been previously
subjected to semantic processing. In this sense, it is a confirmation of the
fact that there has to differentiate between contextual information associated
to a stimulus and the physical and semantic features that describe it, in accordance
with the ideas pointed out by Jacoby (1983; also see Feustel et al, 1993) and
already implemented in other simulation models (Murre, 1992, Ratcliff, 1990;
Schreiber et al, 1991). In this way, while semantic and physical features associated
to the processing of an stimulus are usually constant, to a certain extent,
from a processing to another, contextual features vary. A connectionist model
is able to integrate into one trace all the information (contextual, graphemic,
and semantic) associated with the repeated processing of a stimulus, in such
a way that assertions like that semantic knowledge comes from the massive processing
of episodic traces (see e.g., Feustel et al, 1983) can be explained in a easy
way from such kind of networks. This is what occurs in the second simulation
shown aboye, where the change of context associated to the different processing
of each item does not seem to affect its recognition afterwards.
The explanation of this would be, according to the theory of
Graf and collaborators (Graf and Mandler, 1984; Graf and Ryan, 1990), that,
while in an implicit task there is just automatic activation, strengthening
and integration of pre-existent traces already consolidated through previous
learning, without new connections being created, in a recall or recognition
task there is an elaboration of new links that associate the context
in which an item appears with the memory trace that represents it. The terms
integration and elaboration have been identified with the synonym terms of familiarity,
activation or perceptual fluency, and memory for new associations, respectively
(see, e.g., Light et al, 1992). According to Graf and Ryan (1990), these two
different types of processes would be responsible for the results found in both
types of memory tasks. Our data totally confirm this idea.
To summarize, the simulations on the whole have revealed that a multilayer neural network based on the back-propagation learning algorithm (or generalized delta rule) is valid to simulate the experimental evidence concerning the dichotomy explicit memory/implicit memory. Morever, it has been shown that it is not necessary to postulate two different stores to explain the dissociations between explicit and implicit tasks. These results are in accordance with those presented in other areas of Psychology where similar connectionist models have been applied to fields as diverse as the conversion of written characters in voice (e.g., Seidenberg and McClelland 1989), category learning (Kruschke, 1992), implicit learning of events structure (Cleeremans and McClelland, 1991), language learning in children (Plunkett and Marchman, 1991); or identification of faces depending on the context (Schreiber et al, 1991), among others. Negative results with back-propagation networks, such as those found by Ratcliff (1990), seem to be due to formal defects in the training of these models rather than in its incapacity to simulate the overall functioning of human memory.
Future research should validate the model here proposed, using
other experimental dissociations regarding the controversy of explicit memory
vs implicit memory.
Bowers, J.S. and Schacter, D.L. (1990). Implicit memory and
test awareness. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 16; 404-416.
Challis, B.H. and Sidhu, R. (1993). Dissociative effect of
massed repetition on implicit and explicit measures of memory. Journal of
Experimental Psychology: Learning, Memory, arad Cognition, 19, 115-127.
Cleeremans, A. and McClelland, J.L. (1991). Learning the structure
of event sequences. Journal of Experimental Psychology General, 120,
Feustel, T.C., Shiffrin, R.M. and Salasoo, A. (1983). Episodic
and lexical contributions to the repetition effect in word identification. Journal
of Experimental Psychology: General, 112, 309-346.
Graf, P. and Mandler, G. (1984). Activation makes words more
accesible, but not necessarily more retrievable. Journal of Verbal Learning
and Verbal Behavior, 23, 55-3-568.
Graf, P and Ryan, L. (1990). Transfer-Appropiate processing
for implicit and explicit memory. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 16, 978-992.
Jacoby, L.L. (1983). Perceptual enhancement: Persistent effects
of an experience. Journal of Experimental Psychology: Learning, Memory; and
Cognition, 9, 21-38.
Kolen, J.F. and Goel, A.K. (1991). Learning in parallel distributed
processing networks: Computational complexity and information content. IEEE
Transactions on Systems, Man and Cybernetics. 21, 359-367.
Kruschke, J.K. (1992). ALCOVE: An exemplar-based connectionist
model of category learning. Psychological Review, 99, 22-44.
Light, L.L., La Voie, D., Valencia-Laver, D., Owens, S.A.A.
and Mead, G. (1992). Direct and indirect measures of memory for modality in
young and older adults. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 18, 1284-1297.
Masson, M.E.J. (1991). A distributed memory model of context
effects in word identification. En D. Besner and G.W. Humphreys (eds.). Basic
processes in reading: Visual word recognition. LEA, N.J.
McClelland, J.L. & Rumelhart, D.E. (1985). Distributed
memory and the representation of general and specific information. Journal
of Experimental Psychology General, 114, 159-188.
Murre, J.M.J. (1992). Learning and categorization in modular
neural networks. Harvester Wheatsheaf. N.Y.
Pitarque, A., Algarabel, S. and Meseguer, E. (1992). Degree
of elaborative processing in two implicit and two explicit memory tasks. Bulletin
of the Psychonomic Society, 30, 217-220.
Plunkett, K. and Marchman, V. (1991). U-shaped and frecuency
effects in a multi-laye-red perceptron. Cognition, 38, 43-102.
Ratcliff, R. (1990). Connectionist models of recognition memory:
Constraints imposed by learning and forgetting functions. Psychological Review,
Richardson-Klavhen, A., and Bjork, R.A. (1988). Measures of
memory. Annual Review of Psychology; 39, 475-543.
Roediger III, H.L. (1990). Implicit memory: Retention without
remembering. American Psychologist, 45, 1043-1056.
Roediger, H.L., Srinivas, K., and Weldon, M.S. (1989). Dissociations
between implicit measures of retention. En S. Lewandosky, J.C. Dunn and K, Kirsner.
(eds.). Implicit memory: Theoretical issues. Hillsdale, N.J.: Erlbaum.
Roorda-Hrdlicková, V., Wolters, G., Bonke, B. and Phaf,
R.H. (1989). Unconscious perception during general anaesthesia demonstrated
by an implicit memory task. En B, Bonke, W, Fitch and K, Millar (eds.), Memory
and awareness in Anaesthesia. Amsterdam: Swets & Zeitlinger.
Ruiz, J.C., Pitarque. A., Dasí, C. and Algarabel, S.
(1994): Simulaciones en una red conexionista de propagación hacia atrás.
Psicológica, 15, 121-140.
Schacter, D.L. (1987). Implicit memory: History and current
status, Journal of Experimental Psychology: Learning, Memory and Cognition,
Schacter, D.L. (1992). Understying implicit memory: A cognitive
neuroscience account. American Psychologist, 47, 559-569.
Schacter, D.L., Chiu, C.T. and Ochner, K.N. (1993). Implicit
memory: A selective review. Annual Review of Neurosciences, 16, 159-182.
Schreiber, A., Rosset, S, and Tiberghien, G. (1991). FACENET:
A connectionist model of face identification in context. European Journal
of Cognitive Psychology, 3,177-198.
Schwartz, B.L., Rosse, R.B. and Deutsch, S.I. (1993). Limits
of the processing view in accountine for dissociations among memory measures
in a clinical population. Memory and Cognition, 21, 63-72.
Seidenberg, M.S. and McClelland, J.L. (1989). A distributed,
developmental model of word recognition and naming. Psychological Review,
Shimamura, A.P. (1986). Priming effects in amnesia: Evidence
for a dissociable memory function. Quarterly Journal of Experimental Psychology,
Squire, L.R., Knowlton, B., and Musen, G. (1993). The structure
and organization of memory. Annual Review of Psychology, 44, 453-495.
Tulving, E. and Thompson, D.M. (1973). Encoding specificity
and retrieval processes in episodic memory. Psychological Review, 80,
Warrington, E.K. and Weiskrantz, L. (1970). Amnesic syndrome:
Consolidation or retrieval?. Nature, 228, 628-630.
Wolters, G. and Phaf, R.H. (1990). Implicit and explicit memory:
Implications for the symbol-manipulation versus connectionist controversy. Psychological
Research, 52, 137-144.