Social Cognition and Connectionism 1
Connectionist Exploration in Social Cognition
Frank Van Overwalle
Vrije Universiteit Brussel, Belgium
Christophe Labiouse
Belgian NFSR Research Fellow & University of Liège, Belgium
Robert French
Université de Liège, Belgium
This research was supported by Grant OZR423 of the Vrije Universiteit Brussel to Frank Van
Overwalle and Grant HPRN-CT-2000-00065 of the European Commission to Robert French.
Address for correspondence: Frank Van Overwalle, Department of Psychology, Vrije Universiteit
Brussel, Pleinlaan 2, B - 1050 Brussel, Belgium; or by e-mail: Frank.VanOverwalle@vub.ac.be.
Running Head: Connectionism and Social Cognition
[PUBSOCO]
5 February, 2016
Social Cognition and Connectionism 2
Connectionist Exploration in Social Cognition
Abstract
Major findings in social cognition are reviewed and modeled from a connectionist
perspective. These findings are in the areas of categorization and base-rate neglect, impression
formation, primacy and recency in impression formation, assimilation and contrast, increased recall
for inconsistent information, discounting in causal attribution, attitude formation, central and
heuristic processing, cognitive dissonance and the use of reasoning heuristics. The majority of
these phenomena are illustrated with well-known experiments, and simulated with an autoassociative network architecture with linear activation update and delta learning algorithm for
adjusting the connection weights. All of the phenomena considered were successfully reproduced
in the simulations. Moreover, the proposed model is shown to be consistent with algebraic models
of impression formation (Anderson, 1981), causal attribution (Cheng & Novick, 1992) and attitude
formation (Ajzen, 1991). The discussion centers on how the particular simulation specifications
may be used to develop novel hypotheses for testing the connectionist modeling approach and,
more generally, for improving and unifying theorizing in the field of social cognition.
Social Cognition and Connectionism 3
Connectionist modeling of theoretical and empirical data in social cognition has only
emerged during the last decade. This new approach arose from a certain dissatisfaction with
mainstream models and a growing concern for the limitations of these models. In particular, the
field suffered from a lack of theoretical integration. Inspired by the ever-increasing success of
connectionist models in cognitive psychology, a number of authors have turned to these models in
an attempt to provide a unified framework for social psychology research. In 1993 Read and
Marcus-Newhall wrote the first major article describing a connectionist model of causal reasoning.
Later, Smith (1996) forcefully argued for the application and the development of connectionist
ideas in social psychology.
Researchers have since made substantial progress in developing connectionist models of
diverse social psychological phenomena including person perception and group stereotyping
(Kunda & Thagard, 1996; Smith & DeCoster, 1998; Labiouse & French, 2001), causal attribution
(Van Overwalle, 1998, Read & Montoya, 1999), cognitive dissonance (Shultz & Lepper, 1996; Van
Overwalle & Jordens, 2001), group impression formation and change (Kashima, Woolcock, &
Kashima, 2000) and illusory correlation (Van Rooy & Van Overwalle, 2001c). However, there are
still large domains of social psychology that remain untouched by this new approach. Further,
insofar as each of the above articles focuses largely on a single domain of social psychology, the
field is still waiting for an overarching theoretical perspective.
In an attempt to provide an integrative account of the various fields within social
psychology, we will examine a number of mainstream findings in social cognition and will analyze
them from a common connectionist perspective. In the past, many findings in social cognition have
been explained by appeals to what often appear to be rather ad-hoc hypotheses and theories.
Moreover, various areas of the field, such as the “person perception”, the “impression formation”,
and the “intergroup relations” traditions, have unfortunately developed largely independently of
each other, despite the close conceptual connections of the topics. This has left the field with a
fragmentary theoretical basis. Our connectionist approach is an attempt to integrate some of these
theoretical areas into a more comprehensive whole. While we are aware that this is a major
undertaking, given the great power and flexibility of connectionist networks, as well as previous
successful attempts to model social data within this framework, we believe it is possible to use
these models to take a modest step towards the goal of unification of the field.
Social Cognition and Connectionism 4
Many mainstream processes and findings in social cognition can be explained within a
connectionist framework and, in many cases, better than the statistical or associative memory
models developed in the past. What are the main characteristics of the connectionist models that
accomplish this?
First, connectionist models exhibit emergent properties such as prototype extraction, pattern
completion, generalization, constraint satisfaction, and graceful degradation. (All of these are
extensively reviewed in Smith, 1996, and Rumelhart & McClelland, 1986). It is clear that these
characteristics are potentially useful for any account of social cognitive phenomena. In addition,
connectionist models assume that the development of internal representations and the processing of
these representations are done in parallel by simple and highly interconnected units, contrary to
traditional models where the processing is inherently sequential. As a result, these systems have no
need for a central executive, which eliminates the requirement of previous theories of explicit
(central) processing of relevant social information. Consequently, information can, in principle, be
processed in an implicit and automatic manner without recourse to explicit conscious reasoning.
This does not, of course, preclude people’s being aware of the outcome of these preconscious
processes.
Second, neural networks are not fixed models but are able to learn over time, usually by
means of a simple learning algorithm that progressively modifies the strength of the connections
between the units making up the network. The fact that most traditional models in social
psychology are incapable of learning is a significant restriction. Interestingly, the ability to learn
incrementally puts connectionist models in broad agreement with developmental and evolutionary
pressures.
Third, connectionist networks have a degree of neurological plausibility that is generally
absent in previous statistical approaches to information integration and storage (e.g., Anderson,
1981; Cheng & Novick, 1992; Fishbein & Ajzen, 1975). While it is true that connectionist models
are highly simplified versions of real neurological circuitry and processing, it is commonly assumed
that they reveal a number of emergent processing properties that real human brains also exhibit.
One of these emergent properties is the integration of long-term memory (i.e., connection weights),
short-term memory (i.e., internal activation) and outside information (i.e., external activation).
There is no clear separation between memory and processing as there is in traditional models. Even
Social Cognition and Connectionism 5
if biological constraints are not strictly adhered to in connectionist models of social cognition (i.e.,
persuasion, prejudice, …), concerns of the biological implementation of social cognitive
mechanisms have indeed started to emerge (Adolphs & Damasio, 2001; Allison, Puce & McCarthy,
2000; Ito & Cacioppo, 2001; Cacioppo, Berntson, Sheridan & McClintock, 2000; Phelps,
O’Connor, Cunningham, Funayama, Gatenby, Gore & Banaji, 2000) and parallel the increasing
attention paid to neurophysiological determinants of social behavior. Other emergent properties of
the connectionist approach will be explained in more depth in the next section.
This article is organized as follows: First, we will describe the proposed connectionist
model in some detail, giving the precise architecture, the general learning algorithm and the specific
details of how the model processes information. In addition, a number of other less well-known
emergent properties of this type of network will be discussed. We will then present a series of
simulations, using the same network architecture applied to a number of significantly different
phenomena. These phenomena involve categorization (base-rate neglect), impression formation
(primacy, recency and memory advantage for inconsistencies), assimilation and contrast (of traits
and exemplars), causal attribution (covariation, discounting and augmentation), attitudes (formation
and cognitive dissonance) and judgmental heuristics. We will also very briefly discuss related work
on group judgments (illusory correlation, ingroup versus outgroup differences).
Our review of empirical phenomena in the field is not meant to be exhaustive, but is rather
designed to illustrate how connectionist principles can be used to shed light on the processes
underlying social cognition. While the emphasis of the present article is on the use of a particular
connectionist model to explain a wide variety of phenomena in social cognition, previous
applications of connectionist modeling to social psychology (Smith & DeCoster, 1998; Read &
Montoya, 1999; Van Overwalle, 1998) are also mentioned. In addition, we will perform a
comparison of different models.
Finally, we will discuss the limitations of the proposed
connectionist approach and discuss areas where further theoretical developments are under way or
are needed. Ultimately, what we would like to accomplish in this paper is to create a greater
awareness that connectionist principles could potentially underlie diverse social cognitive
phenomena.
A Recurrent Model
Throughout this paper, we will use the same basic network model - namely, the recurrent
Social Cognition and Connectionism 6
auto-associator developed by McClelland and Rumelhart (1985). This model has already gained
some familiarity among social psychologists studying person and group impression (Smith &
DeCoster, 1998) and causal attribution (Read & Montoya, 1999). We decided to apply a single
basic model to emphasize the theoretical similarities that underlie a great variety of processes in
social cognition. In particular, we chose this model because it is capable of reproducing a wider
range of phenomena than other connectionist models, like feedforward networks (see Read &
Montoya, 1999) or constraint satisfaction models such as Thagard's ECHO (Van Overwalle, 1998).
The auto-associative network can be distinguished from other connectionist models on the
basis of its architecture (the elements of the model) and its learning algorithm (how information in
processed in the model). We will discuss these points in turn.
Architecture
The generic architecture of an auto-associative network is illustrated in Figure 1A. Its most
salient property is that all nodes are interconnected with all of the other nodes. Thus, all nodes send
out and receive activation.
Information Processing
In a recurrent network, processing information takes place in two phases. During the first
activation phase, each node in the network receives activation from external sources. Because the
nodes are interconnected, this activation is spread throughout the network in proportion to the
weights of the connections to the other nodes. The activation coming from the other nodes is called
the internal input (for each node, it is calculated by summing all activations arriving at that node).
This activation is further updated during a number of cycles through the network. Together with
the external input, this internal input determines the final pattern of activation of the nodes, which
reflects the short-term memory of the network. Typically, activations and weights have lower and
upper bounds of –1 and +1.
In the linear version of activation spreading in the auto-associator that we use here, the final
activation at each cycle is the linear sum of the external and internal input. In non-linear versions
used by other social researchers (Smith & DeCoster, 1998; Read & Montoya, 1999), the final
activation is determined by a non-linear combination of external and internal input. During our
simulations, however, we found that the linear version with a single internal updating cycle often
Social Cognition and Connectionism 7
reproduced the observed data better. Therefore, we used the linear variant of the auto-associator for
all the reported simulations. We will discuss later why the linear variant might have been more
efficient.
After the first activation phase, the recurrent model enters the second learning phase in
which the short-term activations are consolidated in long-term weight changes of the connections.
Basically, these weight changes are driven by the error between the internal input generated by the
network and the external input received from outside sources. This error is reduced in proportion to
the learning rate that determines how fast the network changes its weights (typically between .01
and .50).
This error reducing mechanism is known as the delta algorithm (McClelland &
Rumelhart, 1988).
Thus, when the network overestimates the external input of a node, this means that this node
received too much internal input from the other nodes through their connections. To adjust this, the
delta algorithm decreases the weights of these connections.
Conversely, when the network
underestimates the external input, this means that it received too little internal input and the weights
are increased. These weight changes allow the network to better approximate the external input.
Thus, the delta algorithm strives to match the internal predictions of the network as closely as
possible to the actual state of the external environment, and stores this information in the
connection weights.
Structural and Dynamic Connections
The social phenomena that we analyze can be subdivided in two main classes: structural
processes that focus on stable attributes of social actors and objects (categorization, impression
formation, generalization, assimilation and contrast) and dynamic processes that focus on causal
consequences (attributions and attitudes). Although these two processes may differ somewhat at
the surface level, we believe that they are very similar in their logical structure in that they both
reflect predictions from features to categories (e.g., from behaviors to trait categories and from
causes to events).
For structural processes, the prediction involves a category or attribute, for instance, the trait
of a person or the stereotype of a group. For dynamic processes, the prediction involves the
outcome of a cause or the behavior toward an attitude-object. To illustrate these two types of
predictions, imagine that when meeting an unfamiliar individual, a perceiver may wonder to what
Social Cognition and Connectionism 8
type of group the individual belongs (stereotyping or structural prediction) and what he or she might
do next (behavioral outcome or dynamic prediction).
Typically, the prediction goes from low-level features, exemplars, causes or attitude-objects
to higher-level abstractions such as categories and outcomes. These predictions are graphically
illustrated in Figure 1B—C for structural and dynamic relations. To visualize the direction of
prediction, we have drawn the (low-level) features that serve as input at the bottom layer of the
architecture, and the (high-level) predicted categories that serve as output at the top layer. We will
consistently use this direction of representation in the illustrations.
Basic Emergent Connectionist Principles
Before moving on to the social phenomena of interest, it is essential to briefly discuss the
basic principles or mechanisms that drive many of our simulations. These principles are the
emergent properties of the delta-learning algorithm and include acquisition, competition, and
diffusion. Some of these principles have already been documented in prior social connectionist
work (Van Overwalle, 1998; Van Overwalle & Van Rooy, 1998; 2001a, 2001b). However, because
they are essential for understanding our examples, we will describe these principles first and discuss
their application for social cognition in more detail later during the simulations.
Acquisition and Sample Size Effect
The acquisition principle involves sample size effects that have been documented in many
areas of social cognition. For instance, when receiving more supportive information, people tend to
hold more extreme impressions about other persons (Anderson, 1967, 1981), make more extreme
causal judgments (Baker, Berbier & Vallée-Tourangeau, 1989; Försterling, 1992; Shanks, 1985,
1987, 1995; Shanks, Lopez, Darby & Dickinson, 1996), make more polarized group decisions
(Fiedler, 1996; Ebbesen & Bowers, 1974), endorse more firmly an hypothesis (Fiedler, Walther &
Nickel, 1999), make more extreme predictions (Manis, Dovalina, Avis & Cardoze, 1980) and agree
more with persuasive messages (Eagly & Chaiken, 1993).
One of the most striking characteristics of connectionist models using the delta algorithm is
that learning is modeled as a gradual on-line process of adjusting existing knowledge to novel
information.
This characteristic has already been exploited in the earlier associative learning
models that preceded connectionism, such as the popular Rescorla-Wagner (1972) model of animal
Social Cognition and Connectionism 9
conditioning and human contingency judgments.
This model predicts that when a cue (i.e., conditioned stimulus) is followed by an effect
(i.e., unconditioned stimulus), the organism integrates this information resulting in a stronger cueeffect association and more vigorous responding when the cue is present. In humans, this also
results in stronger judgments of the causal influence of the cue (see Baker et al. 1989; Shanks,
1985, 1987, 1995; Shanks et al., 1996; Van Overwalle & Van Rooy, 2001a). Likewise, the delta
algorithm predicts that the more information that is received on the joint presence of a feature and a
category, the stronger their connection weight will become. This results in a pattern of increasing
weights as more pieces of information are processed, or a sample size effect (see illustration in
Figure 2A). In contrast, conventional probabilistic and statistical models of causality (e.g., Cheng
& Novick, 1992; Försterling, 1989) and attitude formation (Ajzen, 1991) do not predict a gradual
increase of judgments and remain silent with respect to sample size effects.
How is on-line learning and sample size effect achieved in connectionist models? Given the
assumption that the connection weights are initially set to zero (or any arbitrary low scale value),
the effect is that in the beginning phases of learning, the connection weights are relatively modest
and often inaccurate, and only grow more accurate (stronger or weaker, positive or negative) when
more information is received. The reason for this incremental learning is that the error in the delta
algorithm is only gradually minimized as regulated by the learning rate. Even when the covariation
between a feature and a category is perfect, the learning rate dictates that the weights connecting the
two will increase by only a small fraction.
Thus, it takes multiple repetitions of the same
information before a strong weight emerges.
Figure 2A depicts a system with a learning rate of 0.20. This means that the error of
underestimating a perfect correlation was corrected gradually by increasing the weights with 20% of
the error. As can be seen, because feature A is always paired with a category (i.e., perfect
correlation), its connection weight will gradually increase at each trial starting with 0.20 to reach
eventually its maximum value of +1 after a number of trails. Several researchers have noted that
given a sufficient number of trials, the delta learning algorithm converges to the same predictions as
conventional probabilistic and statistical models (Chapman & Robbins, 1990; Van Overwalle,
1996; Sarle, 1994).
Social Cognition and Connectionism 10
Competition and Discounting
Another essential property of the delta algorithm is that it gives rise to competition between
connections. This competition principle favors features or causes that are more predictive or
diagnostic than others, which are disfavored. The term competition stems from the associative
learning literature on animal conditioning and causality judgments mentioned earlier (Rescorla &
Wagner, 1972; Shanks, 1995), and should not be confused with other usages in the connectionist
literature such as competitive networks (McClelland & Rumelhart, 1988). A typical example of
competition is discounting in causal attribution. When one cause acquires strong causal weight,
perceivers tend to ignore alternative causes (Hansen & Hall, 1985; Kruglanski, Schwartz, Maides,
& Hamel, 1978; Rosenfield & Stephan, 1977; Van Overwalle & Van Rooy, 1998; Wells & Ronis,
1982).
Competition is a basic property of associative learning models like that of Rescorla and
Wagner (1972), where it is known as blocking. In fact, one of the reasons of the wide popularity of
the Rescorla-Wagner model is that it was among the first conditioning models that were able to
predict this property. As noted by several researchers (Read & Montoya, 1999; Van Overwalle,
1998), the delta algorithm makes similar predictions.
How does this property work in connectionist models?
The basic mechanism behind
competition is that the internal activation of an outcome node is determined by the sum of the
activations received from all connecting causal nodes (see Figure 2B). As the connection of cause
A is already relatively strong, it sends a great deal of activation to the outcome node. Any
additional activation from an alternative cause B leads to over-activation of the outcome node and
increased error, and therefore blocks any growth of connection strength of cause B.
Diffusion and Memory for Inconsistent Information
Still another property of the delta algorithm is that it is responsible for the weakening of
connections when a single node is connected to many nodes that are only occasionally activated. In
spreading activation models of memory, this property is known as the fan effect (Anderson, 1976),
although its underlying mechanism is fundamentally different from the diffusion property. In the
associative learning and connectionist literature, this is a novel property that — to our knowledge
— was not detected or mentioned earlier. The diffusion property is introduced to explain lower
Social Cognition and Connectionism 11
recall for consistent as compared to inconsistent information in impression formation (Hastie &
Kumar, 1979) and illusory correlation (Hamilton, Dugal & Trollier, 1985).
A typical example is a trait that implies many behaviors of which only a few are actually
present at a given time. The more behaviors that imply the same trait, the weaker each of the traitbehavior associations become. From a connectionist view, the reason is that while only a few
behaviors are activated (and their connections strengthened), all other possible trait-implying
behaviors are absent and thus remain inactivated, leading to a reduction of their connection weight
with the trait.
How does this diffusion principle explain enhanced recall of inconsistent information? As
illustrated in Figure 2C, compared to many consistent behaviors that imply trait T1, inconsistent
behaviors that imply trait T2 are, by definition, smaller in number. Hence, there is less often
inactivation and thus weakening of inconsistent connections (with T2) than of consistent
connections (with T1). This unequal weakening or diffusion therefore leads to enhanced recall for
inconsistent information.
Overview of the Simulations
Simulated Phenomena
We applied the three emergent connectionist processing principles to a number of classic
findings in the social cognition literature. For explanatory purposes, most often, we replicated a
well-known experiment that illustrates a particular phenomenon, although we occasionally also
simulated a theoretical prediction. Table 1 lists the topics of the simulations to be reported shortly,
the relevant empirical study or theory that we attempted to replicate, as well the major underlying
processing principle responsible for reproducing the data. Although not all relevant data in social
cognition can be addressed in a single paper, we are confident that we have included some of the
most relevant phenomena in the current literature.
General Methodology
We basically used the same methodology throughout the simulations.
The particular
conditions and trial orders of the focused experiments were reproduced as faithfully as possible,
although sometimes minor changes were introduced to simplify things (e.g., fewer trials than in the
actual experiments). When a random trial order was used, we ran the auto-associative network 50
Social Cognition and Connectionism 12
times with a different random order and averaged the results.
All parameters of the auto-associative model, except the learning rate, were kept fixed for
all simulations (Estr = Istr = Decay = 1, and internal cycles = 1, see McClelland & Rumelhart,
1988). We did not impose a common learning rate because of the different contexts, measures and
procedures used in the experiments. Rather, we freely selected a learning rate value that provided
the highest correlation with the observed data of each simulation, after examining all admissible
parameter values (see Gluck & Bower, 1988; Nosofsky, Kruschke & McKinley, 1992). In most
cases, the selected learning rate was quite robust. In other words, increasing or decreasing this
parameter had little substantial effect on the simulations. Only in a few cases where the original
learning rate was already high ( .28), increasing the rate further was problematic because the
weights grew out of bound (e.g., far beyond +1). The technical details on the auto-associative
model are given in Appendix A.
At the end of each simulated experiment or experimental condition, test trials were run in
which certain nodes of interest were turned on and the resulting activation in other nodes was
recorded to evaluate our predictions or to compare with observed experimental data. This will be
explained in more detail for each simulation. Except when otherwise noted, the obtained test
activations were projected onto the observed data using linear regression (with a positive slope), to
visually demonstrate the fit of the simulations to data. The reason for the use of this technique is
that most often only the pattern of test activations is of interest, rather than the exact values.
Structural Relationships
Categorization
Perhaps one of the most basic learning processes in social cognition is categorization, or the
grouping of diverse information into meaningful concepts or categories that contain similar features
(e.g., objects), functions (e.g., roles) or members (e.g., social groups). The categorization process
promotes cognitive economy and organization, which enables us to go beyond the current
information given and to plan our behavior and interaction with the external environment.
In recent approaches to categorization, members of a (social) category are not defined by
strict criteria of necessity or sufficiency, but rather by a degree of typicality or representativeness.
The process by which typicality is derived is most often described in terms of either a prototype or
Social Cognition and Connectionism 13
an exemplar approach. According to the prototype approach, learners abstract a central tendency of
each category and then classify instances according to their similarity to the category's central
prototype (e.g., Rosch, 1978). In contrast, no such average or ideal prototype is assumed in the
exemplar approach where categorizing of an object depends on the similarity with memory traces of
all instances in the category (Fiedler, 1996; Hintzmann, 1986; Medin & Shaffer, 1978; Nosofsky,
1986; Smith & Zárate, 1992).
Simulation 1: Categorization
How does a recurrent model simulate categorization? As we have seen, during learning, the
delta algorithm changes the weights between the object's features and the category so that they
better predict category membership. By this error-reducing process, the weights reflect a sort of
average link between the features of a category, that is, all instances are effectively "superimposed"
or abstracted into a prototype.
Let us illustrate the connectionist properties of feature similarity and prototype abstraction
with the network example shown in Figure 3. This network has four feature nodes and two
category nodes. Imagine that we are on a visit in Brussels and that we want to know whether an
inhabitant is Flemish or Walloon. Probably the best criterion is the language being used: either
Dutch (Flemish) or French (Walloon). However, because we may not always be able to hear these
people talk, there are other less perfect features we may rely on: A Fleming is often perceived as
simple-tasted and less sophisticated, refined or cultured than a Walloon.
Table 2 shows a simulated learning experience in which we perceive each of these features a
number of times and are also told the correct category (e.g., by our host). As can be seen, the
perfect features are always paired with their own category, whereas the imperfect features are often
absent even when the person can be categorized as Flemish or Walloon.
The results of this simulation are illustrated in the top panel of Figure 4. In addition, this
figure also depicts predictions from probabilistic theory and empirical data from Gluck and Bower
(1988, Experiment 1). In this experiment, subjects were given a medical diagnosis task in which
they had to learn to diagnose one of two diseases (i.e., categories) on the basis of four symptoms
(i.e., features). Our simulation is a simplified version of the learning trails given to the subjects in
Gluck and Bower's experiment. As can be seen, in Gluck and Bower's (1988) data, there was a
clear preference for the perfect category ("Dutch" in our example) that went above the 50% base-
Social Cognition and Connectionism 14
rate, illustrating base-rate neglect.
In the simulation, to measure the typicality of features with respect to a category, each
feature is activated or primed (see bottom panel of Table 2). This activation is automatically spread
upward to the category nodes, and the degree of category activation reflects the typicality of the
feature for that category. For instance, speaking Dutch is strongly related to the Flemish category
so that priming of this feature will strongly activate the Flemish category. Likewise, speaking
French is strongly related to the Walloon category and priming of this feature will strongly activate
the Walloon category.
To measure the preference of one category over the other, we considered the difference
between the resulting activation of the Flemish category node and the activation of the Walloon
category node (see Table 2). To map these simulation results on the proportional data of Figure 4
(where 50 % reflects an equal preference for both categories), the simulated data were regressed on
the observed data. The intercept in this simulation was held constant at .50. Hence, simulation
results above .50 reflect a preference for the Flemish category, while simulation results below .50
reflect a preference for the Walloon category. Not fixing the intercept in this manner would conceal
the relative preference for the Flemish or Walloon category.
As one would expect, Figure 4 (top panel) reveals that a perfect Flemish feature (e.g.,
Dutch) gives rise to the highest activation in the Flemish category. Similarly, a perfect Walloon
feature (e.g., French) gives rise to the highest activation in the French category (as indicated by the
lowest score in Figure 4). Imperfect features show activations that lie between these two extremes,
as they are ambivalent predictors of category belongingness. As can be seen, the simulations fit
nicely with research findings from Gluck and Bower (1988, Experiment 1) and better than
probabilistic predictions.
The bottom panel of Figure 4 illustrates the prototype of each category. To measure the
category prototype, the category node is primed (see bottom panel of Table 2). This activation
automatically spreads downward to the relevant feature nodes, and the resulting activation pattern
of the features reflects the prototype of the primed category. (These downward connections are not
shown in Figure 3, but are roughly equivalent to the upward connections). Thus, for instance, to
measure the prototypical Flemish features, the Flemish category is primed and the resulting
activation of the features reflects the prototype.
Social Cognition and Connectionism 15
As one might expect, Figure 4 (bottom panel) reveals that the prototype consists
predominantly of the category's perfect feature and less so of its imperfect feature. Because in our
simulation features of the other category were also present, they are also part of the prototype
although to a much weaker degree. Thus, the prototype is quite flexible as it may include features
that are relatively rare, although these features are clearly less relevant or typical of the prototype.
Base-Rate Neglect in Categorization
One of the reasons why we took the Brussels example is that the distribution of the social
categories is unequal: There are many more Walloons than Flemish living in Brussels. In the
original study of Gluck and Bower (1988), they had a similar imbalance between common and rare
disease categories. Given such an unequal distribution, the connectionist approach makes an
interesting prediction that is intuitively plausible, but difficult to explain by other approaches.
This prediction is base-rate neglect.
Perceivers often place more emphasis on the
diagnosticity or similarity of features and neglect the normative probabilities of the features'
occurrence in making categorical judgments (Gluck & Bower, 1988; Kruschke, 1996).
For
instance, although Table 2 shows that the probability that a Dutch-speaking Brussels inhabitant is
Flemish is equal to the probability that he or she is Walloon (e.g., 30 cases in both categories),
people tend to rely more on Dutch as an informative cue to make a Flemish categorization.
Intuitively, this makes sense. Dutch is a better predictor of being Flemish, because this feature is
quite often a good predictor of the Flemish category on its own. In contrast, Dutch is not a good
predictor of being Walloon as it is always contaminated by the presence of other Walloon features.
People's reliance on the most predictive feature regardless of normative probability has been
explained in the past by the operation of the representativeness heuristic, or by people's use of
similarity rather than probability to categorize.
In the social domain, the preference for diagnosticity of information is revealed in trait
inferences where people rely more on some types of behaviors than on others (Reeder & Brewer,
1979). For morality traits, people draw inferences more readily from negative behavior (e.g., lying)
whereas for ability traits, people draw inferences more readily from positive (e.g., successful)
behavior. This has been explained by the fact that immoral behaviors are more rare and unique and
thus more informative for morality judgments, while high-ability behaviors are more unique and
thus more informative for ability inferences.
Social Cognition and Connectionism 16
A connectionist network can predict this base-rate neglect in a straightforward manner.
Consider the results of the simulation in the top panel of Figure 4. As noted earlier, a score above
.50 indicates a preference for the Flemish category, while a score below .50 indicates a preference
for the Walloon category.
Of particular interest is that the perfect feature “Dutch-speaking”
exceeds the normative probabilistic prediction of .50. This reveals base-rate neglect.
From a
connectionist perspective, this is due to the competition principle. For a Flemish inhabitant, the
Dutch-speaking feature is the best predictor available so that the other features are discounted. In
contrast, because a Walloon most often possesses better predicting features than speaking Dutch,
this feature must compete against these better predictors and is discounted. As a result, the
connection weight of the Dutch feature is stronger with the Flemish category (.21) than with the
Walloon category (.05), resulting in a substantial proportion above .50 in favor of the Flemish
category.
Limitations and Future Directions
Although the present recurrent model is capable of explaining base-rate neglect, it is not
able to account for the inverse base-rate effect (Medin & Edelson, 1988; Kruschke, 1996; Shanks,
1992). Whereas base-rate neglect reflect the tendency to select a rare category (e.g., Flemish) when
tested with a single feature (e.g. speaking Dutch) for which the objective probably was equal for all
categories, an inverse base-rate effect reflects the tendency to select a rare category when tested
with a combination of conflicting features (e.g. speaking Dutch and French). An explanation for
this phenomenon is that the combination of both perfect features is quite distinctive and rare, and so
is more indicative of a rare category (Shanks, 1992).
A number of authors have developed connectionist network models to account for this
inverse base-rate effect. Gluck (1992) claimed that a distributed representation of the present
recurrent network approach was capable of explaining inverse base-rates.
However, our
simulations showed that this claim is incorrect, as this network cannot explain all data. Other
proposals are more effective and account for the inverse base-rate effect by increasing the attention
given to uncommon features. Shanks (1992) developed a simple extension of the standard delta
algorithm to give more attention to uncommon features and Kruschke (1996; Kruschke & Johansen,
1999) developed a network model that learns to attend more to features that distinguish them from
the already learned (frequent) category.
Social Cognition and Connectionism 17
Impression Formation
In a social context, getting to know others often involves drawing inferences about
characteristics and traits of individuals and groups. This process of impression formation, we will
argue, obeys connectionist principles similar to those underlying categorization processes in
general.
In a typical impression formation experiment, participants receive a series of trait adjectives
about a person and are requested to make overall trait or likability impressions (categorization) of
that person (e.g., Asch, 1946; Anderson, 1981; Kashima & Kerekes, 1994).
Sometimes the
adjectives are close synonyms that imply one or more specific traits, sometimes they are very
diverse and imply an overall likeability impression. Anderson (1981) argued that impressions of a
person are abstracted from trait adjectives as if people average these adjectives, and proposed a
weighted average model to explain person impression judgments.
Although his claim was
supported by an impressive amount of research, the model was criticized on the grounds that it
seems unlikely that people would perform all the necessary weighting and averaging calculations in
their mind to arrive at an impression, and many researchers abandoned Anderson's model for this
reason.
However, the connectionist metaphor used here can revive Anderson's model. The weighted
averaging principle can easily be implemented by implicit and automatic connectionist processes
based on the delta algorithm, without recourse to explicit arithmetical calculations (see Appendix B
for an algebraic proof). We will illustrate how a recurrent network can model impression formation
with two typical findings from person impression research.
Simulation 2: On-line integration and Recency
First, consider an experiment by Stewart (1965) in which adjectives describing a high trait
(e.g., talkative) were followed by opposite (or low) trait adjectives (e.g., reticent).
This
experimental manipulation was modeled using a network architecture consisting of a person node
and a task context node (which reflects instructions and other experimental context variables)
connected to a trait node. The simulations start from the assumption that the trait implied by an
adjective is already learned and recruited from semantic and social knowledge. Specifically, we
assume here that adjectives associated with the trait are denoted by an activation value of +1 for
Social Cognition and Connectionism 18
that trait, whereas adjectives associated with the opposite trait have an activation value of -1 (this is
equivalent to Anderson's scale values).
What is of interest here is how this trait-implying information is applied to build up an
impression of a specific person, by changing the weights linking the person with the trait. Table 3
depicts a schematic list of the information given in Stewart's (1965) experiment, where some
subjects received high trait information about a person in the first half of the experiment and low
trait information in the second half, and other subjects received the reverse low-high order. When
the person is described by a high trait, the connection weight is increased according to the
acquisition principle of the delta algorithm. In contrast, when the person is described by a low trait,
the weight is decreased according to the acquisition principle. After training, the person node in the
network is primed and the resulting activation of the trait node indicates what trait the person
conveys (see bottom panel of Table 3).
The results of a recurrent simulation are shown in Figure 5. As can be seen, there is a close
fit between the simulation and the data. Of particular interest is the crossover at the end of training
as the last presented adjectives win over the earlier presented adjectives, in both data and
simulations. This reflects a recency effect and suggests that the revision and adjustment of person
impressions is an on-line acquisition process where novel information often "overwrites" older
information previously stored in the connection weights.
Simulation 3: Recency in Concurrent Judgments, Primacy at Final Judgments
As a second example, consider research in which disconfirmatory information is given
during a single specific position in a series of trials. By comparing the effect of disconfirmatory
with confirmatory information at the same position in the trial series (denoted as serial position),
one can estimate the weight each trait takes at a given position (Anderson, 1979; Anderson &
Farkas, 1973; Busemeyer & Myung, 1988; Dreben, Fiske & Hastie, 1979; Kashima & Kerekes,
1994). Early disconfirmatory trait information might be important in crystallizing an impression
(primacy effect), while late information might be influential because it sheds new light on traits
presented earlier on (recency effect).
Research uniformly suggests that when participants give their trait ratings continuously after
each adjective is presented, then item weights are relatively equal in all but the last position, at
which point they rise sharply. This reflects a recency effect. However, it is most important to note
Social Cognition and Connectionism 19
that this recency effect attenuates when more trait information is given. Thus, when given only a
few pieces of trait-implying information, disconfirmatory information has a stronger effect than
when given more trait-implying information. It is as if increasing the amount of confirmatory
information shields the perceiver from the disconfirmatory information. In order to simulate this
result, we used the same recurrent architecture as before. A simulation of the experiment by
Dreben, Fiske and Hastie (1979) is schematically listed in Table 4 for the case when four adjectives
are given (the logic is similar for other frequencies).
The simulation results are shown in the top panel of Figure 6, where the dotted line depicts
the attenuation of recency. The recurrent network was clearly able to reproduce the predicted
attenuation, although attenuation was somewhat less steep in the simulations than in the data of
Dreben, Fiske and Hastie (1979). How did the recurrent model attain attenuation of recency? One
possible interpretation suggested by an analysis of the simulation is that the person node in the
recurrent network receives internal activation from the trait node (e.g., “When someone is talking
that loud, it must be John”). Because the trait node becomes positively linked with the person node
after confirmatory trials, it compensates for the disconfirmatory information, and it does so
increasingly better with more trials. Stated differently, a robust impression as a consequence of
earlier confirmatory information makes the perceiver more resistant to change his or her impression
given one disconfirmatory item. This explanation differs from Anderson's reasoning based on a
distinction between item-specific and abstract aspects of impression formation (Anderson & Farkas,
1973; Dreben et al., 1979). Similar recurrent simulations also fitted well with recent serial position
data from Kerekes (1991: in Kashima and Kerekes, 1994).
In contrast to the previous findings, trait weights show a typical primacy effect when
impression judgments are given at the end of the series of trials rather than continuously (Anderson,
1979). This primacy effect can also be simulated by the same recurrent network, as shown in the
bottom panel of Figure 6. Of specific interest is the much greater learning rate for this simulation,
which suggests that primacy might be a consequence of building up a prediction of the trait very
quickly in only a few trials, so that later information has little effect on the impression. This
interpretation shares with Anderson's (1981) attention decrement hypothesis the idea that there is
the most attention paid to and the most uptake of information during the earliest trials, thus
allowing little impact of information presented later. This hypothesis was also incorporated in an
Social Cognition and Connectionism 20
alternative connectionist model, termed the tensor product model, developed for impression
formation by Kashima and Kerekes (1994).
In sum, based on our connectionist simulations, we can explain the different effects of
continuous and final judgments by differences in learning rate.
This seems plausible, as
information uptake and processing is probably less interrupted by final trait judgments than by
continuous judgments, resulting in a faster learning rate for final judgments and hence primacy
rather than recency.
Limitations and Future Research
Although Kashima and Kerekes (1994; see also Busemeyer & Myung, 1988) correctly
pointed out that attenuation of recency cannot be simulated with a feedforward network, we have
demonstrated here that it can easily be reproduced with a recurrent model. This contradicts the
claim made by Kashima, Woolcock and Kashima (2000, p. 924) that this effect cannot be obtained
with a recurrent model. Moreover, unlike the tensor product model proposed by Kashima and
Kerekes (1994), our simulations do not require additional ad-hoc assumptions such as a changing
context after each judgment, to obtain attenuation of recency.
That our recurrent model can reproduce both recency and primacy effects is encouraging,
but as long as we cannot verify which independent conditions actually determine both effects, the
idea that both effects are driven by a different learning rate remains at best suggestive. The novel
hypothesis that grew from the simulations is that people build more robust impressions of a person,
either through a growing positive expectancy that shields them from disconfirming information
(attenuation of recency) or by building an impression very quickly and disregarding subsequent
information (primacy). However, we are not aware of any research that has explored this potential
explanation in depth.
Simulation 4: Higher recall for inconsistent information
In the previous research paradigms, participants received trait adjectives and were instructed
to form an impression about a person. This seems to reflect the manner in which we routinely
communicate about others. However, when we learn about others from our own observations, we
do not see traits but rather the behaviors that are associated with them. An intriguing finding given
this type of learning is that inconsistent or unexpected behavioral information is often better
Social Cognition and Connectionism 21
recalled than information that is consistent with the dominant trait expectation (for a review see
Stangor & McMillan, 1992). Thus, we better recall a hooligan helping an older lady cross the street
than a nurse performing the same act.
Hastie (1980, Hastie & Kumar, 1979) reasoned that the inconsistent information requires an
extra cognitive effort to explain and to make sense of the inconsistency, and is therefore elaborated
more deeply. This leads to extra links between the inconsistent information and other locations in
memory, and, thereby, to better recall. Hastie (1980) supported this interpretation by research
indicating that inconsistent information leads to more causal elaborations of the behavioral
sentences. However, these sentence elaborations were explicitly requested from the participants
after the initial phase of impression formation was over. It is thus not clear whether they were
generated spontaneously during initial encoding or only constructed after the request (cf., Nisbett &
Wilson, 1977).
Can connectionist principles account for the enhanced memory of inconsistent information
without recourse to explicit elaborative processes? Yes, and to illustrate this, we simulated a wellknown experiment by Hamilton, Katz and Leirer (1980, Experiment 3).
Participants read
information concerning several fictional persons. For each person, they read a list of 10 consistent
and 1 inconsistent behavioral descriptions about that person, after which they had to recall as many
behavioral sentences as possible. Half of the participants were given the instruction to form an
impression of the person, whereas the other half was given the instruction to memorize the
behavioral information. Under impression formation instructions, participants were more likely to
recall inconsistent items, whereas this difference disappeared under memory instructions.
To understand enhanced memory for inconsistent behavioral information, consider a
network architecture with a person node and a trait node, as in the previous simulations, as well as
separate nodes for each behavioral sentence. Thus, categorical trait information implied by the
behavior as well as the individual behavioral exemplars are represented in a sort of semi-distributed
manner. Table 5 provides a simplified simulation of Hamilton et al.'s experiment with 4 consistent
behaviors and 1 inconsistent behavior.
To simulate impression formation, each behavior was activated together with the associated
trait and the person node (see Table 5). As predicted by the diffusion principle, however, each time
a behavior is not present but expected due to the presence of the trait, this weakens the trait-
Social Cognition and Connectionism 22
behavior connection. Thus, the more behaviors confirm the expected trait, the less indicative each
behavior becomes for that trait or person. This is especially true for consistent behaviors, which
appear much more often not than with the trait. As a result, the behavioral links will be weaker for
consistent as opposed to inconsistent behaviors.
In contrast, in the memorizing condition, subjects are not motivated to form a unified trait
impression of the person. We assumed that this would result in a much shallower encoding of
person and trait information, which was simulated by setting the activation of these nodes to 0.10
instead of the typical 1. As a result, all links between the person or trait and the behaviors would
reduce sharply.
Figure 7 shows the results of the recurrent simulation. It was assumed that the person or the
traits would serve as cue to recall the specific behavioral episodes (see bottom panel of Table 5).
The simulations give very similar results when only the person or only the traits are primed to
retrieve the behavioral information. As can be seen, the simulations replicated the basic finding
that inconsistent information was better recalled than consistent information under impression
formation instructions. However, under memorizing instructions, enhanced memory disappeared.
It is important to note that the same simulation was able to replicate the well-known finding
that recognition measures produce the opposite tendency to report more consistent information
(Stangor & McMillan, 1992). This was accomplished by running the same simulation followed by
a recognition test that was biased by searching only for behaviors congruent with the consistent trait
(see bottom panel of Table 5). This reflects the idea that consistent traits guide recognition when
the perceiver relies on guessing.
However, if this bias was removed (by deleting ? for the
"common" trait), inconsistent behaviors were better recognized than consistent behaviors in line
with the improved recognition sensitivity measures reported by Stangor and McMillan (1992).
Limitations and Future Research
The simulation of higher recall of inconsistent behavioral information suggests that this
effect may be due to relatively stronger direct links of unique behavioral information. Thus, the
present connectionist account emphasizes the direct connections from a particular trait or person to
behavioral exemplars, while Hastie (1980, see also Srull, 1981) argued that stronger associations
between consistent and inconsistent behaviors after resolving the inconsistency were what produced
this higher recall. Our simulation does not rule out other processes, such as deeper and more
Social Cognition and Connectionism 23
elaborated processing (Hastie, 1980), that may contribute to the effect of better recall of incongruent
information. But is this more elaborated processing necessary?
Some support for the effortful generation of elaborations was demonstrated in studies that
found decreased recall for inconsistent behaviors when mental resources were limited by reducing
answering time, by making the task more complex, or by adding distracter tasks (Bargh & Thein,
1985; Hamilton, Driscoll & Worth, 1989; Macrae, Hewstone & Griffiths, 1993; Stangor & Duan,
1991). However, these results can be easily simulated with our connectionist network by simply
assuming that load decreased the encoding of the behavioral episodes or even all information (e.g.,
with an activation of 0.10). This suggests that poorer encoding of information, rather than less
inconsistency reduction and elaboration might have reduced recall of inconsistent information.
Hence, there seems to be no need to postulate explicit elaborations to explain higher recall of
inconsistent behavior.
The present perspective is also consistent with other findings that report less enhanced recall
for inconsistent information

when an impression is formed for a non-meaningful group of individuals, by assuming a
decreased activation of the person and trait nodes, based on the fact that perceivers are less
willing to invest cognitive effort in encoding an overall impression (Srull et al., 1985,
experiment 7),

for behavioral items at the beginning of a list compared to the end of a list (Srull et al.,
1985, experiments 5 & 6; Hastie & Kumar, 1979, experiment 3),

when the number of inconsistent items increases, thus making them less unique and
unexpected (Hastie & Kumar, 1979, experiment 3; Srull, 1981, experiments 1—3; Srull,
Lichtenstein & Rothbart, 1985, experiment 3).
Overall, it appears that the proposed model is broadly consistent with a relatively large
spectrum of research findings. This suggests that the diffusion principle provides an interesting
alternative hypothesis explaining increased recall for inconsistent information.
Assimilation and Contrast in Person and Group Perception
An important feature of recurrent models is their capacity to generalize. A trained network
exposed to an incomplete pattern of information will fill in the missing information on the basis of
the complete pattern learned previously. This generalization process can be seen as a type of
Social Cognition and Connectionism 24
assimilation in that past experiences influence how we perceive and interpret novel information that
is similar or closely related to it.
For instance, when seeing a photo of Hitler, we might
immediately complete this image with activated memories on his aggressive wars, mass
annihilation of Jews and so on. There is abundant evidence showing that accessible knowledge like
traits, stereotypes, moods, emotions and attitudes is likely to result in the generalization to
unobserved features. In the next simulations, we will explore some applications of this capacity to
generalize, as well the opposite capacity to generate contrast effects in person perception (Anderson
& Cole, 1990; Smith & DeCoster, 1998).
Simulation 5: Assimilation of Unobserved Attributes
To demonstrate generalization in a recurrent network, imagine that the network learns that
Hitler was a cruel German Nazi leader who was responsible for the mass annihilation of Jews.
When the network is then tested with a Hitler probe and a few related attributes (e.g., German, nazi,
cruel), would it use this knowledge to activate the missing feature of mass annihilation? Recall that
activation in a recurrent network is determined not only by external input, but also by internal input
coming from related nodes in proportion to their connection weights. This implies that although
the missing annihilation node receives no external activation, it does receive internal activation
through its links with the Hitler and other related nodes.
Table 6 shows a schematic description of a simulation that combines several simulations by
Smith and DeCoster (1998). In this simulation, we presented information on three individual
exemplars such as Hitler, Goebbels, and Himmler, each defined by five features (labeled E1—E5,
E6-E10, and E11-E15) as well as on their group (e.g., Nazi) that was characterized by three features
(labeled G1—G3).
To increase the realism of the simulation, like Smith and DeCoster, we represent features in
a distributed manner, that is, each feature is represented by a set of micro-features (unlike our
previous simulations in which a localist coding scheme was used with each feature being
represented by a single node). Distributed representations are more realistic because we know that
symbolic concepts are not represented by single neurons but rather by assemblies of neurons.
Specifically, each feature was represented by 5 micro-features or nodes. For instance, Hitler was
not represented by five nodes, but rather by a series of 25 nodes that reflected several micro-features
of his physical appearance, character and so on. In addition, we also use random noise in the
Social Cognition and Connectionism 25
presentation of background context and features to simulate the imperfect conditions of perception.
Although these latter aspects appear in our simulation mainly for purposes of comparison, they are
not essential in understanding the generalization process or in producing the results (e.g., the noise
cancels out given enough simulation runs).
After going through the learning history of Table 6, all but one feature of the individual
exemplars or group were primed (see bottom panel of Table 6). Figure 8 depicts the resulting
activation of the remaining feature (represented by five nodes). As can be seen, the internal
activation of the other nodes in the network allows the network to reconstruct the missing
information of the original learned pattern almost perfectly, for both the individual exemplar and
for the group. This indicates that the recurrent network is capable of integrating and utilizing both
individualized and schematic (i.e., group) information.
Further research
Smith and DeCoster (1998) demonstrated that a recurrent network can reproduce other very
interesting phenomena of social cognition. Perhaps one of the most intriguing properties is the
creation of new emergent attributes by combining parts of existing attributes (see Smith &
DeCoster, 1998, simulation 3). Traditional theories of categorization assume that people use a
single schema, stereotype or knowledge structure to make inferences about a target person or a
group. Even if multiple schemas are relevant, each of them is independently activated and applied.
However, people can combine many sources of knowledge in order to construct new emergent
properties to describe subtypes or subgroups of people. For instance, a militant feminist who is also
a bank teller may become subtyped as a feminist bank teller with specific idiosyncratic attributes
(Smith & DeCoster, 1998; Asch & Zukier, 1984). Previous connectionist models like ECHO
(Thagard, 1989) were unable to model this process.
Simulation 6: Assimilation with Traits, Contrast with Exemplars
The abundance of assimilation effects in social cognition research may generate the
suggestion that filling in unobserved characteristics is the default or most natural process. Thus,
when primed with “violent,” we judge a non-descript or ambiguous target person as more hostile,
and when primed with “nice,” we judge that same target as less hostile. However, under some
circumstances, the opposite effect may occur. Sometimes primed features may lead to contrast
Social Cognition and Connectionism 26
rather than assimilation.
For instance, when primed with the exemplar Gandhi, people may judge a target person as
relatively more hostile, whereas primed with Hitler, they may judge the same target as relatively
less hostile. Under these conditions, the exemplars Gandhi and Hitler serve as an anchor against
which the target is judged, and so leads to contrast effects. In sum, contextually (or chronically)
primed information may not only serve as an interpretation frame, but also as a comparison
standard during impression formation.
What produces assimilation or contrast? According to Stapel, Koomen and Van der Pligt
(1997), trait concepts are more likely to serve to interpret an ambiguous person description
(assimilation), because traits carry with them only conceptual meaning.
On the other hand,
exemplars -- if sufficiently extreme -- will be used as comparison standard (contrast) because both
the exemplar and the target are persons that can be compared with each other. An experiment by
Stapel et al. (1997) confirmed this proposition. Participants were asked to form an impression of an
ambiguous friendly or hostile target person. Before they were exposed to the description of the
target, they were primed with names of traits (e.g., violent or nice) or with names of extreme
exemplars (e.g., Hitler or Gandhi). Assimilation was found in the trait priming condition, whereas
contrast was found in the person priming condition.
A recurrent network can simulate this combination of assimilation and contrast. As listed in
Table 7, the network first builds up background knowledge about average persons (who are at times
more or less violent or nice), extreme exemplars like Hitler and Gandhi, as well as about the
relationships between traits (e.g., nice is the opposite of hostile and violent).
The essential idea of the simulation is that during priming, the primed stimulus and the
target description are temporarily activated together. This is represented by programming a single
learning trial for each priming condition (see Table 7). Because testing a trait category involves
connections from person to trait, there is competition between exemplars, but not between traits and
exemplars. Hence, when a trait concept is primed, this leads to the usual assimilation of the trait
impression through the acquisition principle. In contrast, when an exemplar such as Hitler is
primed, competition arises between this exemplar and the target exemplar (which is so nondescript
that it is assumed to be taken as an instance of an average person). This competition arises when
both exemplars predict hostility and their summed activation overestimates the observed degree of
Social Cognition and Connectionism 27
hostility. This error leads to a decrease in the connection weights between target and hostility, and
results in a contrast effect.
The full learning history of this simulation is listed in Table 7. Distributed coding and noise
were used in the priming trials to implement the idea that slightly different instances of traits and
exemplars were used in the priming and prior knowledge phases. As can be seen in Figure 9, the
simulation replicated the empirical assimilation and contrast effects as reported by Stapel, Koomen
and Van der Pligt (1997).
Limitations and Future Research
The exemplars that serve as a comparison standard need to be sufficiently extreme, because
otherwise little overestimation would occur in the network, and thus little contrast. This prediction
is supported by a recent study by Moskowitz and Skurnik (1999). In two experiments, they found
that that moderate exemplars (e.g., Kissinger) lead to less contrast than extreme exemplars (e.g.,
Hitler). As one might expect from the recurrent network's generalization property, they also found
that moderate trait primes lead to less assimilation than extreme primes. The present recurrent
network was able to reproduce the findings of Moskowitz and Skurnik (1999).
However, the present network is, at the moment, not capable of reproducing the effect of
cognitive load on assimilation and contrast. Moskowitz and Skurnik (1999) showed that cognitive
interference (i.e., increasing task load or interrupting the current task) minimized the effects of trait
assimilation, but left the effects of exemplar contrasts relatively untouched.
If we simulate
decreased resources during priming by decreasing node activation, however, we would expect the
opposite effect to occur. Future research is necessary to ensure that Moskowitz findings are robust,
and if so, how task load can be implemented in a recurrent network so that it can approximate their
findings.
Causal Judgments and Attitudes
In this section we discuss causal judgments and attitude formation from an attributional
perspective. A first question is how causes, attitudes and effects are represented in a connectionist
network. As with the recurrent networks of social judgment described earlier, we represent causes
and attitude-objects as features, and outcomes or behaviors as categories. However, whereas noncausal features in social inference are rather passive descriptors or predictors of category
Social Cognition and Connectionism 28
membership, intuitively, causes and attitude-objects have a more active role, in that they also tend
to play a causal role about the outcomes they predict. For instance, an angry face does not only tell
something about the person (social inference), but also warns the observer to defend him- or herself
for possible attacks (causal inference). Likewise, an attitude-object like a toy may not only look
attractive (trait inference), but may also increase approach behavior (causal inference).
This
difference between the descriptive nature of social inference and the more active role or power of
causes and attitudes is not explicitly modeled in connectionist models, but is evident from the
typical sort of categories which reflect social events (behaviors, outcomes) rather than social
entities (traits, groups, family, etc.).
Causal Attributions
Recent research has demonstrated that there are many parallels between human models of
causal attribution and animal conditioning models (for overviews see Allen, 1993; Shanks, 1995;
Read & Montoya, 1999). To cite a few important parallels, one of the most popular models in
animal learning, the Rescorla-Wagner (1972) model, is identical to the delta learning algorithm
(implemented in a feedforward network), and it has also been shown that this model asymptotically
converges to another popular model of human causality based on probabilistic principles (Cheng &
Novick, 1992; see Chapman & Robbins, 1990; Van Overwalle, 1996).
In a recent article, Read and Montoya (1999) successfully simulated a number of
phenomena from the animal learning literature with a recurrent network (see their Table 2, p. 735).
This simulation work demonstrated that a recurrent model can reproduce competition between
alternative causes such as

discounting (where one cause blocks the causal influence of an alternative cause),

augmentation (where one inhibitory cause increases the influence of an alternative cause
that facilitates the outcome),

inhibition (where an alternative cause develops inhibitory effects that prevent the outcome
from occurring),

overshadowing (where the causal strength of two causes is less than that of a single cause in
predicting the same outcome).
Like many researchers in the social and animal learning domain, we apply the terms
discounting and augmentation quite broadly to denote causal competition both during and after
Social Cognition and Connectionism 29
causal learning, that is, during or after novel, causally relevant information is received and
processed. Thus, competition may occur between information taken in at any time, either with
novel information (or novel causes) or with earlier material reactivated from memory (or known
causes). This differs from the position taken by other authors (Morris & Larrick, 1995; Read &
Montoya, 1999) who reserve the terms discounting and augmentation exclusively for reasoning
processes based on prior causal learning in the original sense of Kelley (1972).
Because several authors (Van Overwalle, 1998; Van Overwalle & Van Rooy, 2001a, 2001b;
Read & Montoya, 1999) have already provided many illustrations of connectionist modeling of
causal attribution, we will present only a single simulation of this phenomenon.
Simulation 7: Forward Discounting
A common finding in animal and human literature is that when a particular cause has
already explained an outcome, then any alternative cause is always discounted. This is called
forward discounting. The idea of forward discounting is largely consistent with the anchoring
explanation in social psychology, which assumes that people anchor on the first presented
explanation or the first dominant explanation that comes to mind (e.g., the actor's disposition) and
tend to ignore novel information implicating alternative explanations (Shaklee & Fischhoff, 1982;
Gilbert & Malone, 1995). For instance, Van Overwalle, Drenth and Marsman (1999) found that
spontaneous trait inferences were not moderated by covariation information when presented after a
description of the actor's behavior, but only when presented before it.
Thus, when known
personality traits or other situational pressures provide a ready explanation for someone's behavior,
people tend to disregard novel information about additional factors.
In a series of experiments, Van Overwalle and Van Rooy (1998, 2001b) combined the
process of forward discounting with sample size, that is, by increasing the sample size and thus the
strength of a known cause, discounting of a novel cause was made stronger. Thus, they combined
the emergent principles of acquisition and competing. Participants read stories in which several
causes could explain an outcome. For instance, in one of the stories they were first told that Ann
won several single tennis games, and then that Ann (now a known cause) also won several double
games with Troy (a novel cause). As expected, given Ann's previous successes, the contribution of
Troy was decreased or discounted.
However, a crucial manipulation was how often Ann won her single games. When Ann
Social Cognition and Connectionism 30
won her single games only once, she acquired little causal strength and discounting of Troy was
much weaker than when Ann won several times, thus acquiring more causal strength. Thus,
discounting of Troy was indirectly influenced by the weakening or strengthening of Ann. Table 8
shows the design of this experiment, and Figure 10 depicts the simulated and observed results. As
can be seen, the recurrent network conforms nicely to the observed data.
Note that current
statistical models of causality (Cheng & Novick, 1992; Försterling, 1989) are unable to account for
these results.
Van Overwalle and Van Rooy (1998, 2001b) performed similar experiments
involving augmentation, and found parallel results consistent with the combined predictions of
sample size and competition.
Limitations and Future Research
Discounting can occur not only when the alternative explanation is a novel one as in the
simulation, but also when competing causes are processed simultaneously. Thus, competition
effects do not require a fixed sequence of processing of causal information, as assumed in phaselike models of dispositional attribution (e.g., Gilbert, 1989). This implication is in line with recent
research suggesting that the weighting of competing person and environmental attributions
"involves an iterative or even simultaneous evaluation of the various hypotheses before reaching a
conclusion" (Trope & Gaunt, 2000, p. 353).
However, what happens when competition arises after the competing causes have already
gained causal strength? For instance, if Troy and Ann always won their double games, and now we
learn that Ann alone wins all her singles, what do we think about Troy? This now involves
backward revaluation. According to Dickinson and Burke (1996), backward revaluation depends
on the relationship between the two causes. They found that when causes are positively related,
then discounting will take place; when they are independent, there will be no discounting (see also
Van Overwalle & Timmermans, 2000). These results cannot be simulated with the present standard
recurrent network, but requires a modification so that absent causes that are expected (via prior
compound presentation) receive a negative activation rather than the standard "filling up" of
activation from related nodes (for more details see Van Overwalle and Timmermans, 2000, 2001;
Graham, 1999).
Social Cognition and Connectionism 31
Attitudes
The most influential and popular model of attitude formation is the theory of reasoned
action developed by Fishbein and Ajzen (1975) and later refined and relabeled as the theory of
planned behavior (Ajzen, 1991; Ajzen & Madden, 1986). According to this model, an attitude is a
function of

the expectation or belief that the behavior will lead to a certain consequence or outcome
(e.g., various means of transportation, like cars and buses pollute; bicycles don't), and

the person's evaluation of these outcomes (e.g., pollution is bad).
Multiplying the expectancy and value components associated with each outcome and
summing up these products determines an attitude.
The theory of planned behavior has received considerable empirical support in many studies
(see Fishbein & Ajzen, 1975; Ajzen, 1991; Ajzen & Madden, 1986), although it has been found that
other factors besides attitudes may exert an influence on behavior. A major criticism leveled
against the theory, however, is its assumption that humans make rational decisions, and carefully
elaborate and compare alternative behavioral options before they engage in a particular behavior. It
seems unlikely that people engage in extensive processing of the pros and cons of specific
behavioral alternatives for every opinion or attitude they have (Fazio, 1990). Although Ajzen and
Fishbein (1980) acknowledged that people may simply reactivate and employ attitudes formed
previously, they still assumed that these prior attitudes had been formed explicitly.
Simulation 8: Attitude Formation
We propose, however, that attitudes may also be developed implicitly. Recent research by
Betsch et al. (2001) indicates that the encoding of value-charged stimuli is sufficient to prompt an
on-line process by which values are implicitly summed and stored in memory. A process of
implicit attitude formation, representation and retrieval in memory without deliberative processing
can be modeled by a connectionist implementation of the theory of planned behavior.
To illustrate, let us return to the above example. The first attitude component, the belief that
one's choice will result in certain outcomes, can be represented as causal expectations linking the
choice of transportation with likely outcomes such as how fast a car will be, how dry the trip will
be, and how polluting.
The likelihood of these outcomes is expressed in the weight of the
Social Cognition and Connectionism 32
connections, acquired during prior experiences. Thus, the more often a particular consequence is
observed, the stronger the weight becomes. Conversely, the less often a particular consequence is
observed, the weaker this weight will be.
The second attitude component, the value, can be represented by concurrent emotional or
evaluative responses to these outcomes, such as, how much the person likes or dislikes being in a
fast and polluting car, in a dry place, again acquired during prior experiences. Thus, in line with
Ajzen (1991, p. 191), we assume that the outcomes linked with a behavior are "valued positively or
negatively", and that they are further modified during actual experiences.
In a connectionist network, the activation sent out by each means of transportation is
multiplied by the weight of the connections associated with the outcomes, including the value node.
We suggest that a person's attitude is reflected in the activation of this value node after the relevant
attitude-object (i.e., means of transportation) was activated. This proposition is mathematically
very similar to the multiplicative function in Ajzen's (1991) theory of planned behavior (i.e., where
expectations are replaced by connection weights, and values are replaced by value node activations;
see Appendix C for a formal proof).
However, it does not require the less plausible assumption of
deliberate weighting of all alternatives, as only the beliefs and evaluations that are accessible in
memory at the time of judgment will determine the attitude. Note that outcomes other than the
value (fast, dry and polluting) are not taken into account for measuring an attitude because they
reflect cognitions related to the attitude-objects rather than evaluations. This is consistent with the
dominant view in the attitude literature that takes attitudes primarily as evaluative responses.
Table 9 depicts a recurrent simulation of this example. The likelihood of the outcomes is
determined by the frequency that a causal factor co-occurs with an outcome, and the value is
determined by the degree of satisfaction or dissatisfaction experienced during this outcome.
Although we used extreme +1 and -1 evaluative values for simplicity, moderate values are also
possible. In Figure 11, the simulated values are compared with predictions of the theory of planned
behavior. As can be seen, the simulated and predicted data match almost perfectly.
Simulation 9: Dual-Process Models of Persuasion
The theory of planned behavior (Ajzen, 1991; Ajzen & Madden, 1986) assumes that people
systematically scrutinize all relevant information for making an attitude judgment. Although this
might be the preferred approach when forming an initial opinion about an important issue
Social Cognition and Connectionism 33
(Gollwitzer, 1990), in many cases attitudes are created or changed in a more shallow or heuristic
manner. This distinction has been captured in the heuristic—systematic model of Chaiken (1980,
1987; Chen & Chaiken, 1999) and the elaboration likelihood model (Petty & Cacioppo, 1986; Petty
& Wegener, 1999). According to these dual-process models, systematic processing implies that
people have formed or updated their attitudes by actively attending to and cognitively reflecting
upon persuasive argumentation. In contrast, heuristic processing implies that people have formed
or changed their attitudes by using heuristic cues that give rise, automatically, to stored decision
rules such as "experts can be trusted", "majority opinion is correct", and "long messages are valid
messages".
Dual-process theories regard systematic processing of information as requiring more effort
and cognitive capacity than heuristic processing.
Hence, when motivation or capacity for
systematic scrutiny of information is low, such as when the issue is of low personal relevance or
when time is limited, people use heuristics like source credibility, other people's attitudes or the
length and number of arguments.
These two processing modes are not necessarily mutually
exclusive. For instance, systematic and heuristic processing may co-occur when the arguments are
too ambiguous to form an opinion by extensive processing alone, that is, heuristic cues may
additionally help to form an opinion by biasing the selection of ambiguous information (Chen &
Chaiken, 1999).
Such an interaction between systematic and heuristic processing was investigated by
Chaiken and Maheswaran (1994). They presented a message about a fictitious answering machine
in which different features were described with varying importance.
This information was
ostensibly published either in a highly regarded magazine specialized in scientific testing of new
products or in a promotional pamphlet prepared by sales personnel. The results revealed that with
low task importance (i.e., respondents' opinion would have little bearing on the manufacturer's
product distribution), source credibility was the only determinant of people's attitude. In contrast,
with high task importance (i.e., respondents' opinion would count heavily), the quality of the
machines' features was the only determining factor, except when the message was ambiguous and
source credibility alone influenced the attitude. This study is important, because it demonstrates
several predictions of dual-process models. It documents how heuristic cues can bias message
arguments that are ambiguous (the biasing hypothesis; Chen & Chaiken, 1999), and how systematic
Social Cognition and Connectionism 34
processing can overrule heuristic cues when the arguments are unambiguous (the attenuation
hypothesis; Chen & Chaiken, 1999).
We simulated the interactive nature of systematic and heuristic processes as investigated by
Chaiken and Maheswaran (1994) with our recurrent model (see Table 10). According to Bohner,
Ruder and Erb (1999), heuristic cues like source credibility may lead people to form expectations
about message valence or strength.
We assume that these expectations are driven by prior
experiences of good and bad argumentation with the same or similar sources, or by communicated
opinions about such experiences. As can be seen in the top panel of Table 10, this assumption of
prior knowledge on quality of argumentation was incorporated by setting the value node to +1 (high
credibility) or .10 (low credibility) during an initial prior learning phase.
Next, we ran one of three message types, involving strong, weak and ambiguous arguments,
which were simulated by different activation levels of the value node. That is, the strength and
direction of the arguments was determined by setting the activation of the value node to either
positive or zero.
(No negative activation was used as weak arguments actually involved
descriptions of available features for which other products were however superior). Table 10
depicts a simplified version of the actual design used by Chaiken and Maheswaran (1994).
More importantly, given heuristic processing, in line with the basic assumptions of dual
process models, we assumed that these arguments would not be encoded or elaborated sufficiently
(i.e., we ran one trial only with all activation levels divided by 10). In contrast, given systematic
processing we assumed that these arguments would be processed more extensively (i.e., two trials
as shown in Table 10). The results depicted in Figure 12 reveal that our simulation reproduced the
predicted pattern as observed by Chaiken and Maheswaran (1994).
Thus, the simulation
reproduced heuristic and systematic processing, as well the predicted interaction between both (i.e.,
biasing and attenuation effects; Chen & Chaiken, 1999).
Siebler, Bohner and Weinerth (1998) proposed an alternative connectionist constraint
satisfaction model (i.e., ECHO, Thagard, 1989) to account for the same data. However, this latter
type of connectionist model suffers from shortcomings (e.g., no weight adjustments, no permanent
attitude change) to be discussed later. Nevertheless, for the present simulation, it should be noted
that small deviations in the learning rate destroyed the predicted interaction between the systematic
and heuristic processing of the ambiguous message. Specifically, a higher learning rate caused the
Social Cognition and Connectionism 35
novel arguments to overwrite all memory of the source's credibility, whereas a lower learning rate
caused the source's credibility to be the only determinant of the attitude. This seems to suggest that
the interaction between systematic and heuristic processing depends on a critical balance between
source credibility and (rate of) systematic elaboration of argument quality, which is in line with the
sparse reports on this interaction in the attitude literature (Chaiken, Liberman & Eagly, 1989, p.
233; but see Bohner, Moskowitz, Chaiken, 1995).
Simulation 10: Cognitive Dissonance
Sometimes, our attitudes are not so much driven by immediate evaluations of attitude
objects, but rather by reactions to our own behaviors, especially when these behaviors go against
our initial preferences. This phenomenon is captured in Festinger's (1957) theory of cognitive
dissonance, which predicts that discrepant behavior generates dissonance or uneasiness that "will
exert pressures in the direction of bringing the appropriate cognitive elements into correspondence"
(p. 11). For instance, when induced to write an essay that runs counter to one's initial attitude (e.g.,
a student defending stricter exam criteria), an individual will tend to reduce dissonance by changing
his or her attitude in the direction of the position taken in the essay. This tendency is stronger when
alternative explanatory factors or justification such as high payment or social pressure, are absent.
In contrast, when external demands (e.g., payment or pressure by the experimenter) provide
sufficient justification for engaging in the dissonant behavior, then dissonance reduction does not
occur (e.g., Linder, Cooper & Jones, 1976).
Cooper and Fazio (1984) have proposed an attributional analysis of the process of cognitive
dissonance reduction. They suggested that individuals attempt to understand and justify their
discrepant behavior ("Why did I behave this way?"). When alternative causal explanations for the
discrepant behavior are absent, then participants conclude that they must have liked writing the
essay more than they initially thought, and this results in attitude change. Conversely, when
sufficient external explanations are available, no dissonance is experienced and no attitude change
will occur. Thus, for instance, more attitude change is expected given a low rather than high
monetary reward.
However, the reverse effect has been observed when individuals are forced to engage in
discrepant behavior. Indeed, in this case, there is more attitude change with a high rather than low
monetary reward (Linder, Cooper & Jones, 1976). To explain this opposite effect, Van Overwalle
Social Cognition and Connectionism 36
and Jordens (2001) extended the attributional analysis by assuming that individuals try to
understand not only their discrepant behavior, but also their concurrent feelings ("Why do I feel this
way?"). In the case of high external constraints like strong pressure towards discrepant behavior
and low payment, the experimental situation will be experienced as particularly unpleasant.
According to Van Overwalle and Jordens (2001), these negative feelings will counteract and reduce
the attitude discrepancy, as if a person concludes that although having done something wrong, he or
she was already sufficiently punished for it by feeling very bad about it.
We conducted a connectionist implementation of the original experiment by Linder et al.
(1967), which is very similar to the simulation by Van Overwalle and Jordens (2001). The learning
history is shown in Table 11. To simulate the idea that prior experiences are only roughly similar to
the experimental manipulations, we used a distributed representation with noise added (see top
panel). The experimental manipulation was implemented as a single trial, to reflect the assumption
that attributional thoughts were raised at least once during the experiment (see middle panel). The
attitude toward the essay was measured by priming the attitude-object (i.e., essay) and reading off
the activation of both the behavioral and affective outcomes. Thus, an attitude is seen here not only
as an affective response, but also as a behavioral approach—avoidance response.
How is attitude change under induced choice simulated? Given that pressure from the
experimenter is absent, only the attitude-object (the essay) and payment serve as potential causes in
explaining discrepant behavior and concurrent feelings.
A lowered reward is simulated by
decreasing the activation of the payment node to .20. This results in compensatory augmentation
(i.e., competition principle) of the connections from the essay node to the behavioral and affective
outcomes, with as consequence an increased positive attitude toward the essay.
How was the no-choice condition simulated? In addition to the influence of payment, the
negative feelings arising from low payment combined with experimenter pressure drive the
connection between the essay and evaluative outcome downward, resulting in a decreased attitude
toward the essay. The results of this simulation are shown in Figure 13 and compared with
empirical data by Linder et al. (1967). As can be seen, the fit between simulated and observed data
was excellent.
Limitations and further research
The present simulations encompass a wide variety of models and data in the attitude
Social Cognition and Connectionism 37
literature. We are just beginning to uncover the implications of connectionist modeling for this area
in social cognition. However, from this initial sketch it appears that seemingly different modes of
processing and types of persuasion and information may all be driven by the same underlying
connectionist mechanisms. In addition, our analysis paints a somewhat different picture of heuristic
cues in attitude formation, in particular, and in social cognition in general. It is to this discussion
that we now turn.
A Note on Judgmental Heuristics
The mainstream theoretical approach to judgment in social psychology is that information
processing is rarely exhaustive or guided by logical norms, but rather reveals a compromise
between rationality and economy. In this approach, effortless judgments are typified by judgmental
heuristics that enable individuals to make rapid and easy judgments by rules-of-thumb that require
little explicit thinking but, overall, provide adequate responses most of the time. According to this
view, the price of such rapid judgments can be observed in a series of biases.
Rather than viewing heuristics and biases as exceptions to the rules of logical thinking, we
would like to argue that they actually reflect how the brain — as a connectionist device — works.
Take, for example, the heuristics assumed to influence judgments under uncertainty (Kahneman,
Slovic & Tversky, 1982), or the heuristic rules of dual-process models of persuasion (Chaiken,
1980; Petty & Cacioppo, 1987).
Heuristics under Uncertainty
We will consider three major heuristics used in judgment making under uncertainty. These
are:
Availability. The availability heuristic reflects the finding that many judgments are biased
by information about facts and arguments available in memory, either due to frequent (chronic)
utilization in the past or to recent priming. This is exactly what a recurrent network would predict.
Information that is recently primed or activated is spread to other related concepts, influencing
judgments about them as we have seen in the earlier assimilation examples (Simulation 5). A
dramatic demonstration of chronic accessibility was given by Smith and DeCoster (1998,
Simulation 5). They demonstrated that people who used a particular concept frequently in the past
might lose this information if it was "overwritten" by novel information, but this concept could be
Social Cognition and Connectionism 38
quickly be recovered after a few presentations of the original information.
Representativeness. The representativeness heuristic has been invoked to explain the
finding that categorization is often guided by resemblance between concepts rather than by
statistical base rates. As we have seen in the section on categorization, this is exactly what one
would expect from a connectionist view. In Simulation 1, we demonstrated that a category is
chosen on the basis of its most unique (diagnostic) feature, even if that feature has the same base
rate as another, less unique feature.
Anchoring. The anchoring and adjustment heuristic has been proposed to explain why
judgments are often biased toward an initial anchor and has been taken as evidence that judgments
are often made and adjusted on-line. Again, anchoring can be simply taken as a consequence of online or incremental connectionist learning. According to the delta learning algorithm, weight
adjustments are often stronger initially because of the greater error in the network, while later
adjustments become increasingly smaller because the error is reduced.
It is interesting to ask why adjustments in later phases of learning are often insufficient to
engender a change of opinion. For instance, why are situational constraints and pressures seldom
taken into account when making dispositional inferences about an actor? The answer may be found
in the backward revaluation hypothesis proposed by Dickinson and Burke (1996). As noted in the
causal attribution section, this hypothesis posits that backward discounting of causal factors will
occur only when there is a strong unique relationship between these causes.
Because the
relationship between an actor and a situation is seldom unique (e.g., many different actors may
appear in the same situation), this revaluation hypothesis predicts that correction of personal
judgments by situational information will often be insufficient (see also Van Overwalle &
Timmermans, 2000, 2001).
Heuristics in Persuasion
Dual-process models of persuasion (Chaiken, 1987; Petty & Cacioppo, 1986) posit that
people revert to heuristic processing when their motivation or their capacity to analyze message
content in detail is minimal. Heuristic cues are characterized as salient, easily processed pieces of
stimulus information that gives rise, automatically, to the activation of a stored decision rule
(Chaiken, Duckworth & Darke, 1999).
These heuristic rules are developed through past
experiences and observations and include, for example, beliefs that "experts can be trusted",
Social Cognition and Connectionism 39
"multiple arguments are stronger", "high consensus implies correctness", and "things I like are
good". Heuristic processing involves automatic processing of these rules with little awareness of
their occurrence and their impact on judgments.
Although we agree that heuristic processing is automatic and often beyond awareness, we
would argue that it does not necessarily involve the application of well-learned rules. We do not
exclude this possibility, but we propose that connectionist principles provide a much more
convincing and parsimonious account of the implicit nature of heuristic processes. To support this
contention, let us review a number of these heuristic rules and see how connectionist processes can
explain them. As we shall see, the proposed connectionist mechanisms differ markedly from the
original hypotheses in the literature on how heuristics are "activated" and "applied" to influence
attitude judgments (Chen & Chaiken, 1999).
Expertise.
This heuristic was already simulated in the previous section (Chaiken &
Maheswaran, 1994; Petty, Cacioppo & Goldman, 1981). It was shown that expertise involves an
expectation about argument quality and value, based on prior learning from the same or similar
sources. During prior learning, the activation of the value node is high when the source is an expert
(with a standard activation level of 1) or low when the source is not an expert (e.g., 0.10). Most
importantly, we contend that the effect of knowledge resulting from prior learning on source quality
is naturally integrated with novel information about an attitude-object through the principle of
acquisition. Hence, this heuristic does not require activation of any explicit rule or belief.
Consensus. This heuristic (Maheswaran & Chaiken, 1991) functions in very similar ways
to the expertise heuristic. Consensus information involves an expectation about the positive or
negative value of features (i.e., implemented by a positive or negative activation of the value node),
based on prior learning from other sources. In a connectionist framework, this expectation or prior
knowledge is naturally integrated with novel information, without any recourse to additional rules
or beliefs.
Length. Lengthy messages tend to contain more arguments (Petty & Cacioppo, 1984) or
tend to repeat the same arguments in different words with more detail (Wood, Kallgren & Preisler,
1985). According to the principle of acquisition, greater sample size of arguments should result in
stronger effects on attitude judgments. Thus, the more often an argument that an attitude-object
possesses a particular feature is repeated, the stronger the connection will grow between the
Social Cognition and Connectionism 40
attitude-object and this feature (and its associated value).
Nevertheless, it is also possible that people are mislead by the sheer length of a message
(i.e., by use of larger fonts), even if it does not include more persuasive arguments (Wood et al.,
1985). This seems to suggest that superficial characteristics of a message can influence processing,
rather than the arguments themselves. A connectionist network can account for this effect by
assuming that such superficial characteristics of length are often correlated with actual differences
in message length, and so may influence attitude indirectly. More generally, heuristic processing
may sometimes be influenced by issue-irrelevant aspects of the information, and so reflect
qualitative rather than merely quantitative variations in processing (Petty & Wegener, 1999).
Mood. As is evident from the network architecture used to simulate attitude formation, in
our view, mood is just another outcome component that determines attitudes together with other
behavioral approach or avoidance outcomes (i.e., discrepant behaviors). In our simulations, we
tended to equate mood with evaluation (i.e., value node), but admittedly, this might prove to be an
oversimplification (e.g., Perugini & Conner, 2000), useful only as a first approximation of this
issue. Nevertheless, we do believe that mood and affect in general are outcomes or pieces of
information that determine one's attitude and other judgments, although we did not differentiate
strongly between implicit mood priming (Bower, 1981) and explicit processing of mood as part of
relevant information (Schwartz, 1990). We return to this issue of implicit and explicit information
processing in the general discussion.
Fit and Model Comparisons
The simulations that we have reported all replicate the empirical data or theoretical
predictions reasonably well. However, it is possible that this fit is due to some procedural choices
of the simulations rather than conceptual validity. The aim of this section is to demonstrate that
changes in these choices generally do not invalidate our simulations. To this end, we explore a
number of issues, including the localist versus distributed encoding of concepts, and the specific
recurrent network used. We will address each issue in turn.
Distributed Coding
The first issue is whether the nodes in the auto-associative architecture encode localist or
distributed features. Localist features reflect “symbolic” pieces of information, that is, each node
Social Cognition and Connectionism 41
represents a concrete concept. In contrast, in a distributed encoding, a concept is represented by a
pattern of activation across an array of nodes, none of which reflect a symbolic concept but rather
some sub-symbolic micro-feature of it (Thorpe, 1994). Although we most often used a localist
encoding scheme to facilitate this introduction to the most important processing mechanisms
underlying connectionism, we admit that localist encoding is far from realistic. Unlike distributed
coding, it implies that each concept is stored in a single processing unit and, except for differing
levels of activation, is always perceived in the same manner. Given the advantages of distributed
coding, is it possible to replicate our localist simulations with a distributed representation?
To address this question, we reran all localist simulations with a distributed encoding
scheme much like the previous distributed simulations (see Table 12 for details). As can be seen,
all distributed simulations attained a good fit to data and, in all cases, the pattern of results from the
localist simulations was reproduced. This suggests that the underlying principles and mechanisms
that we put forward as being responsible for the major simulation results can be obtained not only
in the more contrived context of a localist encoding, but also in a more realistic context of a
distributed encoding.
Feedforward Model
We claimed earlier that feedforward connections were responsible for replicating most of
the phenomena of interest, with the exception of serial position in impression formation
(Simulation 3) and generalization (Simulation 5). To substantiate this claim, we ran all simulations
with a feedforward pattern associator (McClelland & Rumelhart, 1988) that consists only of
feedforward connections. As can be seen in Table 12, for all simulations except those mentioned
earlier, as predicted, a feedforward architecture did almost equally well as the original simulations.
The only exception was the interaction between heuristic and central processing of attitude
information (Simulation 9) that was less robust, as noted earlier.
This suggests that for most phenomena in social cognition, the feedforward connections in
the network were most crucial. Only for serial position, generalization and interaction of heuristic
and central processing (simulations 3, 5 & 9), the other lateral or backward connections were also
important for obtaining the predicted effects.
Social Cognition and Connectionism 42
Non-linear Recurrent Model
We also claimed earlier that a recurrent model with a linear updating activation algorithm
and a single internal updating cycle (for collecting the internal activation from related nodes) was
sufficient for reproducing the social phenomena of interest.
This contrasts with other social
researchers who used a non-linear activation updating algorithm and many more internal cycles
(Smith & DeCoster, 1998; Read & Montoya, 1999). Are these model features necessary? To
answer this question, we ran all our simulations with a non-linear activation algorithm and 10
internal cycles.
As can be seen from Table 12, although the non-linear model yielded an adequate fit, most
simulations did not improve substantially compared to the original simulations. This suggests that
the present linear activation update algorithm with a single internal cycle is sufficient for simulating
many phenomena in social cognition. This should not come as a surprise. In recurrent simulations
of other issues, such as the formation of semantic concepts, multiple internal cycles were useful to
perform "cleanup" in the network so that the weights between, for instance, a perceptual and
conceptual level of representation were forced to eventually settle into representations that had preestablished conceptual meaning (e.g., Sitton, Mozer & Farah, 2000). Such a distinction between
perceptual and conceptual levels was not made here, and, as a result, multiple internal cycles had no
real function.
The Tensor Product Model
Kashima and his colleagues recently presented a tensor product model, an alternative
connectionist model of person and group impression formation and change (Kashima & Kerekes,
1994; Kashima, Woolcock, Kashima, 2000). As noted earlier, contrary to their claims, the present
recurrent network was able to successfully reproduce the phenomena of impression formation
simulated with their model, including recency and primacy effects. A major difference with the
present recurrent approach, however, is that the tensor product model uses a Hebbian learning
algorithm. Even though this type of learning has the advantage of neurobiological plausibility, it
has the significant disadvantage that it does not reproduce the competition property. Hence, social
cognition phenomena explained by this property such as base-rate neglect, discounting, cognitive
dissonance and so forth, can presumably not be simulated with this model, at least not without
Social Cognition and Connectionism 43
additional assumptions. And indeed, to simulate, for instance, attenuation of recency in impression
formation, this model requires the ad-hoc assumption of different context presentations before and
after a judgment (Kashima & Kerekes, 1994). This assumption was not required with the present
simulations.
General Discussion
In this article, we have presented an overview of a number of major findings in social
cognition and have shown how they might be able to be accounted for within a connectionist
framework. This connectionist perspective offers a novel view on how information could be
encoded in the brain, how it might be structured and activated, and how it could be retrieved and
used for social judgment. This view differs from earlier theories in social cognition which have
relied on metaphors such as associative networks or constraint satisfaction networks with fixed
weights (Kunda & Thagard, 1996; Read & Marcus-Newhall, 1993; Shultz & Lepper, 1996), phaselike integration of information (Gilbert, 1989; but see Trope & Gaunt, 2000) or a formulation in
algebraic or probabilistic terms (Cheng & Novick, 1992, Anderson, 1981; Ajzen, 1991). The
problem is that these various metaphors give a fragmentary account of social cognitive
mechanisms.
In contrast, the connectionist approach proposed in this paper, while it relies on the same
general auto-associative architecture and processing algorithms, has been used in such a way as to
be applicable to a wide-ranging number of phenomena in social cognition. Moreover, we have
shown that this model provides an alternative interpretation of earlier algebraic models in social
psychology (Cheng & Novick, 1992, Anderson, 1981; Ajzen, 1991). In addition, this model can
also account for the learning of social knowledge structures. Hence, this approach could potentially
be used to investigate the development among infants and children of the structures underlying
social cognition.
We have focused to a large extent on the model as a learning device, that is, as a mechanism
for associating patterns that reflect social concepts by means of very elementary learning processes.
One major advantage of a connectionist perspective is that complex social reasoning and learning
can be accomplished by putting together an array of simple interconnected elements, which greatly
enhance the network’s computational power, and by incrementally adjusting the weights of the
connections with the delta learning algorithm. We have demonstrated that this learning algorithm
Social Cognition and Connectionism 44
gives rise to a number of novel properties, among them the acquisition property which accounts for
sample size effects, the competition property accounting for discounting and augmentation, and the
diffusion principle accounting for higher recall for inconsistent information. These properties are
able to explain most of our simulations of social judgment and behavior. In contrast, introductory
textbooks on the auto-associator (e.g., Fausett, 1994; McClelland & Rumelhart, 1988) emphasize
other capacities of the auto-associator including its content-addressable memory, its ability to do
pattern completion (see also Simulation 5) and fault and noise tolerance.
Implications
What are the implications of the present work for social cognitive theories? The key
contribution of this paper is that a wide range of social cognitive phenomena was simulated with
the same overall network model, suggesting that these phenomena are based, at least during early
processing, on the same fundamental information processing principles. Providing a common
framework for these different phenomena will hopefully generate further research and extend to
new areas of social psychology usually seen as too different to be brought under a single theoretical
heading. In addition to the present model’s ability to account for empirical data, it can generate new
hypotheses that can be tested in a classical experimental setting. We briefly discuss some potential
questions that emerge from this model.
Knowledge Acquisition
To what extent is the learning history assumed in our simulations correct?
What
mechanisms and architectural considerations are necessary to preserve the network’s knowledge
base? Perhaps, these answers can in part be answered by laboratory replications of the assumed
learning histories that should reveal equivalences with the (prior) knowledge of participants.
Heuristic versus Central Processing
Our approach does not make a principled distinction between heuristic and central
processing. Quite often, setting activation to a lower or higher (default) level made it possible to
simulate this distinction, suggesting that heuristic and central processing is mainly a matter of
shallow versus focused attention to information. This differential attention gives rise to a
differential emphasis on, for example, prior information versus novel information, and may result in
different judgments. In previous sections, we explained in detail how several reasoning heuristics
Social Cognition and Connectionism 45
could be viewed from this connectionist perspective and these suggestions are immediately open to
empirical tests.
Automatic versus Conscious Reasoning
As noted in the introduction, the present model does not draw a sharp distinction between
automatic and conscious processing, implicit and explicit processing, or associative and symbolic
information processing. While some may view this as a disadvantage of the model, recent research
has revealed that this distinction is far from clear-cut, as unconscious intuitions and insights may
underlie conscious decisions. To resolve this quandary, some researchers (e.g., Smith & DeCoster,
2000) have proposed a distinction between two processing modes: a slow-learning (connectionist)
pattern-completion mode and a more effortful (symbolic) mode that involves explicit symbolically
represented rules and inferences. The present approach seems to suggest that such sharp distinction
is perhaps not necessary.
The Role of Affect
In the final section on attitudes, the role of affect and evaluation was prominent in our
simulations. As noted above, our model makes no distinction between affect as unconscious
priming or explicit information (Schwartz, 1990), although it is clear that this distinction is crucial
to understanding people’s reaction to mood changes. Assuming that evaluation and affect play a
crucial role in attitude judgment, simply inducing a positive or negative mood unobtrusively will
change these judgments in the direction predicted by the model. Recent research findings seem to
support these predictions (Jordens & Van Overwalle, 2001).
Limitations and Future Directions
Given the breadth of social cognition, we inevitably were not able to include many other
interesting findings and phenomena. Perhaps the most interesting area omitted involves group
processes. Connectionist modeling may well help to explain how group identity is created, how
perceptions of group homogeneity is changed, how accentuation of correlated features is enhanced,
how illusory correlation and unrealistic negative stereotypes of minority groups are developed.
These questions are addressed in Van Rooy and Van Overwalle (2001c). However, other
phenomena, such as motivation, love, and violence, remain far beyond the current scope of
connectionist modeling.
Social Cognition and Connectionism 46
While we believe we have shown that a connectionist framework can potentially provide a
parsimonious account of a number of disparate phenomena in social psychology, we are not
suggesting that this is the only valid means of modeling social cognitive phenomena. On the
contrary, we defend a multiple-view position in which connectionism would play a key role but
would co-exist alongside other viewpoints. We think that a strict neurological reductionism is
untenable, especially in personality and social psychology, where it is difficult to see how one could
develop a connectionist model of such high-level abstract concepts as “need for closure”,
“prejudice”, and the like.
There are other limitations to connectionist models. Researchers who may agree with our
overall auto-associative approach, may remain unpersuaded by a specific application of the model
to a particular phenomenon. These applications merely reflect our current thinking and will almost
certainly be replaced by improved models in the future. We believe, however, that the essence of
the approach proposed here will survive.
Our model suggests a number of possible directions for further investigation. Even though
the simple auto-associative model presented here does, indeed, apply to a wide-range of social
phenomena, it would be ridiculous to assume that the whole of higher cognition could be modeled
by auto-association alone. This simple paradigm quickly reveals its limits when we try to apply the
results obtained to other mechanisms.
First, given the importance of attention and motivation in social perception and cognition, it
will ultimately be necessary to incorporate these factors into an improved model. For the time
being, attentional aspects of human information processing are not part of the dynamics of our
network (variations were simply hand-coded as differences in activation states), which focuses
almost exclusively on learning and pattern association. Another issue that remains to be resolved is
how concepts are initially represented when presented to the network. This was not modeled in the
present simulations, but is certainly critical for context-dependent learning and judgment, which
involve combinations of features and context.
Second, another improvement to the present recurrent network might be the inclusion of an
array of hidden (McClelland & Rumelhart, p. 121—126) or exemplar nodes (e.g., Kruschke &
Johansen, 1999) that may potentially increase its power and capacity, for instance, to process nonlinear interactions. Although non-linearity was not an issue in the present simulations, it may
Social Cognition and Connectionism 47
become more critical for combinations of features (e.g., when only a combination of causes
produces an effect), or of features and context.
Third, a more modular architecture will almost certainly be necessary to produce a better fit
of the model to empirical data. For example, one severe limitation of most connectionist models is
known as “catastrophic interference” (McCloskey & Cohen, 1989; Ratcliff, 1990), which is the
tendency of neural networks to forget abruptly and completely previously learned information in the
presence of new input. This limitation is untenable for a realistic model of social cognitive
processes, in general, and for a model of the formation and use of stereotypes, in particular, since
one of the basic properties of stereotypes is their resistance to change in the presence of new
information. In response to such observations, it has been suggested that, to overcome this problem,
the brain developed a dual hippocampal-neocortical memory system in which new information is
processed in the hippocampus and old information is stored and consolidated in the neocortex
(McClelland, McNaughton, & O’Reilly, 1995; Smith & DeCoster, 2000). Various modelers
(French, 1997; Ans & Rousset, 1997) have proposed modular connectionist architectures
mimicking this dual-memory system with one sub-system dedicated to the rapid learning of
unexpected and novel information and the building of episodic memory traces and the other subsystem responsible for slow incremental learning of statistical regularities of the environment and
gradual consolidation of information learned in the first subsystem. There is considerable evidence
for the modular nature of the brain, in particular for the complementary learning roles of
hippocampal and neocortical structures (McClelland, McNaughton & O’Reilly, 1995), the
predominant role of the amygdala in social judgment and perception of emotions (Adolphs, Tranel
& Damasio, 1998), and so forth. It strikes us that the next step in connectionist modeling of social
cognition will involve exploring connectionist architectures built from separate but complementary
systems.
Conclusion
Connectionist modeling of social cognition fits seamlessly into a multilevel integrative
analyses of human behavior (Cacioppo & al., 2000). Given that cognition is intrinsically social,
connectionism will ultimately have to begin to incorporate social constraints into its models. On the
other hand, social psychology will need to be more attentive to the biological underpinnings of
social behavior. Social and biological approaches to cognition can therefore be seen as
Social Cognition and Connectionism 48
complementary endeavors with the common goal of achieving a clearer and deeper understanding
of human behavior. We hope that connectionist accounts of social cognition will provide the
common ground for this exploration.
Social Cognition and Connectionism 49
Appendix
A. The Linear Auto-Associative Model
In an auto-associative network, features and categories, or causes and outcomes are
represented in nodes that are all interconnected. Processing information in this model takes place in
two phases. In the first phase, the activation of the nodes is computed, and in the second phase, the
weights of the connections are updated (see also McClelland & Rumelhart, 1988).
Node Activation
During the first phase of information processing, each node in the network receives
activation from external sources. Because the nodes are all interconnected, this activation is then
spread throughout the network where it influences all other nodes. The activation coming from the
other nodes is called the internal input. Together with the external input, this internal input
determines the final pattern of activation of the nodes, which reflects the short-term memory of the
network.
In mathematical terms, every node i in the network receives external input, termed exti. In
the auto-associative model, every node i also receives internal input inti which is the sum of the
activation from the other nodes j (denoted by aj) in proportion to the weight of their connection, or
inti = (aj * wij),
(1)
for all j  i. Typically, activations and weights range between –1 to +1. The external input
and internal input are then summed to the net input, or
neti = E * exti + I * inti,
(2)
where E and I reflect the degree to which the net input is determined by the external and
internal input respectively. Typically, in a recurrent network, the activation of each node i is
updated during a number of cycles until it eventually converges to a stable pattern that reflects the
network's short-term memory.
According to the linear activation algorithm, the updating of
activation is governed by the following equation:
ai = neti - D * ai,
(3)
where D reflects a memory decay term. In the present simulations, we used only one
internal updating cycle and the parameter values D = I = E = 1.
Given these simplifying
assumptions, the final activation of node i reduces simply to the sum of the external and internal
Social Cognition and Connectionism 50
input, or:
ai = neti = exti + inti
(3')
Weight Updating
After this first phase, the auto-associative model then enters in its second learning phase,
where the short-term activation is consolidated in long term weight changes to better represent and
anticipate future external input. Basically, weight changes are driven by the discrepancy between
the internal input from the last but one updating cycle of the network and the external input
received from outside sources, formally expressed in the delta algorithm (McClelland & Rumelhart,
1988, p. 166):
wij = (exti - inti)aj,
(4)
where wij is the weight of the connection from node j to i, and  is a learning rate that
determines how fast the network learns.
The presence of a feature or a category was typically encoded by setting the external input to
+1, and -1 for opposite features or categories (lower values were also used, see appropriate tables);
otherwise the external activation remained at resting level 0. The weights of the connections were
updated after each trial. At the end of each simulation, the judgment of interest was tested by
turning on the external input of the appropriate nodes and reading off the resulting activation of the
nodes that represent the judgment of interest (see also appropriate tables).
B. Anderson's Averaging Rule and the Delta algorithm
This appendix demonstrates that the delta algorithm converges at asymptote to Anderson's
(1981) averaging rule of impression formation, which expresses a rating about a person as:
rating = isi / i,
(5)
were i represents the weights and si the scale values of the trait.
This proof uses the same logic as Chapman and Robbins (1990) in their demonstration that
the delta algorithm converges to the probabilistic expression of covariation. In line with the
conventional representation of covariation information, person impression information can be
represented in a contingency table with two cells. Cell a represents all cases where the actor is
ascribed a focal trait, while cell b represents all cases where the actor is ascribed the opposite trait.
For simplicity, I use only two trait categories, although this proof can easily be extended to more
Social Cognition and Connectionism 51
categories.
In a recurrent connectionist architecture with localist encoding as used in the text, the target
person j and the trait categories i are each represented by a node, which are connected by adjustable
weights wij. When the target person is present, its corresponding node receives external activation,
and this activation is spread to each trait node. We assume that the overall activation received at
the trait nodes i (or internal activation) after priming the person node, reflects the impression on the
person.
According to the delta algorithm in Equation 4, the weights wij are adjusted proportional to
the error between the actual trait category (represented by its external activation ext) and the trait
category as predicted by the network (represented by its internal activation int). If we substitute in
Equation 4 ext by Anderson's scale values (s1 for the focal trait, and s2 for the opposite trait) and if
we take the default activation for aj (which is 1), then the following equations can be constructed
for the two cells in the contingency table:
For the a cell: wi = (s1 - int),
(6)
For the b cell: wi = (s2 - int).
(7)
The change in overall impression is the sum of Equations 6 and 7, weighted for the
corresponding frequencies a and b, in the two cells, or:
wi = a[(s1 - int)] + b[(s2 - int)]
(8)
These adjustments will continue until asymptote, that is, until the error between actual and
expected category is zero. This implies that at asymptote, the changes will become zero, or wi =
0. Consequently, Equation 8 becomes:
0 = a[(s1 - int)] + b[(s2 - int)]
= a[s1 - int] + b[s2 - int]
= [a * s1 + b * s2] – [a + b]int
so that:
int = [a * s1 + b * s2] / [a + b],
Because the internal activation of the trait nodes reflects the trait impression on the person,
this can be rewritten in Anderson's terms as:
impression = fisi / fi
(9)
where f represents frequencies with which a person and the traits co-occur. As can be seen,
Social Cognition and Connectionism 52
Equation 6 has the same format as Equation 1. This demonstrates that the delta algorithm predicts a
weighted averaging function at asymptote for making overall impression judgments, where
Anderson's weights  are determined by the frequencies by which person and traits are presented
together.
C. Ajzen's Expectancy-Value model and the Delta algorithm
This appendix demonstrates that the delta algorithm converges at asymptote to the
expectancy-value model of attitude formation by Ajzen (1991). According to Ajzen's (1991)
expectancy-value model, an attitude is formed by summing the multiplicative combination of (a)
the strength of a salient belief that a behavior will produce a given outcome and (b) the subjective
evaluation of this outcome, or (Ajzen, 1991, p. 191):
attitude  biei,
(10)
were bi represents the strength of the belief and ei the evaluation. Beliefs and evaluations
are typically scored on 7- point scales. However, because there is "no rational a priori criterion we
can use to decide how the belief and evaluation scales should be scored" (Ajzen, 1991, p. 193), the
preceding formula can be normalized by diving it by the mean belief strengths, or:
attitude  biei / bi
(11)
Using the same logic of the proof above, it can be shown that the delta algorithm results in
asymptote in an equation similar to Equation 9, or:
attitude = fiei / fi
(12)
where f represent the frequencies that the attitude-object leads to a given outcome (which
we assume determine the belief strengths b), and where e represents the activation values +1
(desirable outcomes), -1 (undesirable outcomes) or 0 (neutral). The equivalence between Equations
4 and 12 demonstrates that the delta algorithm predicts a (normalized) multiplicative function at
asymptote for making attitude judgments, where the strength of the beliefs are determined by the
frequencies by which the attitude-object and outcomes co-occur.
Social Cognition and Connectionism 53
References
Adolphs, R., & Damasio, A. (2001). The interaction of affect and cognition: A neurobiological
perspective. In J.P. Forgas (Ed.). Handbook of affect and social cognition (pp. 27-49).
Mahwah, NJ: Lawrence Erlbaum Associates.
Adolphs, R., Tranel, D., & Damasio, A. (1998). The human amygdala in social judgment. Nature,
393, 470-474.
Ajzen, I. & Madden, T.J. (1986). Prediction of goal-directed behavior: Attitudes, intentions, and
perceived behavioral control. Journal of Experimental Social Psychology, 22, 453—474.
Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human Decision
Processes, 50, 179—211.
Ajzen, I., & Fishbein, M. (1980).
Understanding attitudes and predicting social behavior.
Englewood Cliffs, NJ: Prentice-Hall.
Allan, L. G. (1993). Human contingency Judgments: Rule based or associative? Psychological
Bulletin, 114, 435-448.
Allison, T., Puce, A., & McCarthy, G. (2000). Social perception from visual cues: role of the STS
region. Trends in Cognitive Sciences, 4, 267-278.
Anderson, J. R. (1976). Language, memory and thought. Hillsdale, NJ: Erlbaum.
Anderson, N. H. (1967). Averaging model analysis of set size effect in impression formation.
Journal of Experimental Psychology, 75, 158—165.
Anderson, N. H. (1979). Serial position curves in impression formation. Journal of Experimental
Psychology, 97, 8—12.
Anderson, N. H. (1981). Foundations of information integration theory. New York: Academic
Press.
Anderson, N. H., & Farkas, A. J. (1973). New light on order effect in attitude change. Journal of
Personality and Social Psychology, 28, 88—93.
Anderson, S. M. & Cole, S. W. (1990). "Do I know you?": The role of significant other in general
social perception. Journal of Personality and Social Psychology, 59, 384—399.
Ans, B., & Rousset, S. (1997). Avoiding catastrophic forgetting by coupling two reverberating
neural networks. Académie des Sciences de la vie, 320, 989-997.
Asch, S. E. & Zukier, H. (1984). Thinking about persons. Journal of Personality and Social
Psychology, 46, 1230—1240.
Asch, S. E. (1946).
Forming impressions of personality.
Journal of Abnormal and Social
Psychology, 41, 258—290.
Baker, A. G., Berbier, M. W., & Vallée-Tourangeau, F. (1989). Judgments of a 2 x 2 contingency
Social Cognition and Connectionism 54
table: sequential processing and the learning curve. The Quarterly Journal of Experimental
Psychology, 41B, 65—97.
Bargh, J. A. & Thein, R. D. (1985). Individual construct accessibility, person memory, and the
recall-judgment link: The case of information overload. Journal of Personality and Social
Psychology, 49, 1129—1146.
Betsch, T., Plessner, H., Schwieren, C., & Gütig, R. (2001). Personality and Social Psychology
Bulletin, 27, 242—253.
Bohner, G., Moskowitz G. B., & Chaiken, S. (1995).
The interplay between heuristic and
systematic processing of social information. European Review of Social Psychology, 6,
33—68.
Bohner, G., Ruder, M., & Erb, H.-P. (1999).
When expertise backfires: Contrast versus
assimilation in the interplay of heuristic and systematic processing.
Unpublished
manuscript.
Bower, G. H. (1981) Emotional mood and memory. American Psychologist, 36, 129—148.
Busemeyer, J. R. & Myung, I. J. (1988). A new method for investigating prototype learning.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 3—1.
Cacioppo, J.T., Berntson, G.G., Sheridan, J.F., & McClintock M.K. (2000). Multilevel integrative
analyses of human behavior: social neuroscience and the complementing nature of social
and biological approaches. Psychological Bulletin, 126, 829-843.
Chaiken, S. (1980). Heuristic versus systematic information processing and the use of source
versus message cues in persuasion. Journal of Personality and Social Psychology, 39,
752—766.
Chaiken, S. (1987). The heuristic model of persuasion. In M. P. Zanna, J. M. Olson, & C. P.
Herman (Eds.). Social influence: The Ontario Symposium (Vol. 5., pp. 3—39). Hillsdale,
NJ: Erlbaum.
Chaiken, S., & Maheswaran, D. (1994). Heuristic processing can bias systematic processing:
effects of source credibility, argument ambiguity, and task importance on attitude judgment.
Journal of Personality and Social Psychology, 66, 460—473.
Chaiken, S., Duckworth, K. L., & Darke, P. (1999). When parsimony fails… Psychological
Inquiry, 10, 118—123.
Chaiken, S., Liberman, A. & Eagly, A. H. (1989). Heuristic and systematic information processing
within and beyond the persuasion context.
In J. S. Uleman & J. A. Bargh (Eds.)
Unintended thought (pp. 212—252). New York, NY: Guilford.
Chapman, G. B. & Robbins, S. J. (1990).
Cue interaction in human contingency judgment.
Social Cognition and Connectionism 55
Memory and Cognition, 18, 537—545.
Chen, S. & Chaiken, S. (1999). The Heuristic-systematic model in its broader context. In S.
Chaiken & Y. Trope (Eds.). Dual-process theories in social psychology (pp. 73—96). New
York, NY: Guilford Press.
Cheng, P. W., & Novick, L. R. (1992). Covariation in natural causal induction. Psychological
Review, 99, 365—382.
Cooper, J. & Fazio, R. H. (1984). A new look at dissonance theory. In L. Berkowitz (Ed.).
Advances in experimental social psychology (Vol. 17, pp. 229-266). New York: Academic
Press.
Dickinson, A. & Burke, J. (1996).
Within-compound associations mediate the retrospective
revaluation of causality judgments. Quarterly Journal of Experimental Psychology, 49B,
60-80.
Dreben, E. K., Fiske, S. T., & Hastie, R. (1979). The independence of evaluative and item
information: Impression and recall order effects in behavior-based impression formation.
Journal of Personality and Social Psychology, 37, 1758-1768.
Eagly, A. H. & Chaiken, S. (1993). The psychology of Attitudes. San Diego, CA: Harcourt Brace.
Ebbesen, E. B., & Bowers, R. J. (1974) Proportion of risky to conservative arguments in a group
discussion and choice shifts. Journal of Personality and Social Psychology, 29, 316—327.
Fausett, L. (1994). Fundamentals of neural networks: architectures, algorithms and applications.
Englewood Cliffs, NJ: Prentice-Hall.
Fazio, R. H. (1990). Multiple processes by which attitudes guide behavior: the MODE model as an
integrative framework. In M. P. Zanna (Ed.) Advances in Experimental Social Psychology
(vol. 13, pp. 75—109). San Diego, CA: Academic Press.
Festinger, L. (1957) A theory of cognitive dissonance. Evanston, IL: Row, Peterson.
Fiedler, K. (1996). Explaining and simulating judgment biases as an aggregation phenomenon in
probabilistic, multiple-cue environment. Journal of Personality and Social Psychology, 103,
193-214.
Fiedler, K., Walther, E. & Nickel, S. (1999).
The auto-verification of social hypotheses:
Stereotyping and the power of sample size. Journal of Personality and Social Psychology,
77, 5-18.
Fishbein, M., & Ajzen, I. (1975). Belief attitude, intention and behavior an introduction to theory
and research. London, UK: Addison-Wesley.
Försterling, F. (1989). Models of covariation and attribution: How do they relate to the analogy of
analysis of variance? Journal of Personality and Social Psychology, 57, 615—625.
Social Cognition and Connectionism 56
Försterling, F. (1992). The Kelley model as an analysis of variance analogy: How far can it be
taken? Journal of Experimental Social Psychology, 28, 475—490.
French, R. (1997). Pseudo-recurrent connectionist networks: An approach to the “sensitivity–
stability” dilemma. Connection Science, 9, 353-379.
Gilbert, D. T. & Malone, P. S. (1995). The correspondence bias. Psychological Bulletin, 117, 21—
38.
Gilbert, D. T. (1989). Thinking lightly about others: Automatic components of the social inference
process In J. S. Uleman & J. A. Bargh (Eds.) Unintended thoughts: Limits of awareness,
intention, and control (pp. 189-211). New York: Guilford.
Gluck, M. A. & Bower, G. H. (1988). From conditioning to category learning: An adaptive network
model. Journal of Experimental Psychology: General, 117, 227—247.
Gluck, M. A. (1992). Stimulus sampling and distributed representations in adaptive network
theories of learning. In A. F. Healy, S.M. Kosslyn, R. M. Shiffrin (Eds.) From learning
theory to connectionist theory: Essays in honor of William K. Estes (pp. 169—199).
Hillsdale, NJ: Erlbaum.
Gollwitzer, P. M. (1990) Action phases and mind-sets. In E. T. Higgins and R. M. Sorrentino
(Eds.), Handbook of motivation and cognition: Foundations of social behavior (Vol. 2, pp.
53—92). New York: Guilford Press.
Graham, S. (1999). Retrospective revaluation and inhibitory associations: Does perceptual learning
modulate our perceptions of the contingencies between events?
Quarterly Journal of
Experimental Psychology, 52B, 159-185.
Hamilton, D. L., Driscoll, D. M., & Worth, L. T. (1989). Cognitive organization of impressions:
Effects of incongruency in complex representations. Journal of Personality and Social
Psychology, 56, 925—939.
Hamilton, D. L., Katz, L. B., Leier, V. O. (1980).
Cognitive representation of personality
impressions: Organizational processes in first impression formation. Journal of Personality
and Social Psychology, 39, 1050—1063.
Hamilton, D.L., Dugan, P.M., & Trollier, T.K. (1985). The formation of stereotypic beliefs:
Further evidence for distinctiveness-based illusory correlation. Journal of Personality and
Social Psychology, 48, 5-17.
Hansen, R. D. & Hall, C. A. (1985). Discounting and augmenting facilitative and inhibitory forces:
the winner takes all. Journal of Personality and Social Psychology, 49, 1482--1493.
Hastie, R. & Kumar, P. A. (1979) Person memory: Personality traits as organizing principles in
memory for behaviors. Journal of Personality and Social Psychology, 37, 25—38.
Social Cognition and Connectionism 57
Hastie, R. (1980). Memory for behavioral information that confirms or contradicts a personality
impression. In R. Hastie, T. M. Ostrom, E. B. Ebbesen, R. S. Wyer, D. L. Hamilton, & D.
E. Carlston (Eds.). Person Memory: The cognitive basis of social perception (pp. 155—
177). Hillsdale, NJ: Erlbaum.
Hintzmann, D. L. (1986). "Schema abstraction" in a multi-trace memory model. Psychological
Review, 93, 411—428.
Ito, T.A., & Cacioppo, J.T. (2001). Affect and attitudes: A social neuroscience approach. In J.P.
Forgas (Ed.) Handbook of affect and social cognition (pp. 50-74). Mahwah, NJ: Lawrence
Erlbaum Associates.
Kahneman, D., Slovic, P., & Tversky, A. (1982). Judgments under uncertainty: Heuristics and
biases. Cambridge, England: Cambridge University Press.
Kashima, Y, & Kerekes, A. R. Z. (1994). A distributed memory model of averaging phenomena in
person impression formation. Journal of Experimental Social Psychology, 30, 407—455.
Kashima, Y., Woolcock, J., & Kashima, E. S. (2000). Group impression as dynamic configurations:
The tensor product model of group impression formation and change. Psychological
Review, 107, 914-942
Kelley, H. H. (1972). Causal schemata and the attribution process. In E. E. Jones, D. E. Kanouse,
H. H. Kelley, R. E. Nisbett, S. Valins & B. Weiner (Eds.) Attribution: Perceiving the causes
of behavior (pp. 151-174). Morristown, NJ: General Learning Press.
Kruglanski, A. W., Schwartz, S. M., Maides, S., & Hamel, I. Z. (1978). Covariation, discounting,
and augmentation : Towards a clarification of attributional principles.
Journal of
Personality, 76, 176–189.
Kruschke, J. K. (1996). Base rates in category learning. Journal of Experimental Psychology:
Learning, Memory and Cognition, 22, 3—26.
Kruschke, J. K., & Johansen, M. K. (1999). A model of probabilistic category learning. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 25, 1083—1119.
Kunda, Z., & Thagard, P. (1996). Forming impressions from stereotypes, traits, and behaviors: A
parallel-constraint-satisfaction theory. Psychological Review, 103, 284-308
Labiouse, C. L. & French, R. M. (2001).
A connectionist model of person perception and
stereotype formation. In R. French & J. Sougné (Eds.) Proceedings of the sixth Neural
Computation and Psychology Workshop: Learning, Development, and Evolution, pp.209218. London: Springer Verlag.
Linder, D.E., Cooper, J. & Jones, E.E. (1967). Decision freedom as a determinant of the role of
incentive magnitude in attitude change. Journal of Personality and Social Psychology, 6,
Social Cognition and Connectionism 58
245-254.
Macrae, C. N., Hewstone, M., & Griffiths, R. J. (1993).
Processing load and memory for
stereotype-based information. European Journal of Social Psychology, 23, 77-87.
Maheswaran, D. & Chaiken, S. (1991).
Promoting systematic processing in low-motivation
settings: Effect of incongruent information on processing and judgment.
Journal of
Personality and Social Psychology, 61, 13—25.
Manis, M., Dovalina, I., Avis, N. E., & Cardoze, S. (1980). Base rates can affect individual
predictions. Journal of Personality and Social Psychology, 38, 231—248.
McClelland, J. L. & Rumelhart, D. E. (1985). Distributed memory and the representation of
general and specific information. Journal of Experimental Psychology, 114, 159—188.
McClelland, J. M. & Rumelhart, D. E. (1988). Explorations in parallel distributed processing: A
handbook of models, programs and exercises. Cambridge, MA: Bradford.
McClelland, J., McNaughton, B., & O’Reilly, R. (1995). Why there are complementary learning
systems in the hippocampus and neocortex: Insights from the successes and the failures of
connectionist models of learning and memory. Psychological Review, 102, 419-457.
McCloskey, M., & Cohen N.J. (1989). Catastrophic interference in connectionist networks: the
sequential learning problem. The Psychology of Learning and Motivation, 24, 109-165.
Medin, D. L. & Edelson, S. M. (1988). Problem structure and the used of base-rate information
form experience. Journal of Experimental Psychology, General, 117, 68—85.
Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological
Review, 85, 207—238.
Morris, M. W. & Larrick, R. P. (1995). When one cause casts doubt on another: A normative
analysis of Discounting in causal attribution. Psychological Review, 102, 331—355.
Moskowitz, G. B., & Skurnik, I. W. (1999). Contrast effects as determined by the type of prime:
trait versus exemplar primes initiate processing strategies that differ in how accessible
constructs are used. Journal of Personality and Social Psychology, 76, 911—927.
Nisbett, R.E., & Wilson, T.D. (1977). Telling more than we can know: Verbal reports on mental
processes. Psychological Review, 84, 231-259.
Nosofsky, R. M., Kruschke, J. K. & McKinley, S. C. (1992). Combining exemplar-based category
representations and connectionist learning rules.
Journal of Experimental Psychology:
Learning, Memory and Cognition, 18, 211—233.
Nosofsky, R.M. (1986). Attention, similarity, and the identification-categorization relationship.
Journal of Experimental Psychology: General, 115, 39-57.
Perugini, M. & Conner, M. (2000). Predicting and understanding behavioral volitions: The
Social Cognition and Connectionism 59
interplay between goals and behaviors. Journal of Experimental Social Psychology, 30,
705—731.
Petty, E. & Wegener, D. T. (1999).
The elaboration likelihood model: Current status and
controversies. In S. Chaiken & Y. Trope (Eds.). Dual-process theories in social psychology
(pp. 41—72). New York, NY: Guilford Press.
Petty, R. E. & Cacioppo, J. T. (1986). The elaboration likelihood model of persuasion. In L.
Berkowitz (Ed.). Advances in experimental social psychology (Vol. 19, pp. 123—205). San
Diego, CA: Academic Press.
Petty, R. E., Cacioppo, J. T. (1984). The effects of involvement on responses to argument quantity
and quality: Central and peripheral routes to persuasion. Journal of Personality and Social
Psychology, 46, 69—81.
Petty, R.E., Cacioppo, J. T., & Goldman, R. (1981). Personal involvement as a determinant of
argument-based persuasion. Journal of Personality and Social Psychology, 41, 847—855.
Phelps, E.A., O’Connor, K.J., Cunningham, W.A., Funayama, S., Gatenby, C., Gore, J.C., & Banaji
M.R. (2000). Performance on indirect measures of race evaluation predicts amygdala
activation. Journal of Cognitive Neuroscience, 12, 729-738.
Ratcliff, R. (1990). Connectionist models of recognition memory: constraints imposed by learning
and forgetting functions. Psychological Review, 97, 285-308.
Read, S. J. & Marcus-Newhall, A. (1993) Explanatory coherence in social explanations: A parallel
distributed processing account. Journal of Personality and Social Psychology, 65, 429—447.
Read, S. J., & Montoya, J. A. (1999). An autoassociative model of causal reasoning and causal
learning: Reply to Van Overwalle's critique of Read and Marcus-Newhall (1993). Journal
of Personality and Social Psychology, 76, 728—742.
Reeder, G. D., & Brewer, M. B. (1979).
A schematic model of dispositional attribution in
interpersonal perception. Psychological Review, 86, 61—79.
Rescorla, R. A. & Wagner, A. R. (1972). A theory of Pavlovian conditioning: Variations in the
effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy
(Eds.) Classical conditioning II: Current research and theory (pp. 64–98). New York:
Appleton-Century-Crofts.
Rosch, E. H. (1978) Principles of categorization. In E. H. Rosch & B. B. Lloyds (Eds.), Cognition
and categorization (pp. 27—48). Hillsdale, NJ: Erlbaum.
Rosenfield, D. & Stephan, W. G. (1977). When discounting fails: An unexpected finding. Memory
and Cognition, 5, 97-102.
Rumelhart, D. E., & McClelland, J. L. (1996). Parallel Distributed Processing: Explorations In The
Social Cognition and Connectionism 60
Microstructure of Cognition - Vol. 1: Foundations. Cambridge, MS: The Mit Press.
Sarle, W. S. (1994). Neural networks and statistical models. Proceedings of the nineteenth annual
SAS users group international conference.
Schwarz, N. (1990). Feelings as information: Informational and motivational functions of affective
states. In E.T. Higgins & R. Sorrentino (Eds.), Handbook of motivation and cognition:
Foundations of social behavior (Vol. 2). New York: Guilford.
Shacklee, H. & Fischhoff, B. (1982). Strategies of information search in causal analysis. Memory
and Cognition, 10, 520-530.
Shanks, D. (1992). Connectionist accounts of the inverse base-rate effect in categorization.
Connection Science, 4, 3—18.
Shanks, D. R. (1985). Forward and backward blocking in human contingency judgment. Quarterly
Journal of Experimental Psychology, 37b, 1—21.
Shanks, D. R. (1987). Acquisition functions in contingency judgment. Learning and Motivation,
18, 147—166.
Shanks, D. R. (1995). Is human learning rational? Quarterly Journal of Experimental Psychology,
48a, 257—279.
Shanks, D.R., Lopez, F. J., Darby, R. J., Dickinson, A. (1996). Distinguishing associative and
probabilistic contrast theories of human contingency judgment.
In D. R. Shanks, K. J.
Holyoak, & D. L. Medin (Eds.) The psychology of learning and motivation (Vol. 34, pp.
265—311). New York, NY: Academic Press.
Shultz, T. & Lepper, M. (1996).
Cognitive dissonance reduction as constraint satisfaction.
Psychological Review, 2, 219-240.
Siebler, F., Bohner, G. & Weinerth, T. (1998). Simulation of implicit and explicit processes in
parallel-constraint-satisfaction networks. Unpublished manuscript.
Sitton, M., Mozer, M. C., & Farah, M. J. (2000). Superadditive effects of multiple lesions in a
connectionist architecture: Implications for the neuropsychology of optic aphasia.
Psychological Review, 107, 709—734.
Smith E. R., & Zárate, M. A. (1992). Exemplar-based model of social judgment. Psychological
Review, 99, 3—21.
Smith, E. R. & DeCoster, J. (1998). Knowledge acquisition, accessibility, and use in person
perception and stereotyping: Simulation with a recurrent connectionist network. Journal of
Personality and Social Psychology, 74, 21—35.
Smith, E. R. (1996). What to connectionism and social psychology offer each other? Journal of
Personality and Social Psychology, 70, 893-912.
Social Cognition and Connectionism 61
Smith, E.R. & DeCoster, J. (2000). Associative and rule-based processing: A connectionist
interpretation of dual-process models. In S. Chaiken & Y. Trope (Eds.) Dual-process
theories in social psychology (pp. 323—338). London, UK: Guilford.
Srull, T. K. (1981). Person Memory: Some tests of associative storage and retrieval models.
Journal of Experimental Psychology: Human Learning and Memory, 7, 440—463.
Srull, T. K., Lichtenstein, M., Rothbart, M. (1985). Associative storage and retrieval processes in
person memory. Journal of Experimental Psychology: Learning, Memory, and Cognition,
11, 316—345.
Stangor, C. & McMillan, D. (1992). Memory for expectancy-congruent and expectancy-incongruent
information: A review of the social and social developmental literatures. Psychological
Bulletin, 111, 42—61.
Stangor, C., & Duan, C. (1991). Effects of multiple task demands upon memory for information
about social groups. Journal of Experimental Social Psychology, 27, 357—378.
Stapel, D. A., Koomen, W. & J. van der Pligt (1997). Categories of category accessibility: the
impact of trait concept versus exemplar priming on person judgments.
Journal of
Experimental Social Psychology, 33, 47—76.
Stewart, R. H. (1965).
Effect of continuous responding on the order effect in personality
impression formation. Journal of Personality and Social Psychology, 1, 161—165.
Thagard, P. (1989). Explanatory coherence. Behavioral and Brain Sciences, 12, 435-467.
Thorpe (1994). Localized versus distributed representations. In M. A. Arbib (Ed.) Handbook of
brain theory and neural networks (pp. 949-952). Cambridge, MA: MIT Press.
Tobena, A., Marks, I., & Dar, R. (1999). Advantages of bias and prejudice: an exploration of their
neurocognitive templates. Neuroscience and Biobehavioral Reviews, 23, 1047-1058.
Trope, Y. & Gaunt, R. (2000). Processing alternative explanations of behavior: Correction or
integration? Journal of Personality and Social Psychology, 79, 344—354.
Van Overwalle, F. & Timmermans, B. (2000) Discounting and augmentation in attribution: The
role of connections between causes. Manuscript submitted for publication.
Van Overwalle, F. & Timmermans, B. (2001). Learning about an Absent Cause: Discounting and
Augmentation of Positively and Independently Related Causes. In R. French & J. Sougné
(Eds.) Proceedings of the sixth Neural Computation and Psychology Workshop: Evolution,
Learning, and Development. London: Springer Verlag.
Van Overwalle, F. (1996). The relationship between the Rescorla-Wagner associative model and
the probabilistic joint model of causality. Psychologica Belgica, 36, 171-192.
Van Overwalle, F. (1998) Causal Explanation as Constraint Satisfaction: A Critique and a
Social Cognition and Connectionism 62
Feedforward Connectionist Alternative. Journal of Personality and Social Psychology, 74,
312-328.
Van Overwalle, F., & Jordens, K. (2001) A Feedforward Connectionist Model of Cognitive
Dissonance: An Alternative to Shultz and Lepper (1996).
Manuscript submitted for
publication.
Van Overwalle, F., & Van Rooy, D. (1998). A Connectionist Approach to Causal Attribution. In S.
J. Read & L. C Miller (Eds.) Connectionist models of Social Reasoning and Social
Behavior. New York: Erlbaum.
Van Overwalle, F. & Van Rooy, D. (2001a). When more observations are better than less : A
connectionist account of the acquisition of causal strength. European Journal of Social
Psychology, 31, 155-175.
Van Overwalle, F., & Van Rooy, D. (2001b). How one cause discounts or augments another: A
connectionist account of causal competition. Personality and Social Psychology Bulletin, in
press.
Van Overwalle, F. & Van Rooy, D. (2001c). A recurrent connectionist model of biases in group
judgments. Manuscript submitted for publication.
Van Overwalle, F., Drenth, T. & Marsman, G. (1999). Spontaneous trait inferences: Are they
linked to the actor or to the action? Personality and Social Psychology Bulletin, 25, 450462.
Wells, G. L. & Ronis, D. L. (1982). Discounting and augmentation: Is there something special
about the number of causes? Personality and Social Psychology Bulletin, 8, 566—572.
Wood, W., Kallgren, C. A., Preisler, R. M. (1985). Access to attitude-relevant information in
memory as a determinant of persuasion: The role of message attributes.
Experimental Social Psychology, 21, 73—85.
Journal of
Social Cognition and Connectionism 63
Table 1
Overview of the Simulations
Nr.
Topic
Evidence / Prediction
Major Processing
Principle
1
Categorization
Gluck & Bower, 1988, exp. 1
Competition
2
Impression formation
Stewart, 1965
Acquisition
3
Serial position
Dreben, Fiske & Hastie, 1979
Acquisition
4
Inconsistent
Hamilton, Driscoll, & Worth, 1980, exp. 3
Diffusion
Smith & DeCoster, 1998, sim. 1 & 2
Spreading
information
5
Generalization
of Internal
Activation
6
Assimilation &
Stapel, Koomen & van der Pligt, exp. 3
Contrast
7
Causal Attribution
Acquisition
Competition
Van Overwalle & Van Rooy, 2001, exp. 1
Acquisition
Competition
8
Attitude Formation
Ajzen, 1991
Acquisition
9
Dual-Process Models
Chaiken & Maheswaran, 1994
Acquisition
10
Cognitive Dissonance
Linder, Cooper & Jones, 1967
Competition
Social Cognition and Connectionism 64
Table 2
Learning Experiences in Categorization (Simulation 1)
Features
——————————————————————————————-———————Dutch
less-sophisticated
refined
French
Categories
———————————————Flemish
Walloon
Flemish (Rare) Category
#10
#5
#5
#10
1
1
1
1
0
1
1
1
R
0
0
1
1
0
0
0
1
1
1
1
1
0
0
0
0
1
1
1
1
0
0
0
0
1
1
1
1
Walloon (Common) Category
#30
#15
#15
#30
1
0
0
0
1
1
0
0
1
1
1
0
Test
Features
Dutch
less-sophisticated
refined
French
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
?
?
?
?
–?
–?
–?
–?
?
?
?
?
?
?
?
?
1
0
0
1
Prototype
Flemish
Walloon
Note. Simplified version of the experimental design of Gluck & Bower (1988); Cell entries denote external
activation; R=Randomized order; #=frequency of trial; the pattern of features were generated according to
the following probabilities: Given a category, the category's own perfect feature was present 100% of the
time and its imperfect feature 67%, the other category's imperfect feature 50% and its perfect feature 33%.
Social Cognition and Connectionism 65
Table 3
Impression Formation: Recency after Reversal of Trait-implying Information (Simulation 2)
Features
——————————————————————
person
context
Category
———————————
trait
High - Low Presentation Order
#4 High trait
#4 Low trait
1
1
1
1
+1
–1
Low - High presentation Order
#4 Low trait
#4 High trait
1
1
1
1
–1
+1
Test
1
0
?
Note. Schematic representation of the experimental design of Stewart (1965); High=adjective implies trait;
Low=adjective implies opposite trait; Cell entries denote external activation; #=number of trials.
Social Cognition and Connectionism 66
Table 4
Impression Formation: Recency and Primacy in Serial Position Weights (Simulation 3)
Features
——————————————————————
person
context
Category
———————————
trait
Confirmatory Information
#4 High
1
1
+1
Mixed Information
#3 High
#1 Lowa
1
1
1
1
+1
–1
Test
1
0
?
Note. Schematic representation of the experimental design of Dreben, Fiske & Hastie (1979) illustrated here
for the four trial condition; High=adjective implies trait; Low=adjective implies opposite trait; Cell entries
denote external activation; Initial weights were set at .10; #=number of trials.
a This disconfirming trial is presented at position 1, 2, 3, or 4 of the four trial series (here it is shown at
position 4), and ratings are then compared with judgments from the confirmatory condition at the same
serial position.
Social Cognition and Connectionism 67
Table 5
Impression Formation: Memory for inconsistent Information (Simulation 4)
person
#1
#1
#1
#1
#1
consistent
consistent
inconsistent
consistent
consistent
1a
1a
1a
1a
1a
Trait
——————————————
common violent
1a
1a
0
1a
1a
0
0
1a
0
0
Behavioral Exemplars
—————————————————————————
consistent inconsistent
1
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
1
0
0
Test
Recall
consistent
inconsistent
1
1
1
1
1
1
? ? ? ?
0 0 0 0
0
?
Biased Recognition
consistent
inconsistent
?
?
?
?
0
0
1 1 1 1
0 0 0 0
0
1
Note. Schematic representation of the experimental design of Hamilton, Katz & Leirer (1980, experiment
3), illustrated for a fictitious 4 / 1 distribution of consistent versus inconsistent behaviors. In the original
experiment, the distribution was 10 / 1, and the inconsistent statement was given at position 2, 6 or 10 of the
list (in most similar experiments order was randomized). Cell entries denote external activation; #=number
of trials.
a Activation set to 0.10 under memorizing instructions.
Social Cognition and Connectionism 68
Table 6
Assimilation: Exemplar and Group Inferences (Simulation 5)
Exemplar Features
————————————————————————————
E1
E2
E3
E4
E5
Group Features
———————————————
G1
G2
G3
Background Knowledge
R
#100 context
#20 exemplar
0̃
̃
0̃
̃
0̃
̃
0̃
̃
0̃
̃
0̃
̃
0̃
–̃
0̃
+̃

0
?
0
0
+
0
–
0
?
Test
exemplar
group

0

0

0
Note. Schematic representation of assimilation of exemplar and group stereotypes; Each feature E or G is
represented by 5 nodes; Cell entries denote external activation; for the exemplar features,  reflects a
randomly drawn Normal distributed pattern with M=0 & SD=.5 (identical across all trials); for the group
features, + reflects M=.50 and – reflects M=-.5; R=Randomized order; #=number of trials. For reasons of
clarity, the other exemplar features E6 to E15 representing two other exemplars (each #20 trials) are not
shown.
~ Noise added randomly at each trial with Normal distribution of M=0 & SD=.5
Social Cognition and Connectionism 69
Table 7
Assimilation and Contrast (Simulation 6)
Exemplars
———————————————————————
person
Hitler
Gandhi
Trait Categories
————————————————
hostile/
violent
nice
Background Knowledge
#10
#10
R #10
#10
#10
Hitler
Gandhi
average person
average person
traits
0
0
+
+
0
+
0
0
0
0
0
+
0
0
0
+
0
+
0
–
0
+
0
+
+
+̃
0
0
0
0
+̃̃
0
0
0
0
+̃
0
0
0
0
+̃̃
0
0
0
?
Priming (each condition only once)
#1
#1
#1
#1
Hitler
Gandhi
hostile
nice
+̃
+̃
+̃
+̃
Test
+
Note. Schematic representation of prior knowledge acquisition and experimental design of Stapel, Koomen
& Van der Plight (1997, Experiment 3); Each feature/category is represented by 5 nodes; Cell entries denote
external activation with + and – reflecting a randomly drawn Normal distributed pattern with M=+.5/-.5 &
SD=.5 (identical across all trials; the simulation was run for 5 such activation patterns and results were
averaged); R=Randomized order; #=number of trials.
~ Noise added randomly at each trial with Normal distribution of M=0 & SD=.5
Social Cognition and Connectionism 70
Table 8
Forward Discounting of a Novel Cause in function of the Sample Size of a Known Cause (Simulation 7)
Causes
——————————————————————
known
novel
(Ann)
(Troy)
Small Sample Size
#1
#5
Large Sample Size
#5
#5
Outcome
———————————
1
1
0
1
1
1
1
1
0
1
1
1
Test
Known (Ann)
Novel (Troy)
1
0
0
1
?
?
Note. Schematic representation of the experimental design of Van Overwalle & Van Rooy (2001); Cell
entries denote external activation; #=number of trials.
Social Cognition and Connectionism 71
Table 9
Attitude Formation (Simulation 8)
Causal Factors
————————————————
car
bicycle
bus
Outcomes
—————————————————————-——
fast
dry pollutes
value
Car
#10
R #10
#10
1
1
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
1
1
1
–1
#5
#10
0
0
Bicycle
1
0
1
0
1
0
0
–1
0
0
1
–1
#10
R #5
#5
0
0
0
0
0
0
1
1
1
–1
0
0
0
1
0
0
0
1
–1
1
–1
0
0
1
0
0
0
0
0
0
0
0
0
?
?
?
R
Bus
Test
attitude toward car
attitude toward bicycle
attitude toward bus
1
0
0
0
1
0
Note. Schematic representation of attitude formation on the basis of beliefs on outcome consequences and
value (likeability of consequences; cf., theory of Planned behavior); Cell entries denote external activation;
R=Randomized order; #=number of trials.
Social Cognition and Connectionism 72
Table 10
Dual Processes in Attitude Formation (Simulation 9)
Causal Factors
———————————————
product source
Outcomes
———————————————————————-——
featuresa
value
#20 Low credibility
#20 High credibility
Prior Knowledge on Source
1
1
1
1
0 0 0 0
0 0 0 0
.1
1
#2
R #2
#2
1
1
1
Strong arguments
1
1
1
1 0 0 0
0 1 0 0
0 0 1 0
1
1
0
#2
R #2
#2
1
1
1
Weak arguments
1
1
1
1 0 0 0
0 1 0 0
0 0 1 0
0
0
.1
#2
R #2
#2
#2
1
1
1
1
Ambiguous arguments
1
1
1
1
1
0
0
0
0
0
0
1
0
1
0
.1
0 0 0 0
?
0
1
0
0
0
0
1
0
Test
attitude toward product
1
0
Note. Simplified representation of attitude formation on the basis of heuristics (source credibility) and
systematic processing (argument quality; Chaiken & Maheswaran, 1994); Cell entries denote external
activation; R=Randomized order; #=number of trials (for heuristic processing, the number of trials for all
arguments was set to 1 with all activation levels divided by 10).
a The first two features of the product are of high importance, the last two of low importance (as can be
seen from the value component).
Social Cognition and Connectionism 73
Table 11
Cognitive Dissonance following Induced Compliance (Simulation 10)
Causal factors
—————————————————————
essay
topic
payment
forced
Outcomes
———————————————
write
essay
value
Background Knowledge
#10 topic (T)
#10 T + payment
#10 T + 20% P
R
#10 T + forced (F)
#10 T + P + F
#10 T + 20% P + F
+̃
+̃
+̃
+̃
+̃
+̃
0
+̃
+̃a
0
+̃
+̃a
0
0
0
+̃
+̃
+̃
0
+̃
+̃
+̃
+̃
+̃
0
0
0
–̃̃
0
–̃
0
0
+
+
+
+
+
+
0
0
–
0
0
?
?
Induced Compliance (each condition only once)
#1 choice & low payment
#1 choice & high payment
#1 forced & low payment
#1 forced & high payment
+
+
+
+
+a
+
+a
+
Test
attitude toward essay
+
0
Note. Schematic representation of induced compliance experiment by Linder, Cooper & Jones (1967); Each
factor/outcome is represented by 5 nodes; Cell entries denote external activation with + and – reflecting a
randomly drawn Normal distributed pattern with M=+1/-1 & SD=.2 (identical across all trials; the
simulation was run for 5 such activation patterns and results were averaged); R=Randomized order;
#=number of trials.
a M=+.2 to reflect low payment; ~ Noise added randomly at each trial with Normal distribution of M=0 &
SD=.2
Social Cognition and Connectionism 74
Table 12
Fit and Robustness of the Simulations, including Alternative Encoding and Models
Distributed
Feedforward
Non-linear
Recurrent
.97 (.01)
.98 (.32b)
.98
.97
.99
.94
.97
.96
.94 (.40c)
.99 (.89c)
.90
—
.86
.78
—
.99
4
.97 (.28)
.91
.97
.96
5 persone
groupe
.99 (.01)
—
—
.99
.99 (.01)
—
—
.99
6e
—
.73
.53
7
.86 (.05)
.99 (.10d)
.99
.99
.99
8
.99 (.20)
.98
9
.94 (.30)
.94
.99
.82x
.97
.84x
10e
.99 (.02)
—
.95
.66x
Nr
1
2
3 contn
final
Original
Simulationa
Note. Cell entries are correlations between mean simulated values (averaged across randomizations) and
empirical data or theoretical predictions. For the distributed encoding, each concept was represented by 5
nodes and an activation pattern drawn from a Normal distribution with M = activation of the original
simulation & SD = .20 (5 such random pattern were run and averaged) and additional noise at each trial
drawn from a Normal distribution with M = 0 & SD = .20. For the Non-linear auto-associative model, the
parameters were: E = I = Decay = .15 and internal cycles = 9 (McClelland & Rumelhart, 1988). For all
alternative models, we searched for the best fitting learning parameter.
a Learning rate between parentheses; b-d The contextual node's learning rate was (b) 25%, (c) 33% or (d)
166% of this learning rate; e Distributed encoding; x Predicted pattern was not reproduced.
Social Cognition and Connectionism 75
Figure Captions
Figure 1. (A) Architecture of an auto-associative recurrent network, applied for (B) structural
relations and (C) causal relations.
Figure 2. Graphical illustration of the principles of (A) acquisition [with learning rate 0.20], (B)
competition and (C) diffusion.
O=outcome, C=consistent information, I= inconsistent
information, T=trait. Filled nodes are activated at a single trial, empty nodes are not
activated.
Full lines denote strong connection weights, broken lines denote moderate
weights while dotted lines denote weak weights.
Figure 3. Categorization and prototype abstraction: Network architecture with 4 feature nodes
connected to 2 category nodes (only the important connections are shown). Connection
weights are shown after the learning history listed in Table 2, where stronger weights are
depicted by solid lines and weaker weights by broken lines.
Figure 4. Categorization and prototype abstraction: Observed data from Gluck and Bower (1988)
and simulation results of categorization (top panel) and prototype abstraction (bottom panel;
learning rate = .01). Note that the simulation results from the top panel were regressed onto
the observed data with an intercept fixed at .50.
Figure 5. Impression formation: Observed data from Stewart (1965) and simulation results
(learning rate for person = .32, for context = .08).
Figure 6. Impression formation: Observed serial position curves from Dreben, Fiske & Hastie
(1979; left panels) and simulation (right panels) of attenuation of recency given continuous
responding (top; learning rate for person = .40, for context = .13) and primacy given final
responding (bottom; learning rate for person = .89, for context = .29).
Figure 7. Higher recall of inconsistent behavioral information after impression formation and
memory instructions: Observed data from Hamilton, Katz and Leirer (1980, exp. 3) and
simulation results (learning rate = .28).
Figure 8. Generalization: Simulation of exemplar and group stereotype assimilation (learning rate =
.01). The original external activation of the 5 nodes (reflecting micro-features) is given by
solid lines, while the reconstructed activation through internal input is given by broken
lines.
Figure 9. Assimilation and contrast effects after priming with a trait or person: Observed data from
Social Cognition and Connectionism 76
Stapel, Koomen and Van der Pligt (1997, exp. 3) and simulation results (learning rate =
.05).
Figure 10. Causal attribution and discounting: Observed data from Van Overwalle and Van Rooy
(2000) and simulation results (learning rate = .10).
Figure 11. Attitude formation: Prediction from the theory of planned behavior by Ajzen (1991) and
simulation results (learning rate = .20).
Figure 12. Dual processes of attitude formation: Observed data from Chaiken and Maheswaran
(1994; top panel) and simulation results (learning rate = .30; bottom panel).
Figure 13. Cognitive dissonance: Observed data from Linder, Cooper and Jones (1967) and
simulation results (learning rate = .02).
Figure 1
A. Recurrent Architecture
Output
Internal
input
External input
B. Structural Connections
Category
feature
exemplar
C. Causal Connections
Outcome
cause
attitude-object
Figure 2
A. Acquisition
feature A
1
0
0 1 2 3 4 5 6 7 8 9
Trials
B. Competition
A
B
F
F
D
F
F
F
F
I
D
C. Diffusion
B
BB
AT
1
F
F
F
F
F
F
I
I
I
Figure 3
Flemish
Walloon
(Rare)
(Common)
.21
.05
Dutch
lessrefined
sophisticated
French
Figure 4
Judged Probability of Flemish
Base-rate N eg lect
0.7
Gluck & Bower
Probabilistic
Simulation
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Dutch
lesssophisticated
refined
French
Feature
Simulated Prototypes
1.0
0.8
Flemish
Walloon
0.6
0.4
0.2
0.0
-0.2
-0.4
Dutch
lesssophisticated
Feature
refined
French
Figure 5
Impression Formation
8
High-Low
Low-High
Simulation
Mean Impression
7
6
5
4
3
2
1
2
3
4
5
Successive Adjectives
6
7
8
Figure 6
D ata
Simulation
Impact of position on
1st rating
2nd rating
3rd rating
4th rating
0.80
0.70
Weight
0.60
0.50
0.40
0.30
0.20
1
2
3
4
Serial Position -
1
2
3
4
Continuous Responding
Impact of position on
final rating
0.45
Weight
0.40
0.35
0.30
0.25
1
2
3
4
1
2
Serial PositionFinal
- Responding
3
4
Figure 7
R ecall of Inconsistent Information
0.8
Consistent
Inconsistent
Simulation
% Free Recall
0.7
0.6
0.5
0.4
0.3
Impression
Memory
Condition
) and Internal (
) Activation
Figure 8
Generalization of Persons and Groups
1.4
1.0
Group
0.6
0.2
-0.2
Person
Original (
-0.6
-1.0
1
2
3
4
Distributed Nodes of Unobserved Feature
5
Figure 9
Priming Trait versus Person Exemplars
Impression of Ambigious Person
5.0
Positive Prime
Negative Prime
Simulation
4.5
4.0
3.5
3.0
2.5
2.0
Trait
Person
Prime
Figure 10
D iscounting and Sample Size
80
75
Small Size
Large Size
Simulation
Causal Rating
70
65
60
55
50
45
40
Known
Novel
Cause
Figure 11
Theory of Reasoned Action
0.2
Prediction
Simulation
Attitude
0.1
0.0
-0.1
-0.2
-0.3
Car
Bicycle
Attitude-object
Bus
Figure 12
D ual Processes in Attitude Formation
3.0
Strong
Observed Data
2.5
2.0
1.5
Ambiguous
1.0
0.5
0.0
Credibility:
Weak
Low
High
Low Importance
Low
High Importance
Strong
50
Internal Activation
High
40
30
Ambiguous
20
10
Weak
0
Credibility:
Low
High
Low Importance
Low
High Importance
High
Figure 13
Induced Compliance
Attitude toward Essay Topic
4.0
Low Payment
High Payment
Simulation
3.5
3.0
2.5
2.0
Yes
No
Choice
Figure 14