Modeling Cognitive Dissonance as a Parallel Constraint Satisfaction Network With Learning
Mathematical Models, Neural Activations, and Affective Responses
Stephen J. Read and Brian M. Monroe
One of the most famous and productive theories in social psychology has been Festinger’s (1957) theory of cognitive dissonance. The term cognitive dissonance has entered everyday language and is often used to describe inconsistencies among cognitions. According to cognitive dissonance theory, when two cognitions are dissonant with each other, an individual is motivated to reduce that dissonance. Researchers have used dissonance theory to generate a number of counterintuitive findings, such as finding that people like groups better the more painful the initiation is to join that group (Gerard & Mathewson, 1966), and that people like a boring task better, if they agree to tell someone that the task is actually interesting in exchange for a small amount of money compared to receiving a large amount (Festinger & Carlsmith, 1959).
Several authors (e.g., Read & Miller, 1994; Read & Simon, 2012; Read, Vanman, & Miller, 1997; Shultz & Lepper, 1996, 1998; Simon & Holyoak, 2002; Spellman, Ullman, & Holyoak, 1993) have argued that cognitive dissonance (Festinger, 1957) and related consistency phenomena (Abelson et al., 1968) can be modeled as a parallel constraint-satisfaction process in a neural network, where the relevant cognitions are treated as nodes that have activation levels representing the initial strength of the corresponding cognitions, and the consistent and inconsistent relations between cognitions are treated as excitatory and inhibitory relationships among nodes, respectively. After the nodes in the network are initially activated, activation spreads, over time, across the excitatory and inhibitory links until the activation of the nodes in the network settles into a stable state that best represents the parallel satisfaction of the constraints imposed by the excitatory and inhibitory links and the level of activation of the nodes.
Read et al. (1997) noted that parallel constraint-satisfaction processes provide a computational implementation of the Gestalt processes that are the theoretical underpinnings of dissonance theory. These authors, and particularly Shultz and Lepper (1996, 1998), have shown that a number of different cognitive dissonance findings can be successfully modeled in a parallel constraint-satisfaction network that is implemented as a feedback or recurrently connected network. Many of Shultz and Lepper’s (1999) simulations were reviewed in their modeling chapter in the previous edition of this book.
Van Overwalle and Jordens (2002) noted that these approaches to modeling cognitive dissonance can only model immediate attitude and belief change; because they lack a learning mechanism, they are incapable of modeling long-term change that might result from resolving dissonance. Although learning could be implemented in these models, none did so.
Van Overwalle and Jordens’s (2002) proposed solution was a feedforward neural network, consisting of one input layer (representing the key features of the experimental situation), feeding into one output layer (representing a feeling and a behavior). Because it is a feedforward network, activation only flows from input to output. And because there are no connections among the nodes in the output layer, the feeling and the behavior have no influence on each other. Delta-rule (or error-correcting) learning is used to change the weights.
A learning mechanism is critical for any model of cognitive dissonance that wishes to address long-term attitude change resulting from dissonance reduction. However, a major shortcoming of a two-layer, feedforward network is that it cannot capture the short-term attitude change, represented by changes in activation, that is easily represented in a feedback model. In the current chapter, we present a connectionist model of cognitive dissonance that integrates the strengths of the two types of connectionist models that have been previously proposed and that avoids many of their weaknesses. It is a recurrent or feedback network with learning. Here we successfully model four classic cognitive dissonance experiments: Free Choice, Forbidden Toy, Forced Compliance, and Severity of Initiation.
ISSUES WITH FEEDFORWARD MODELS, SUCH AS THE ONE USED BY VAN OVERWALLE AND JORDENS (2002)
One major issue with a feedforward model is that the characterization of cognitive dissonance is a radical departure from the consensual understanding of cognitive dissonance processes. Rather than treating dissonance reduction processes as a Gestalt-like seeking for good form and coherence (the historical view) or as a constraint-satisfaction process (the modern rendition of Gestalt ideas of coherence; see Read & Simon, 2012; Read et al., 1997; Simon & Holyoak, 2002), van Overwalle and Jordens’s (2002) model treats dissonance reduction purely as an error correcting learning process in which the individual receives explicit feedback as to their behavior toward an object or situation and their evaluative response to that object or situation, and then uses error correcting learning to update its predictions. They model the dissonance process as follows: the network predicts its attitude (measured as the average of the activation of a behavior and a feeling node) and then the researcher tells the network what its actual behavior and feelings were; the network then adjusts its weights to reduce the discrepancy between the predicted attitude and the actual attitude provided by the researcher. The suggestion is that we think about dissonance reduction, not as driven by consistency maintenance processes, but instead as learning to correct errors about our predicted behaviors given explicit feedback.
In contrast, a recurrent model provides an account of consistency seeking or dissonance reduction in terms of parallel constraint-satisfaction processes that are a computational implementation of the Gestalt-like seeking for good form that underlies cognitive dissonance theory and its variants (Read & Simon, 2012; Read et al., 1997). These parallel constraint-satisfaction processes are an inherent part of a recurrent model.
In addition to presenting a different conceptualization of the dissonance process, another major issue is that van Overwalle and Jordens’s (2002) model does not actually infer a new attitude. Instead it is explicitly given the attitude by the researcher. Their network predicts its behavior and feeling and then receives explicit feedback about what its measured behavior and feeling should be. In the simulations we present, the recurrent network receives direct feedback about its behavior, but it does not receive any direct feedback about evaluation or attitude. It uses feedback about the behavior to determine the new activation of the evaluation node. This is consistent with most dissonance studies where participants are encouraged to perform behaviors that are inconsistent with their attitudes. The activation of the evaluation (the new attitude) in our model is a result of a process of parallel constraint-satisfaction, in which the network of nodes pass activation around, until the network settles into a new stable state.
For instance, our network for the Forbidden Toy paradigm can be seen in Figure 10.1. Here, two evaluation nodes (one positive, one negative) are in the path (mediate the relationship) between the attitude object (the toy) and the behavior nodes. These evaluation nodes represent the attitude. We provide feedback only about how the child behaved. In response, the behavior nodes first send activation back to the evaluation nodes, resulting in changes to the momentary attitude, as represented by changes in the activation of the evaluation nodes. The network then subsequently changes its long-term attitude by changing the relevant weights from the toy to the evaluation nodes. That is, we simply tell the network that it performed a behavior counter to what would be expected, given its previous experience. This results in a change in the patterns of activations, including activations of the evaluation nodes, which the learning process then transforms into long-term attitude change by appropriate weight changes. The model changes the evaluations of the object and the relevant weights so that the previously unexpected behavior becomes consistent with the network’s new evaluation and relations.
FIGURE 10.1. Network for Simulation 2: Freedman’s (1965) Forbidden Toy
Further, van Overwalle and Jordens’s (2002) attitude measurement confounds feelings and behavior. Attitude is measured as the average of the activation of the behavior and feeling nodes, so it is impossible to get a measure of behavior that is not confounded with the measure of attitude. By contrast, we rely on the idea that an attitude is the evaluation of the attitude object (Fazio, 2007) and that this evaluation influences behavior toward the attitude object. Thus, we have separate evaluation and behavior nodes and can measure attitude and behavior separately.
Finally, because their model lacks connections between feeling and behavior in the output layer, the behavior node and the feeling node cannot influence each other. In our network the evaluation and behavior nodes are bidirectionally connected, so that activation from the evaluation and the behavior nodes can reciprocally influence each other. An example of why this is important can be seen if we simulated someone’s attitude and behavior toward a social group using van Overwalle and Jordens’s model. Someone could have an extremely negative feeling about a particular social group, yet that feeling could not affect their behavior toward the group because there would be no pathway between the feeling and the behavior.
We present a model that is consistent with the theoretical underpinnings of Dissonance Theory and that we believe appropriately captures the psychological mechanisms involved in dissonance. It is a recurrent (feedback) network with Contrastive Hebbian Learning (CHL), which can capture learning in a multilayer network. Because it is recurrent, it can capture the constraint-satisfaction processes that are central to dissonance processes. Further, because it is recurrent, changes in the behavior nodes can be propagated back to the evaluation nodes and lead to momentary changes in their activation, representing momentary attitude change. Thus, if the network is given information only about its behavior, it uses that information to change its attitudes, without needing direct feedback about its new attitudes. And the network then uses this new pattern of activation of the attitude nodes to modify its weights to capture long-term attitude change represented by changes in the weights in a multilayer network.
In the remainder of the chapter, we first outline a recurrent neural network model with learning that integrates the strengths of previous neural network models of dissonance, while avoiding some of their limitations. Second, we show how this network can simulate four classic dissonance studies.
DESCRIPTION OF THE DISSONANCE MODELING PROCEDURE AND RESULTS
We first provide an overview of the network that we used in our simulations, then describe the general simulation procedure we used in all our simulations, and finally describe the results for the simulations of several classic cognitive dissonance phenomena.
The simulations were done using the constraint-satisfaction module (cs++) in the PDP++ neural network package (Dawson, O’Reilly, & McClelland, 2003).1 This is a freely available, powerful, well-known, and well-supported package. A major advantage of using a package such as emergent is that any simulations can easily be run and investigated by anyone with access to the software. In addition, emergent has been used to build a variety of large-scale cognitive systems, which means that cognitive dissonance ideas could be integrated into large scale cognitive models.
The model used a Contrastive Hebbian Learning (CHL) algorithm developed for the Boltzmann machine and then generalized by O’Reilly (1996). This algorithm compares the activation of the network in a plus phase (when both inputs and desired outputs are presented to the network) to its activation in a minus phase (when only the inputs are presented). CHL then adjusts weights to reduce the difference in activation between the two phases. What CHL is doing is that in the minus phase, the network uses its inputs to predict what its outputs should be whereas the plus phase represents what the inputs actually lead to. This is compatible with the typical dissonance study, where the minus phase can be thought of as what the individual expects to do and the plus phase can be thought of as what the individual actually did. CHL allows for learning in multilayer networks (e.g., networks with hidden units) and adjusts weights in a more biologically plausible way than does back propagation, which requires the use of a nonlocal error term and assumes that the error can be propagated back through multiple layers. CHL uses a local error term and thus does not require the propagation of an error signal. However, we are not making any strong claims about the superiority of CHL over back propagation. The central issue is being able to address learning in a multilayer network, where attitudes can intervene between attitude objects and behavior.
We used the default sigmoidal (S-shaped) activation function for the units in the networks, with activations limited to the range −1 to +1. Bias weights were set at zero. During the predissonance learning phase of the simulations, the learning rate was set to 0.1. Because in some of the experiments, more complicated conjunctive relationships were required to be learned, learning in these simulations proceeded for more epochs (that is, more passes through the learning instances). The number of learning epochs ranged from 20 for Simulation 1: Free Choice to 100 epochs for Simulations 2—4. The specific details of the learning procedure in the predissonance phase are not critical to the model, as this learning was done simply to ensure that each network represented an appropriate pattern of cognitions.
General Overview of the Simulation Procedure
The general logic of the simulations is as follows. First, the network is exposed to a set of learning examples so that it develops a set of weights that correspond to the initial pattern of cognitions in the relevant experiment. For example, in the Forbidden Toy simulation (Freedman, 1965; see Figure 10.1), the model learns such things as that the toy is attractive (that is, the presence of the toy leads to positive activation of a positive evaluation node and negative activation of a negative evaluation node). And the positive evaluation node is positively associated with the “play with the toy” node. Thus, when the toy is presented to the network, it should highly activate the positive evaluation node, which should then highly activate the “play with the toy” node.
In the current simulations, we are using the training simply to make sure that the network represents the appropriate cognitions. Thus, when there are different patterns of relevant cognitions in different conditions, we will capture that by giving the networks different learning histories. In other work it would be interesting to use learning to investigate the impact of different evaluations and beliefs, as a result of different experiences, on dissonance processes.
Once the networks have learned the relevant pattern of cognitions, we expose the network to a set of cues that correspond to a condition in a dissonance experiment. Essentially, we first expose the network to the relevant cues and let it generate a prediction for how it will behave. We then give the network feedback as to how it did behave in the environment. For example, consider the different conditions in the Linder, Cooper, and Jones (1967) Forced Compliance study, where participants are asked to write a counter-attitudinal essay under either “Free Choice” or coercion, with high or low payment (see Figure 10.2 for the network). In the coercion condition, with high payment, features for the attitude object, the coercion, and the high payment would be strongly activated, and we would then let the network generate a prediction. In this case, the node for writing a counterattitudinal essay would be highly activated. The network would then be given feedback indicating that the network did write a counterattitudinal essay. Because there is no discrepancy between the predicted behavior and the actual behavior, and because there would be very little change in the activations of the other nodes in the network, particularly the evaluation nodes, there will be little weight change and therefore little attitude change.
FIGURE 10.2. Network for Simulation 3: Linder, Cooper, and Jones’s (1967) Forced Compliance
However, the behavior of the network in the dissonance condition: the Free Choice, low-payment condition would be different. Here, the nodes corresponding to the attitude object, the Free Choice, and low payment would be activated and the network would generate a prediction, which is that the individual would not write the counterattitudinal essay. However, the teaching feedback is that the individual did write the counterattitudinal essay. Thus, the predicted behavior and the actual behavior are dissonant. This is represented in the network by positive activation of both nodes, which have an inhibitory or negative link between them. Thus, they seek to inhibit each other, with the “writing the counterattitudinal essay” node, because it receives external input, successfully inhibiting the original prediction. Activation is then propagated from the behavior nodes back to the evaluation nodes, resulting in changes of the activation of the evaluation nodes to become more consistent with the activation of the behavior nodes. This change in activation of the evaluation nodes represents short-term attitude change.
The weights in the network are then adjusted, capturing long-term attitude change, so that in the future, the predicted activation of the behavior nodes is closer to the teaching activation and the predicted activation of the evaluation nodes is closer to their actual activation. We then test for the new attitude by activating the attitude object and seeing how strongly activated are the positive and negative evaluation nodes.
Simulation 1: Shultz, Léveillé, and Lepper (1999) Free Choice
This experiment was a refinement of the classic study by Brehm (1956), where participants were asked to rate a set of appliances, then to choose one they wanted from among those items (they were then allowed to keep the item), and finally asked to rate the objects again. The manipulation was whether the choice was easy to make (if the difference in the attractiveness of the items was large, the choice of which item they wanted would be clear) or if the choice was a difficult one (where the attractiveness of the items was very similar).
In the Shultz, Léveillé, and Lepper (1999) version, there were three conditions, as opposed to the two in Brehm’s (1956) experiment. The “easy” condition from Brehm’s study was preserved, but instead of one difficult condition there were two. In Brehm, the difficult condition had two highly rated items. This condition was translated into the difficult-high condition. The new condition was one in which the items were rated similarly but were both poorly rated (difficult-low). The rationale for adding this condition was that Shultz and Lepper’s (1996) consonance model predicted that dissonance would manifest itself differently depending upon whether the difficult choice was between two highly rated objects or between two poorly rated choices. In Brehm’s difficult condition (difficult-high rating), the nonchosen item was derogated to preserve consonance of thought, while Shultz and Lepper’s (1996) model predicted that if the items were both rated poorly to begin with (difficult-low), derogation of the nonchosen item would not be a feasible option, so the chosen item would be bolstered instead.
So, Shultz, Léveillé, and Lepper’s (1999) experiment was set up as follows: In the easy condition, one item was a highly rated poster and the other was a poorly rated poster. In the difficult-low condition, both posters were rated poorly, and in the difficult-high condition, both posters were rated highly. Shultz and Lepper’s (1996) consonance model predicted that in the difficult-high condition, the nonchosen item would be derogated more severely than in the easy condition. In the difficult-low condition, the prediction was that the chosen item would be bolstered more than in the easy condition, because the rejected item was already poorly rated and so the only reasonable way to rearrange cognitions would be to raise the evaluation of the chosen item. These predictions were borne out in the experiment (see Figure 10.3a).
FIGURE 10.3. Results for (a) Shultz, Léveillé, and Lepper’s (1999) Free Choice; and (b) Simulation 1
Structure of the Network
To model this simulation, we used a setup whose basic features are used in all subsequent models in this report. The model had three layers (see Figure 10.4). The first of these layers corresponded to the object(s) in question and served as the primary input layer. The second layer incorporated the evaluation of the object (positive or negative, one node for each). The third layer corresponded to the relevant behaviors in the experiment. The connections between the input layer and the evaluation were feedforward only. This makes sense because the objects/issues under evaluation in dissonance experiments depend purely on external activation, that is, the evaluation comes to mind as a result of activation coming from the object, rather than the object being present as a result of activation from the evaluation. The connections between the evaluation and the behavior were recurrent (bidirectional). Here it is important that the evaluation be allowed to influence the behavior, and vice versa, so activation is allowed to flow in both directions. Thus, in our model information about the behavior and the underlying evaluation can mutually influence each other. This contrasts with van Overwalle and Jordens’s (2002) model in which there is no way for information about the behavior to affect the affective response (or vice versa). This limits the kinds of processing it can do.
FIGURE 10.4. Network for Simulation 1: Shultz, Léveillé, and Lepper’s (1999) Free Choice
The evaluation units were connected to each other with a permanent negative weight (−0.5), and the behavior units were also connected to each other with permanent negative weights (−1), because for all practical purposes these units are mutually exclusive. The behavior nodes had such strong negative weights in order to model the idea that the behaviors were completely mutually exclusive (i.e., you either choose the poster or you don’t). The evaluation nodes had weaker negative weights to allow for the possibility of ambivalence. That is, although these negative weights tend to make the units mutually exclusive, both evaluation nodes can be positively activated at the same time, if they each receive strong enough activation.
In the specific model for this simulation, there were two units in the object layer: one for each of the two posters between which the participant was choosing. The evaluation layer had one unit each for positive and negative evaluations, and the behavior layer had two units, one for each of the behaviors: choose and reject. See Figure 10.4 for a depiction of this model.
Learning Initial Cognitions
To set up the cognitions that were relevant in this simulation, we put the model through a learning process (see Appendix 10.1 at the end of this chapter for details). The associations we had the model learn were the following: (a) a positive evaluation is positively associated with choosing an item; and (b) a negative evaluation is positively associated with rejecting an item. In the difficult-high condition both posters had high evaluations (high activation on the positive evaluation node); in the difficult-low condition, both posters had low evaluations (low activation on the positive evaluation node); and in the easy condition, one poster had a high evaluation and the other poster a low evaluation. Across all conditions, the more attractive item was ensured to be at least very slightly more positively evaluated (activated) than the less attractive item. After initial learning, the model was tested to ensure that these associations were learned correctly.
To insure the generalizability of our results, in this and all subsequent simulations we generated 20 different “individuals.” We did this by generating 20 different networks with different random initial weights and then exposed each network to the learning environment. Each of the 20 individuals or networks was then put through the dissonance phase and attitude change procedure. Thus, the results presented for this and all the subsequent simulations are the average across 20 runs with different random initializations.
Dissonance Phase and Attitude Change
For the dissonance phase of the simulation, the three conditions were implemented identically. In this simulation, the key differences were in terms of the different evaluations for the different items, which were implemented by the different learning histories, resulting in different networks for the different conditions. That is, in the difficult-high condition, the corresponding network had learned high evaluations for both posters; in the difficult-low condition, the corresponding network had learned low evaluations for both posters; and in the easy condition, the corresponding network had learned a high evaluation for one poster and a low evaluation for the other poster.
The following procedure was implemented in each of the three conditions. First, the more attractive item was paired with the behavior of being chosen (i.e., the corresponding nodes were activated), the network was allowed to settle, and the weights were updated. Second, the less attractive item was paired with the behavior of being rejected, the network was allowed to settle, and the weights were updated. This was done once for each pairing. In all the simulations reported in this paper, a learning rate of 0.5 was used in the dissonance conditions.
In the difficult-high condition, the slightly less attractive alternative was still highly evaluated, which would lead the network to predict that it would be chosen. However, the feedback was that it was not chosen, which creates dissonance. Thus, to resolve this dissonance, we expected that the attractiveness of the nonchosen alternative would decrease.
In the easy condition, because the network had learned a high evaluation for the chosen poster and a low evaluation for the nonchosen poster, there shouldn’t be an expectation that the low evaluation poster would be chosen. Thus, there would be no dissonance and we would expect minimal weight change.
Finally, in the difficult-low condition, choosing the slightly more attractive, unattractive alternative was dissonant with the fact that it was unattractive and with the network’s prediction that it would not be chosen. To resolve that dissonance, the weights should change so that the chosen alternative would be more attractive. However, for the nonchosen alternative, because it was unattractive, the network would also predict that it would not be chosen. Thus, there would be no dissonance here and we would predict little weight change for this choice.
After the dissonance process, we assessed the “new” evaluations of the two items by separately activating the respective object units and reading off the activation of the positive evaluation unit (for these simulations, the negative evaluation unit had virtually the exact opposite activation, so only the positive unit activation is reported.) Further, we reasoned that when the person was asked for their evaluation only and did not actually perform a behavior, it did not make sense to allow activation to flow through the behavior units, so those units were deactivated during this assessment by lesioning this portion of the network (that is, setting the weights from the behavior nodes to 0). Results are presented in Figure 10.3b and are the average across the 20 runs. In the difficult-high condition, the rejected item was derogated, while the evaluation of the chosen item increased only very slightly. In the easy choice condition, the chosen item’s evaluation increased very slightly, and the rejected item’s evaluation decreased very slightly. In the difficult-low condition, the rejected item’s evaluation decreased very slightly, yet the chosen item’s evaluation increased much more significantly.
Our simulation is more successful in capturing the results from Shultz et al.’s (1999) experiment than is van Overwalle and Jordens’s (2002) model. Shultz et al. found greater spreading apart of alternatives in both difficult choice conditions (high- and low-attractiveness of the posters), compared to the easy choice condition, whereas van Overwalle and Jordens’s simulation only shows spreading apart of the highly attractive, but not the low-attractiveness alternatives. In contrast, our model demonstrates spreading of alternatives in both high- and low-attractiveness, difficult-choice conditions.
Simulation 2: Freedman (1965) Forbidden Toy
In this paradigm, a young child is brought into the experimental room and shown an attractive toy: an interesting robot. The experimenter than administers either a mild or a severe threat to the child to not play with the robot. The experimenter then either leaves the room or stays in the room. So, the child is in one of four conditions: (a) mild threat, no surveillance; (b) mild threat, surveillance; (c) severe threat, no surveillance; or (d) severe threat, surveillance. The child is then observed through a one-way mirror. In this initial observation, none of the children play with the robot.
Forty days later the child is brought back to the experiment room and again left alone in the room with the toy. The experimenters observed the child through a one-way mirror and recorded whether the child played with the toy. The researchers predicted that all of the children should play with the toy except for those who were initially in the mild threat, no surveillance condition. The rationale is that, in all the other conditions, the child believes that they did not play with the toy because of the threat of punishment (or because of the combination of mild threat and surveillance). Thus, they didn’t need to rationalize why they didn’t play with the attractive toy. There shouldn’t have been any change in the child’s liking for the toy, and when they are later given an opportunity to play with the toy without any possibility of punishment, they will happily play with the toy. In contrast, the argument is that for the children in the mild threat, no surveillance condition, their failure to play with the robot earlier was perceived as inconsistent with the fact that they had only received a mild threat. To justify this inconsistency, they would decrease their liking for the toy. As predicted, the only children who do not play with the toy were those who were given the mild threat and thought they were not watched.
Structure of the Network
The network for this simulation is presented in Figure 10.1. As with Simulation 1, this model has a layer for the object in question (the toy), a second layer for the evaluations of that object, and a third layer for the behaviors relevant to the experiment (playing with the toy, not playing with the toy). However, this experiment was more complicated in that there were additional inputs that were relevant. Specifically, there was an additional input layer that contained additional important features of the situation: one node for the presence of the mild threat given to the child, one for severe threat, and one for the presence of surveillance during the critical phase of the experiment. Finally, in addition to these inputs, a hidden layer with eight nodes was implemented, in order to capture the conjunctive relationship between the threat and surveillance. The connections from the input layers were feedforward and all other connections were recurrent. Again, there were permanent negative connections between the positive and negative evaluations (−0.5), and between the two mutually exclusive behaviors (−1).
Learning Initial Cognitions
The initial learning history for this model was as follows (see Appendix 10.1 for details), across all four conditions: (a) the toy was associated with a positive evaluation, (b) the positive evaluation was associated with playing, and (c) the negative evaluation was associated with not playing. Additionally, the following conjunctive relationships were learned: (a) when the toy was present under surveillance only (and no threat), the toy was still played with; (b) when the toy was present with a mild threat, the toy was still played with; (c) when the toy was present with a severe threat, the toy was not played with; and (d) when the toy was present with surveillance and either a mild or severe threat, the toy was not played with. Testing confirmed these associations were learned properly. As noted above, this and all simulations in this paper are based on averages across 20 randomly initialized networks.
A key aspect of training the network in this way is that when the threat was mild, with no surveillance, the network predicted the child would play with the toy. This behavior is central to the dissonance effect in this network. According to the classic interpretation, the child, under mild threat and no surveillance, was expected to play with the toy, yet did not. Thus, there was dissonance between what the child expected to do and what s/he actually did. It was argued that the child resolved this dissonance by deciding that the toy was less attractive. However, with all other combinations of inputs (i.e., mild threat, surveillance; severe threat, surveillance; or severe threat, no surveillance) the network predicted that the child would not play with the toy.
Dissonance Phase and Attitude Change
The dissonance phase of the simulation presented one event to the network that varied across conditions. In the mild threat, no surveillance condition, the toy and the mild threat were activated and the initial prediction of the network was that the child would play with the toy. However, the feedback was that when the toy and mild threat were present, the output was not playing with the toy (represented by a positive activation on the don’t play unit, and negative activation on the play unit). Thus, there was a discrepancy between what was predicted and what actually happened. The behavior nodes sent activation back to the evaluation nodes, the network then settled and the weights were changed to try to better fit this new pattern of activations.
However, in the other three conditions, the predictions of the network and the actual behavior were consistent. In all three conditions, the network predicted that it would not play with the toy, and the feedback was that it did not play with the toy. In the severe threat, no surveillance condition, the toy and the severe threat node were activated, along with the don’t play node. In the mild threat, surveillance condition, the toy, mild threat, and surveillance were activated, along with the don’t play output. In the severe threat, surveillance condition, the toy, severe threat, and surveillance were activated along with the don’t play node. Because there was no inconsistency or dissonance in these three conditions, we did not expect to see any changes in weights.
Subsequent assessment of the attitude, after the dissonance phase, was done by activating the toy unit by itself, and reading out the activation on the play unit. (The hidden layer was deactivated by lesioning, since the purpose of this layer was to represent conjunctions of the other inputs, which were not present in this phase.) In this simulation, in contrast to the others, we focused on the activation of the behavior units, rather than the evaluation units. We did this because the primary dependent variable in the original study was the child’s behavior, whereas in the other studies simulated here, the dependent variable was the participants’ attitudes.
Results are presented in Figure 10.5a (original study results) and Figure 10.5b (simulation results). Consistent with previous simulations and the experiment itself, only in the mild threat, no surveillance condition did the activation on the play unit decrease after the dissonance phase. This makes sense as only in that condition would there be a discrepancy between how the child was expected to behave and how they actually behaved. To ensure that our results weren’t biased by looking at the activations on the behavior units in this simulation, we also examined the results by deactivating the behavior units and looking at activation of the evaluation units. The pattern of results was virtually the same.
FIGURE 10.5. Results for (a) Freedman’s (1965) Forbidden Toy; and (b) Simulation 2
Simulation 3: Linder et al. (1967): Forced Compliance
This experiment was designed to conceptually extend an experiment by Festinger and Carlsmith (1959) where participants were paid to tell someone that a boring task was in fact interesting, and participants who were paid less actually changed their attitude about the task more in a positive direction. The Linder et al. (1967) experiment examined whether the existence of dissonance in studies such as Festinger and Carlsmith (1959) depended on whether participants perceived that their behavior was freely chosen. Thus, they introduced an additional factor besides how much the participant was paid: whether they had a perceived choice in performing the counterattitudinal act. Students were asked, either under conditions of choice or Forced Compliance, to write an essay supporting a policy that they were not personally in favor of. The low payment was $0.50 and the high payment was $2.50.
Structure of the Network
The critical features for simulating this experiment were the attitude object (e.g., legalization of marijuana, abortion), their evaluation of it (either for or against the policy; promarijuana vs. antimarijuana), the behaviors (either writing a counterattitudinal essay or not writing it), and the situational features of coercion and small and large payments (see Figure 10.2). The form of this model is almost the same as the model in Simulation 2 (the Forbidden Toy experiment), except for changes in the definitions of the input and behavior units. In this model though, we have added a mechanism to account for the mood effects as a result of the payment and coercion that are apparent in the experimental results. In short, the input units representing payment lead directly to a positive mood, which then has an effect on the evaluation. This relationship was structured with fixed weights. The weight between the high payment and positive mood was set at +1, the weight between the low payment and positive mood was set at +0.2, and the weight between the coercion and positive mood was set at +0.7. It is reasoned that some of the negative mood resulting from writing the counterattitudinal essay was relieved by knowing it was done under coercion, thus leading to a more positive mood in these conditions. Weights between positive mood and a favorable evaluation and between positive mood and unfavorable evaluation were set at +0.5 and −0.5, respectively.
The rationale for introducing the mood nodes and their connections arises from the need to capture reward and punishment effects in this and the subsequent simulation. In this study, in the Forced Compliance conditions (the nondissonance conditions), participants in the high-payment condition changed their attitude more than did participants in the low-payment condition. And in the next simulation on Severity of Initiation, participants in the no-initiation condition were more negative toward the discussion group when they received higher levels of shock. One plausible way to capture this is to assume that people who received a higher payment are in a better mood, whereas people who receive a higher level of shock are in a worse mood. There is considerable evidence that such moods can strongly affect people’s evaluations and judgments (Isen, 2000; Schwarz & Clore, 2007).
Learning Initial Cognitions
Here the prior learning history was as follows (see Appendix 10.1 for details): the attitude object was associated with an evaluation against the policy; an evaluation against the policy was associated with not writing the essay supporting the policy; and an evaluation for the policy was associated with writing the essay. In addition, the conjunctive relations were learned: having the attitude, with low payment, was associated with not writing the essay; having the attitude, with high payment, was associated with writing the essay; having the attitude, under coercion, was associated with writing the essay; and having the attitude, under coercion and with high or low payment, was associated with writing the essay. Testing confirmed that all these associations were learned correctly.
Dissonance Phase and Attitude Change
In the dissonance phase of the simulations, events varied across the four conditions. In the Free Choice, low-payment condition, the attitude and low payment were presented as inputs and the network predicted that not writing the essay would be the output. However, the actual output behavior was writing the essay, leading to dissonance between the predicted behavior and the actual behavior, which should lead to attitude change. In the other three conditions, the behavior predicted by the network and the actual behavior were consistent with one another, which should not lead to attitude change. In the Free Choice, high-payment condition, the attitude and high payment were presented as inputs with writing the essay as the output (both predicted and actual). In the Forced Compliance, low-payment condition, the attitude object, coercion, and low payment were presented as input with writing the essay as output (both predicted and actual). Finally, in the Forced Compliance, high-payment condition, the attitude object, coercion, and high payment were presented as inputs with writing the essay as output (both predicted and actual).
The attitude postmeasure was assessed by activating the attitude object unit and the appropriate additional inputs (i.e., type of payment and coercion) for that condition. (The behavior units were deactivated through lesioning, because only the evaluation is being asked for, so it is inappropriate for the behavior activation to influence the evaluation.) This still allowed the payment and coercion to influence the attitude through their effects on mood.
Results are presented in Figure 10.6a (original study results) and Figure 10.6b (simulation results). The results were consistent with both previous simulations and the experimental data. In the Free Choice condition, the attitude increased the most in the low-payment condition, but in the Forced Compliance condition, the attitude increased the most in the high-payment condition. In the Forced Compliance condition, higher payment was more associated with better mood and a more favorable evaluation of the proposal. However, attitudes were still against the proposal in all conditions, which matches the experimental data.
FIGURE 10.6. Results for (a) Linder, Cooper, and Jones’s (1967) Forced Compliance; and (b) Simulation 3
Simulation 4: Gerard and Mathewson (1966) Severity of Initiation
This experiment was an elaboration of the Aronson and Mills (1959) study that showed that the more severe an initiation to join a group, the more the group was liked. Women who were asked to read a more embarrassing passage aloud (explicit sexual language), as their initiation process, before joining a discussion group, rated the group more positively than women who only had to read a mildly sexually embarrassing passage aloud. In the Gerard and Mathewson (1966) version, to ensure that the effect of the initiation in the original study was due to its aversiveness, and not to sexual arousal, the experimenters used electric shock to operationalize Severity of Initiation. Further, to isolate the effects of the severity of the initiation itself from multiple alternatives, ranging from arousal to extraneous expectation effects, a factor was added to separate the initiation from the act of joining the group. In the noninitiation condition, a shock was administered ostensibly as part of an unrelated experiment. Thus, Gerard and Mathewson’s experiment had four conditions: mild shock, no initiation; severe shock, no initiation; mild shock, initiation; and severe shock, initiation. Also, after the shock, all participants listened to a boring discussion by that group. Evaluations of the group were then assessed afterward.
Structure of the Network
The critical features for modeling this experiment were the group, the evaluation of the group, the mild shock and severe shock as inputs, and joining the group as a behavior (see Figure 10.7). A hidden layer was still used, even though it was not necessary to learn any specific conjunctive relations. A positive evaluation would predict joining the group, and shock would predict not joining the group, regardless of the presence of the other feature. This network also contained the path for mood (due to the shock) to affect the evaluation. High shock had a fixed +1 weight to negative mood, and low shock had a fixed +0.1 weight to negative mood. Weights between mood and evaluation were the same as in the Forced Compliance simulation.
FIGURE 10.7. Network for Simulation 4: Gerard and Mathewson’s (1966) Severity of Initiation
Learning Initial Cognitions
The prior learning history was (see Appendix 10.1 for details): the group was mildly associated with the behavior of not joining (because the group discussion turned out to be boring); a positive evaluation of the group was associated with joining and a negative evaluation was associated with not joining; the group combined with low shock was associated with mildly not joining; and the group combined with high shock was associated with strongly not joining. Testing confirmed that all associations were learned properly.
Dissonance Phase and Attitude Change
In the dissonance phase of the experiment, the series of events diverged. In the initiation conditions from the actual experiment performed by Gerard and Mathewson (1966), there was a dissonance-producing event, and in the noninitiation conditions there was not—in the noninitiation conditions, the shock and the exposure to the group discussion were presented as separate events in a sequence that had no meaningful connection. Thus, in our simulations, there was a difference in the events presented to the network: in the mild shock, initiation condition, the group and mild shock were presented as inputs, with joining the group as both the predicted output and the actual output. In the severe shock, initiation condition, the group and severe shock were presented as inputs, with not joining the group as the predicted output and joining the group as the actual output. No analogous dissonance-producing events were used for the noninitiation conditions. The attitude change measure was assessed by activating the group and appropriate level of shock as inputs (behavioral units were deactivated through lesioning again, as only evaluation was requested).
Results are presented in Figure 10.8a (original study results) and Figure 10.8b (simulation results). They confirm both the previous simulation results as well as the experimental results. In the initiation condition, severe shock participants liked the group much more than the mild shock participants, while in the noninitiation condition, severe shock participants liked the group slightly less than mild shock participants.
FIGURE 10.8. Results for (a) Gerard and Mathewson’s (1966) Severity of Initiation; and (b) Simulation 4
It is worth noting that in order to capture the results in the noninitiation condition, van Overwalle and Jordens had to use an event in the dissonance phase where the person joined the group, even though in the actual experiment participants in these conditions did not join the group. In contrast, we were able to capture the effects in the noninitiation conditions without having to include any event in which participants in the noninitiation condition joined the group.
This recurrent neural network model, with CHL, successfully modeled the long-term attitude change that results from reduction of cognitive dissonance, with long-term change represented by changes in the weights among cognitions and evaluations. Further, our multilayer recurrent network used parallel constraint-satisfaction processes to model short-term attitude change, represented by changes in activation.
In all of our simulations, the network first generated an expectation about its behavior (e.g., play with the toy) and then was given feedback about how it actually behaved, which in the dissonant conditions was inconsistent with the expectation generated (e.g., didn’t play with the toy). As our network settles, the activations of the other nodes, especially the evaluation nodes, change to become more consistent with each other. For example, because of the recurrent connections, as the activation of the “didn’t play with toy” node increases, the activation of the negative evaluation node will tend to increase, and when activation of the “play with toy” node decreases, the activation of the positive evaluation node will also tend to decrease. Then once the network settles the weights will change so that activation of the input nodes is more likely to lead to the same pattern of activation in the future.
In addition to capturing the pattern of results in the studies we simulated, our simulation is more successful than van Overwalle and Jordens’ model in capturing the original results from Shultz et al.’s (1999) extension of the choice among alternatives’ paradigm (Simulation 1). Van Overwalle and Jordens’s (2002) simulation only exhibits spreading apart of choices in the difficult, highly attractive condition, but not in the low-attractiveness condition. In contrast, our simulation captures spreading of alternatives in both high- and low-attractiveness, difficult choice conditions. Also, for the Severity of Initiation simulation (Simulation 4) we were able to successfully simulate the results in the noninitiation conditions without having to introduce the behavior of joining the group, whereas van Overwalle and Jordens had to introduce this behavior, despite the fact that participants in the noninitiation condition did not think that the shock had anything to do with being a part of the boring group.
Theoretical Advantages of a Multilayer, Recurrent Model
A multilayer, recurrent network, with its bidirectional connectivity and constraint-satisfaction processing is much more powerful both in terms of its processing mechanisms and its representational ability than is a two-layer feedforward network and, as a result, it better represents many aspects of human cognitive functioning (e.g., O’Reilly & Munakata, 2000). More specifically, recurrent networks have a number of advantages over two-layer feedforward networks as a model of cognitive dissonance and dissonance reduction processes. Recurrent models can handle a broader range of kinds of inconsistency, they can handle reciprocal influences among different concepts, and they do a better job of representing the difference between stored and constructed attitudes. Further, as we discuss in more detail shortly, the parallel constraint-satisfaction processes that are part of recurrent networks capture the Gestalt-based, inconsistency reducing mechanisms that may underlie cognitive dissonance theory.
Can Handle Wider Array of Types of Inconsistency
First, a model that is multilayered and bidirectionally connected can deal with inconsistency between different kinds of concepts (e.g., attitude vs. belief, affect vs. behavior, belief vs. belief), as well as the inconsistency between an expected and an actual behavior or feeling. In contrast, a feedforward model can only deal with the discrepancy or inconsistency between what is expected and what actually happened for the same concept (e.g., expected vs. predicted behavior, expected vs. predicted positive affect). Thus, a recurrent model can deal with a much wider range of types of inconsistency or dissonance. For example, our model simulates the dissonance between two different concepts: an actual behavior and an attitude. This focus is consistent with many different kinds of dissonance studies, where the main theoretical explanation is in terms of inconsistency between an attitude and an actual behavior.
Can Handle a Wider Array of Influences Among Concepts
Because nodes in a recurrent model are bidirectionally connected, the model can deal with reciprocal influences between concepts. A feedforward network cannot. In a feedforward network, such as the one van Overwalle and Jordens (2002) used, there are only unidirectional links from nodes in the input layer to nodes in the output layer. And there are no links among the nodes within a layer. Thus, in van Overwalle and Jordens’s model, because of the absence of links among concepts in the output layer, activation of the behavior node cannot influence the activation of the affect node (or vice versa). Such a claim contradicts considerable data.
Captures the Underlying Theoretical Mechanism Postulated by Cognitive Dissonance Theory
A major characteristic of a recurrent network, because of its bidirectional connectivity, is constraint-satisfaction processes, which implement classic Gestalt-like consistency mechanisms. The original theoretical underpinnings of cognitive dissonance theory owe much to fundamental Gestalt ideas of good form and coherence seeking (see Read & Simon, 2012; Read et al., 1997, for detailed discussions). The central idea is that people have a need for consistency among their cognitions (good form), and that when their cognitions are inconsistent, there is “tension” in the system that acts to move the set of cognitions in a more consistent or coherent direction. As Read et al. (1997) argued (also see Read & Simon, 2012), parallel constraint-satisfaction mechanisms in a recurrent neural network are one way in which such consistency or coherence seeking can be computationally implemented (either in a computer or in a brain). Then once the network has reached a coherent or consistent state, the links among cognitions can be modified to capture these new relationships among the cognitions and beliefs in the network. Thus, the current model is consistent with the fundamental theoretical assumptions of Cognitive Dissonance Theory, and it provides an account of the mechanisms by which the relevant beliefs change.
In contrast, van Overwalle and Jordens (2002) eschew the standard theoretical underpinnings of cognitive dissonance theory. They do not provide an account of how beliefs might change as the result of a seeking after consistency. Rather, they propose that if relevant beliefs change as a result of direct feedback, that the new attitudes and beliefs are recorded by associative learning processes.
Effects of Mood on Cognitive Dissonance
Our current simulations capture some aspects of mood effects. In Simulation 3 of Linder et al.’s (1967) study on Forced Compliance and writing a counterattitudinal essay, we capture an apparent reward effect of a high payment. Specifically, in the high-coercion condition, participants showed more positive attitudes when they received higher payment for writing the counterattitudinal essay. This occurs because of the mood or affect nodes, which are influenced by reward and punishment, that then feed into the attitude or evaluation nodes. Research demonstrates that people’s moods can influence their evaluations of objects (Isen, 2000; Schwarz & Clore, 2007). This network configuration allows us to capture this effect.
And in Simulation 4, Gerard and Mathewson’s (1966) Severity of Initiation study, we captured a punishment effect. Participants who were told that receiving the shock was simply part of the experimental procedure showed greater decreases in liking for the group when they received higher levels of shock. Again, we modeled this as a mood effect.
Is This Just Self-Perception Theory?
Given that this model relies on feedback about the actual behavior of an individual, one might ask whether this isn’t just self-perception theory (Bem, 1972)? It is true that, just like self-perception theory, our model relies on feedback about its behavior to infer its attitudes? But the same is true of cognitive dissonance theory. And our theoretical mechanisms are quite different from self-perception theory. In our model, the central theoretical mechanisms are parallel constraint-satisfaction processes, which we have argued (e.g., Read et al., 1997) can be viewed as a computational implementation of the Gestalt consistency-seeking processes that were the foundation of cognitive dissonance theory. A constraint-satisfaction process changes the pattern of activation of the nodes in the network, specifically the evaluation nodes, to minimize the inconsistency between the predicted and the actual behavior of the individual. The activation of the mediating evaluation nodes is a result of feedforward activation from the attitude object and feedback activation from the behavior nodes. Over time, the activations of the network settle into a stable state. If this new state differs from the predicted state of the network, learning mechanisms (CHL) act to modify the weights so that in the future the actual state of the network is closer to the predicted state. In contrast, in a self-perception network, there would be no notion of inconsistency and the network would not be trying to minimize inconsistency. Thus, inconsistency and a seeking for coherence would not be driving weight change.
Researchers seeking to computationally model cognitive dissonance processes in a connectionist framework have proposed two different types of model: A feedforward model, such as van Overwalle and Jordens (2002), and a recurrent feedback model, such as the one proposed by Shultz and Lepper (1996, 1998) and Read and Miller (1994). Each of these types of models have their limitations. The models proposed by Shultz and Lepper, and by Read and Miller, lacked learning. Although van Overwalle and Jordens’s model captured learning, it did not infer its new attitude, but rather it was explicitly instructed. Perhaps more importantly, it could not capture the constraint-satisfaction or Gestalt-like nature of dissonance and dissonance reduction, and thus was unable to represent the process of short-term attitude change, represented by change in activation. The current model integrates the strength of the two approaches, while avoiding some of their important weaknesses. Because it is a multilayer, recurrent, feedback model it can capture the Gestalt-like nature of dissonance and short-term attitude change. And because the model learns, it can also capture long-term attitude change that results from dissonance processes.
TRAINING EVENTS FOR EACH SIMULATION
SIMULATION 1: FREE CHOICE
Difficult-High Condition Training (20 Epochs, Learning Rate 0.1):
1. attractive poster +1, positive evaluation +.3, negative evaluation —.3
2. less attractive poster +1, positive evaluation +.25, negative evaluation —.25
3. positive evaluation +1, negative evaluation —1, choose +1, don’t choose —1
Easy Condition Training (20 Epochs, Learning Rate 0.1):
1. attractive poster +1, positive evaluation +.3, negative evaluation —.3
2. less attractive poster +1, positive evaluation —.3, negative evaluation +.3
3. positive evaluation +1, negative evaluation —1, choose +1, don’t choose —1
Difficult-Low Condition Training (20 Epochs, Learning Rate 0.1):
1. attractive poster +1, positive evaluation —.25, negative evaluation +.25
2. less attractive poster +1, positive evaluation —.3, negative evaluation +.3
3. positive evaluation +1, negative evaluation —1, choose +1, don’t choose —1
Dissonance Phase (20 Epochs, Learning Rate 0.1):
1. attractive poster +1, choose +1, don’t choose —1
2. less attractive poster +1, choose —1, don’t choose +1
SIMULATION 2: FORBIDDEN TOY
Training (100 epochs, learning rate 0.1):
1. Toy +1, play —0.5, don’t play +0.5
2. Toy +1, positive evaluation +0.2, negative evaluation —0.2
3. positive evaluation +1, negative evaluation —1, play +0.7, don’t play —0.7
4. Toy +1, surveillance +1, play +0.5, don’t play —0.5
5. Toy +1, mild threat +1, play +0.5, don’t play —0.5
6. Toy +1, severe threat +1, play —0.8, don’t play +0.8
7. Toy +1, surveillance +1, mild threat +1, play —0.8, don’t play +0.8
8. Toy +1, surveillance +1, severe threat +1, play —1, don’t play +1
No Surveillance, Mild Threat:
Toy +1, mild threat +1, play —1, don’t play +1
No Surveillance, Severe Threat:
Toy +1, severe threat +1, play —1, don’t play +1
Surveillance, Mild Threat:
Toy +1, surveillance +1, mild threat +1, play —1, don’t play +1
Surveillance, Severe Threat:
Toy +1, surveillance +1, severe threat +1, play —1, don’t play +1
SIMULATION 3: FORCED COMPLIANCE
Training (100 epochs, learning rate 0.1):
1. Attitude object +1, for proposal —0.5, against proposal +0.5
2. Attitude object +1, write essay —0.5, don’t write essay +0.5
3. for proposal +1, against proposal —1, write essay +0.7, don’t write essay —0.7
4. Attitude object +1, low payment +1, write essay —0.5, don’t write essay +0.5
5. Attitude object +1, high payment +1, write essay +0.8, don’t write essay —0.8
6. Attitude object +1, coercion +1, write essay +0.8, don’t write essay —0.8
7. Attitude object +1, coercion +1, low payment +1, write essay +0.8, don’t write —0.8
8. Attitude object +1, coercion +1, high payment +1, write essay +1, don’t write —1
Free Choice, Low Payment:
Attitude object +1, low payment +1, write essay +1, don’t write essay —1
Free Choice, High Payment:
Attitude object +1, high payment +1, write essay +1, don’t write essay —1
Forced Compliance, Low Payment:
Attitude object +1, coercion +1, low payment +1, write essay +1, don’t write essay —1
Forced Compliance, High Payment:
Attitude object +1, coercion +1, high payment +1, write essay +1, don’t write essay —1
SIMULATION 4: SEVERITY OF INITIATION
Training (100 epochs, learning rate 0.1):
1. Group +1, join group —0.2, don’t join group +0.2
2. positive evaluation +1, negative evaluation —1, join group +0.7, don’t join group —0.7
3. Group +1, low shock +1, join group —0.2, don’t join group +0.2
4. Group +1, high shock +1, join group —1, don’t join group +1
Initiation, Low Shock:
Group +1, low shock +1, join group +1, don’t join group —1
Initiation, High Shock:
Group +1, high shock +1, join group +1, don’t join group —1
Abelson, R. P., Aronson, E., McGuire, W. J., Newcomb, T. M., Rosenberg, M. J., & Tannenbaum, P. H. (Eds.). (1968). Theories of cognitive consistency: A sourcebook. Chicago, IL: Rand McNally.
Aisa, B., Mingus, B., & O’Reilly, R. (2008). The emergent neural modeling system. Neural Networks, 21, 1146—1152. http://dx.doi.org/10.1016/j.neunet.2008.06.016
Aronson, E., & Mills, J. (1959). The effect of severity of initiation on liking for a group. Journal of Abnormal and Social Psychology, 59, 177—181. http://dx.doi.org/10.1037/h0047195
Bem, D. J. (1972). Self-perception theory. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 6, pp. 1—62). New York, NY: Academic Press.
Brehm, J. W. (1956). Postdecision changes in the desirability of alternatives. Journal of Abnormal and Social Psychology, 52, 384—389. http://dx.doi.org/10.1037/h0041006
Dawson, C. K., O’Reilly, R. C., & McClelland, J. L. (2003). The PDP++ Software User’s Manual, version 3.0. Pittsburgh, PA: Carnegie-Mellon University.
Fazio, R. H. (2007). Attitudes as object-evaluation associations of varying strength. Social Cognition, 25, 603—637. http://dx.doi.org/10.1521/soco.2007.25.5.603
Festinger, L. (1957). A theory of cognitive dissonance. Palo Alto, CA: Stanford University Press.
Festinger, L., & Carlsmith, J. M. (1959). Cognitive consequences of forced compliance. Journal of Abnormal and Social Psychology, 58, 203—210. http://dx.doi.org/10.1037/h0041593
Freedman, J. L. (1965). Long-term behavioral effects of cognitive dissonance. Journal of Experimental Social Psychology, 1, 145—155. http://dx.doi.org/10.1016/0022-1031(65)90042-9
Gerard, H. B., & Mathewson, G. C. (1966). The effects of severity of initiation on liking for a group: A replication. Journal of Experimental Social Psychology, 2, 278—287. http://dx.doi.org/10.1016/0022-1031(66)90084-9
Isen, A. M. (2000). Positive affect and decision making. In M. Lewis & J. M. Haviland-Jones (Eds.), Handbook of emotions (2nd ed., pp. 417—435). New York, NY: Guilford Press.
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, 245—254. http://dx.doi.org/10.1037/h0021220
O’Reilly, R. C. (1996). Biologically plausible error-driven learning using local activation differences: The generalized recirculation algorithm. Neural Computation, 8, 895—938. http://dx.doi.org/10.1162/neco.1922.214.171.1245
O’Reilly, R. C., & Munakata, Y. (2000). Computational explorations in cognitive neuroscience: Understanding the mind by simulating the brain. Cambridge, MA: The MIT Press.
Read, S. J., & Miller, L. C. (1994). Dissonance and balance in belief systems: The promise of parallel constraint satisfaction processes and connectionist modeling approaches. In R. C. Schank & E. J. Langer (Eds.), Beliefs, reasoning, and decision making: Psycho-logic in honor of Bob Abelson (pp. 209—235). Hillsdale, NJ: Erlbaum.
Read, S. J., & Simon, D. (2012). Parallel constraint satisfaction as a mechanism for cognitive consistency. In B. Gawronski & F. Strack (Eds.), Cognitive consistency: A fundamental principle in social cognition. New York, NY: Guilford Press.
Read, S. J., Vanman, E. J., & Miller, L. C. (1997). Connectionism, parallel constraint satisfaction processes, and gestalt principles: (Re)introducing cognitive dynamics to social psychology. Personality and Social Psychology Review, 1, 26—53. http://dx.doi.org/10.1207/s15327957pspr0101_3
Schwarz, N., & Clore, G. L. (2007). Feelings and phenomenal experiences. In E. T. Higgins & A. Kruglanski (Eds.), Social psychology. Handbook of basic principles (2nd ed., pp. 385—407). New York, NY: Guilford Press.
Shultz, T. R., & Lepper, M. R. (1996). Cognitive dissonance reduction as constraint satisfaction. Psychological Review, 103, 219—240. http://dx.doi.org/10.1037/0033-295X.103.2.219
Shultz, T. R., & Lepper, M. R. (1998). The consonance model of dissonance reduction. In S. J. Read & L. C. Miller (Eds.), Connectionist models of social reasoning and social behavior (pp. 211—244). Hillsdale, NJ: Erlbaum.
Shultz, T. R., & Lepper, M. R. (1999). Computer simulation of cognitive dissonance reduction. In E. Harmon-Jones & J. Mills (Eds.), Science conference series. Cognitive dissonance: Progress on a pivotal theory in social psychology (pp. 235—265). Washington, DC: American Psychological Association. http://dx.doi.org/10.1037/10318-010
Shultz, T. R., Léveillé, E., & Lepper, M. R. (1999). Free choice and cognitive dissonance revisited: Choosing “lesser evils” vs. “greater goods.” Personality and Social Psychology Bulletin, 25, 40—48. http://dx.doi.org/10.1177/0146167299025001004
Simon, D., & Holyoak, K. J. (2002). Structural Dynamics of Cognition: From Consistency Theories to Constraint Satisfaction. Personality and Social Psychology Review, 6, 283—294. http://dx.doi.org/10.1207/S15327957PSPR0604_03
Spellman, B. A., Ullman, J. B., & Holyoak, K. J. (1993). A coherence model of cognitive consistency: Dynamics of attitude change during the Persian Gulf War. Journal of Social Issues, 49, 147—165. http://dx.doi.org/10.1111/j.1540-4560.1993.tb01185.x
van Overwalle, F., & Jordens, K. (2002). An adaptive connectionist model of cognitive dissonance. Personality and Social Psychology Review, 6, 204—231. http://dx.doi.org/10.1207/S15327957PSPR0603_6
1They could also be done in cs++ in emergent (Aisa, Mingus, & O’Reilly, 2008; see https://grey.colorado.edu/emergent/index.php/Main_Page), the newly rewritten version of PDP++.