Modeling Cognitive Dissonance Using a Recurrent Neural Network Model with Learning

This paper presents a recurrent neural network model of long term attitude change resulting from the reduction of cognitive dissonance. The model uses Contrastive Hebbian Learning (CHL) to capture changes in weight strength among cognitions resulting from dissonance reduction. Several authors have presented recurrent network models of dissonance reduction that capture the constraint satisfaction nature of dissonance. But as these models did not learn they could only model short-term attitude change represented by changes in activatio. In response, Van Overwalle and Jordens (2002) presented a feedforward model, with delta rule learning, in an attempt to capture long-term attitude change caused by dissonance reduction. However, the feedforward nature of their model created two problems. First, it could not capture the parallel constraint satisfaction mechanisms that underlie dissonance reduction. Second, and perhaps more important, the network was not able to “reason” backwards from its inconsistent behavior to its new attitude, but instead had to be explicitly taught its new attitude. The present model overcomes the weaknesses of the previous approaches to modeling dissonance reduction. Because it has learning it can represent long-term attitude change by weight change and because it is a recurrent model it can propagate changes from inconsistent behavior to the attitudes linked to that behavior.