Opinion, Distrust and Discrimination in Interacting Neural Networks
Seminário de Física Estatística com João Henrique de Sant'Ana, Universidade de São Paulo
Em 07/10, 4ª feira, às 15h. Sala de Seminários (2061), Departamento de Física Geral.
Abstract: Models of opinion dynamics often represent individuals as simple agents with no internal structure. Here, we consider a society of agents in which each individual is an adaptive neural network to form binary opinions about public issues and maintains an evolving estimate of how much it distrusts every other agent.
During pairwise interactions, agents update both their opinions and their trust relations using a learning rule derived from a Maximum Entropy approximation to Bayesian learning. The rule is optimal in the limit of high-dimensional issues and small societies. Learning occurs primarily in the presence of dissonance, when the receiver agent disagrees with and trusts the emitter, or, conversely, agrees with yet distrusts the emitter.
Despite being locally optimal, this learning dynamics often drives the population toward strongly polarized states. Polarization becomes less frequent when agents introduce a tolerance field that makes them less sensitive to disagreements. We then divide the population into two arbitrary groups and define prejudiced agents as those who apply different levels of tolerance to members of their own group and to members of the other group.
To connect these microscopic interactions to the collective organization of the population, we construct a phase diagram in the space defined by the bias strength (b) and the fraction of prejudiced agents (f_b), based on three pair correlations: opinion–distrust, distrust–class, and class–opinion. The resulting phase diagram displays distinct regimes of frustration, polarization, and discrimination. We also characterize the balance of social relations through measures of ideological and affective frustration, and show that these collective structures depend not only on prejudice, but also on the number of issues discussed by the population.
Our results illustrate how simple and individually optimal learning rules can generate undesirable collective behavior. More broadly, the model provides measurable quantities that may help connect microscopic mechanisms of opinion formation and trust to macroscopic patterns of polarization and discrimination.
