Discrimination in societies of neural networks

• Grant number: 21/05378-8
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: July 01, 2021
• End date: September 30, 2021
• Field of knowledge: Physical Sciences and Mathematics - Physics - General Physics
• Principal Investigator: Nestor Felipe Caticha Alfonso
• Grantee: Mariana Rocha Mercucci
• Host Institution: Instituto de Física (IF). Universidade de São Paulo (USP). São Paulo , SP, Brazil

Abstract

We study polarization and the formation of groups holding antagonistic opinions on a set of issues in the context of agent-based models where the agents are neural network classifiers. Agents exchange binary opinions about multidimensional issues. Several ways may drive polarization and here it is driven by adaptive affective distrust and the inclusion of irrelevant characteristics to the problem being discussed. A central topic in machine learning is how to extract a set of relevant characteristics from a high-dimensional context. Irrelevant characteristics increase the dimension of the input space, leading to the deterioration of performance. Some agents extend the correctly parsed assertion with a set of numbers that are irrelevant to the original classification problem but depend only on the emitter agent. This irrelevant addition acts like disrupting noise, driving agents to effectively learn from a group of similar agents and to unlearn -learn the opposite opinion- from other agents. We use the Entropic Dynamics learning algorithm for neural networks (EDNNA) to introduce information exchange in a model society where the agents are perceptrons. (AU)