Polarization Through the Lens of Learning Theory

We present a fresh perspective on belief polarization, based on two learning-theoretic models. In the first model, two agents learn from training sets drawn from different distributions that slightly disagree on some labels. In the second model, two agents learn from training sets sampled from the very same distribution, but pay a cost for the complexity of the hypotheses they learn. We show that in both cases, even when the agents are exposed to almost identical sources of information, they can learn hypotheses that disagree substantially. However, in the latter model, we demonstrate that this phenomenon can be alleviated by introducing a slight bias into the information selection process.