The Development of Bayesian generative models

Abstract

The predictive processing (PP) framework is one of the leading theories to explain cognition. According to PP, the brain continuously predicts sensory inputs given its generative model of the world. An interesting question is how such a generative model of the world is developed. Two approaches for developing a generative model are model updating and model revision. Model updating refers to updating the probabilities over the hypothesis in the model. Compared to that, model revision can take place by constructing or reducing a generative model. While most research focuses on the model updating, in this Bachelor thesis, we will investigate the development of a Bayesian model by model revision. Particularly interesting is the question, how development of a generative model compares in terms of accuracy and causal relations between the two existing model revision processes. Model reduction and model construction are compared by using a computer simulation. In general, neither of the approaches converge to the `true' model of the environment. However, both approaches developed a model that captures the association rules of the environment. Despite showing that model development can underlie the process of model reduction, as well as model construction, more research in complex areas is necessary to generalize these fi ndings.