Modelling visual decision making using a variational autoencoder

Due to information processing constraints and cognitive limitations, humans necessarily form limited representations of complex visual stimuli when making utility-based decisions. However, it remains unclear what mechanisms humans use to generate representations of visual stimuli that allow them to make predictions of utility. In this paper, we develop a model that seeks to account for the formation of representations in utility-based economic decision making. This model takes the form of a β-variational autoencoder (β-VAE) trained with a novel utility-based learning objective. The proposed model forms representations of visual stimuli that can be used to make utility predictions, and are also constrained in their informational complexity. This representation modelling approach shares common features with related methods, but is unique in its connection to utility in economic decision making. We show through simulation that this approach can account for several phenomena in human economic decision making and learning tasks, including risk averse behaviour and distortion in the calculation of expected utility.