Predictive processing simplified: The infotropic machine

Abstract

On a traditional view of cognition, we see the agent acquiring stimuli, interpreting these in some way, and producing behavior in response. An increasingly popular alternative is the predictive processing framework. This sees the agent as continually generating predictions about the world, and responding productively to any errors made. Partly because of its heritage in the Bayesian brain theory, predictive processing has generally been seen as an inherently Bayesian process. The ‘hierarchical prediction machine’ which mediates it is envisaged to be a specifically Bayesian device. But as this paper shows, a specification for this machine can also be derived directly from information theory, using the metric of predictive payoff as an organizing concept. Hierarchical prediction machines can be built along purely information-theoretic lines, without referencing Bayesian theory in any way; this simplifies the account to some degree. The present paper describes what is involved and presents a series of working models. An experiment involving the conversion of a Braitenberg vehicle to use a controller of this type is also described.