Predictive processing simplified: The infotropic machine.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Information Theory

Background:

• Traditional cognition involves stimulus interpretation and behavioral response.
• Predictive processing (PP) is an alternative framework where agents predict world states and correct prediction errors.
• PP is often viewed as inherently Bayesian due to its roots in Bayesian brain theory.

Purpose of the Study:

• To demonstrate that hierarchical prediction machines can be specified using information theory.
• To propose an alternative to the Bayesian approach for predictive processing.
• To simplify the account of predictive processing by removing reliance on Bayesian theory.

Main Methods:

• Deriving a specification for hierarchical prediction machines directly from information theory.
• Utilizing the metric of predictive payoff as an organizing concept.
• Developing working models of information-theoretic hierarchical prediction machines.

Main Results:

• A specification for hierarchical prediction machines derived from information theory, independent of Bayesian theory.
• Demonstration that predictive processing can be grounded in information theory.
• Successful implementation of working models and an experimental validation using a Braitenberg vehicle.

Conclusions:

• Hierarchical prediction machines can be constructed using purely information-theoretic principles.
• This information-theoretic approach simplifies the understanding and implementation of predictive processing.
• The findings offer a novel perspective on the theoretical underpinnings of cognitive architectures.