A free energy principle for the brain.

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Statistical Physics

Background:

• Helmholtz's theories on perception are reformulated using modern computational frameworks.
• Existing neurobiological data presents challenges for understanding perceptual inference and learning.
• A unified model is needed to explain the brain's ability to infer causal structures from sensory input.

Purpose of the Study:

• To develop a biologically plausible model of perceptual inference and learning.
• To explain neurobiological facts using principles from statistical physics and hierarchical Bayesian models.
• To demonstrate that perception and action are aspects of minimizing free energy.

Main Methods:

• Formulating perceptual inference and learning using constructs from statistical physics.
• Employing Empirical Bayes and hierarchical models to represent sensory input causation.
• Applying the free energy principle to model adaptive exchange with the environment.

Main Results:

• A unified framework explains inference and learning using the same principles, consistent with neurobiological evidence.
• Hierarchical models allow for dynamic, context-sensitive prior expectations.
• Minimizing free energy explains perception, action, and their adaptive role in biological systems.

Conclusions:

• The proposed scheme provides a principled understanding of cortical organization and responses.
• Brain dynamics and structure can be explained by the minimization of free energy.
• Systems conforming to the free energy principle exhibit adaptive behaviors through perception and action.