The graphical brain: Belief propagation and active inference.

Area of Science:

• Computational neuroscience
• Neuroscience
• Artificial Intelligence

Background:

• Functional integration in the brain is crucial for complex cognitive processes.
• Understanding neuronal message passing is key to deciphering brain computation.
• Active inference provides a framework for modeling brain function.

Purpose of the Study:

• To investigate neuronal message passing within the framework of active inference.
• To explore the implications for context-sensitive neural connectivity at various scales.
• To formulate neuronal processing using belief propagation under deep generative models.

Main Methods:

• Formulated neuronal processing as belief propagation under deep generative models.
• Utilized Forney (normal) factor graphs to characterize message passing.
• Developed methods to accommodate mixed generative models with discrete and continuous states.

Main Results:

• Identified distinct belief updating schemes for discrete and continuous states on a shared neuronal architecture.
• Demonstrated how Bayesian model averaging and comparison can be implemented in thalamocortical loops.
• Showcased state-dependent and self-organizing computational connectome.

Conclusions:

• Neuronal message passing, as dictated by active inference, enables context-sensitive brain connectivity.
• Deep generative models offer a unified framework for discrete and continuous information processing in the brain.
• Simulations of reading illustrate the integration of semantic and visual information processing.