Active Inference: A Process Theory.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Bayesian Brain Hypothesis

Background:

• Neuronal processing and action selection are often explained by maximizing Bayesian model evidence or minimizing variational free energy.
• Understanding the precise neuronal dynamics underlying these principles remains a key challenge in neuroscience.

Purpose of the Study:

• To propose a process theory of neuronal dynamics grounded in active inference and belief propagation.
• To investigate whether neuronal responses can be mathematically described as a gradient descent on variational free energy.
• To demonstrate how this framework can explain a wide array of well-characterized neuronal phenomena and behaviors.

Main Methods:

• Utilized a standard generative model based on Markov decision processes.
• Derived neuronal dynamics by formulating neuronal responses as a gradient descent on variational free energy.
• Applied principles of active inference and belief propagation.

Main Results:

• Successfully reproduced diverse neuronal phenomena including repetition suppression, mismatch negativity, and place-cell activity.
• Demonstrated the framework's ability to explain behaviors such as reward seeking and context learning.
• Established variational free energy as a Lyapunov function for neuronal dynamics, linking them to Hamilton's principle of least action.

Conclusions:

• Neuronal activity and action selection can be formally described as a gradient descent on variational free energy.
• This active inference framework provides a unified explanation for a broad spectrum of neuronal responses and adaptive behaviors.
• The derived neuronal dynamics align with fundamental principles of physics, such as the principle of least action.