Neural Dynamics under Active Inference: Plausibility and Efficiency of Information Processing.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Information Theory

Background:

• Active inference provides a normative framework for behavior, rooted in the free energy principle of self-organization.
• It links neuronal dynamics to state-estimation via descent on variational free energy, a measure of model-data fit.
• Prediction error, encoded in neuronal membrane potentials, drives this process, while firing rates reflect state probabilities.

Purpose of the Study:

• To demonstrate the consistency of active inference neuronal dynamics with existing models.
• To establish face validity by synthesizing electrophysiological responses.
• To show that these dynamics approximate natural gradient descent and analyze their metabolic efficiency.

Main Methods:

• Modeling neuronal dynamics within the active inference framework.
• Synthesizing electrophysiological responses to validate the model.
• Comparing information length (metabolic cost) of belief updating between active inference and natural gradient descent.

Main Results:

• Active inference neuronal dynamics are consistent with current neuroscience models and plausible electrophysiological data.
• These dynamics approximate natural gradient descent, an optimization algorithm from information geometry.
• Active inference demonstrates metabolic efficiency in belief updating compared to other schemes.

Conclusions:

• Neural dynamics in active inference are metabolically efficient.
• Biological agents' neural representations may evolve towards steepest descent in information space for optimal inference.