Bayesian Mechanics of Synaptic Learning Under the Free-Energy Principle

Area of Science:

• Computational Neuroscience
• Theoretical Neuroscience
• Cognitive Science

Background:

• The brain operates as a complex biological system, orchestrating perception, behavior, and learning.
• Karl Friston's free-energy principle (FEP) offers a framework for understanding higher-order cognitive functions through self-supervised dynamics.

Purpose of the Study:

• To refine the free-energy principle (FEP) using a physics-guided approach.
• To model synaptic learning as an inference problem within the FEP framework.
• To derive governing equations for synaptic plasticity, termed Bayesian mechanics.

Main Methods:

• Applied a physics-guided formulation to the free-energy principle.
• Treated synaptic learning as an inference problem under the FEP.
• Derived Bayesian mechanics equations governing synaptic plasticity.
• Utilized generative models for likelihood and prior beliefs to explain neural inference.

Main Results:

• Uncovered mechanisms by which the brain infers synaptic weight changes and postsynaptic activity based on presynaptic input.
• Demonstrated that synaptic learning involves the brain deploying generative models of likelihood and prior beliefs.
• Illustrated optimal trajectory organization in neural phase space during continuous-time synaptic learning.
• Showed that this process variationally minimizes synaptic surprisal.

Conclusions:

• Synaptic learning can be understood as an inference problem governed by Bayesian mechanics derived from the FEP.
• The brain actively infers neural states and synaptic modifications to adapt to dynamic environments.
• This framework provides a novel perspective on neural computation and synaptic plasticity.