A neuronal least-action principle for real-time learning in cortical circuits

Area of Science:

• Neuroscience
• Computational Neuroscience
• Physics

Background:

• The principle of least action is fundamental in physics.
• Real-time processing of sensory streams for behavioral output is crucial for survival.
• Understanding the computational principles of the brain remains a significant challenge.

Purpose of the Study:

• To introduce a neuronal least-action principle for cortical processing.
• To explain how neurons minimize errors for real-time behavioral outputs.
• To provide an axiomatic framework for neuronal and synaptic laws.

Main Methods:

• Postulating that voltage dynamics of cortical pyramidal neurons minimize somato-dendritic mismatch error.
• Proposing error minimization for output and deep network neurons to overcome delays and correct errors.
• Describing error extraction in apical dendrites via a cortical microcircuit.
• Incorporating online synaptic plasticity for error reduction and gradient descent.

Main Results:

• The neuronal least-action principle prospectively minimizes local errors within individual neurons.
• Output neurons minimize instantaneous behavioral error, while deep network neurons fire prospectively.
• The framework integrates sensory input, motor output, and feedback for real-time computation.
• Online synaptic plasticity enables gradient descent on output cost.

Conclusions:

• The neuronal least-action principle offers a unified framework for brain computation.
• It provides a basis for deriving local neuronal and synaptic laws for global real-time learning.
• This principle has implications for understanding sensory-motor transformations and neural computation.