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

Abstract

AbstractOne of the most fundamental laws of physics is the principle of least action. Motivated by its predictive power, we introduce a neural least-action principle that we apply to motor control. The central notion is the somato-dendritic mismatch error within individual neurons. The principle postulates that the somato-dendritic mismatch errors across all neurons in a cortical network are minimized by the voltage dynamics. Ongoing synaptic plasticity reduces the somato-dendritic mismatch error within each neuron and performs gradient descent on the output cost in real time. The neuronal activity is prospective, ensuring that dendritic errors deep in the network are prospectively corrected to eventually reduce motor errors. The neuron-specific errors are represented in the apical dendrites of pyramidal neurons, and are extracted by a cortical microcircuit that ‘explains away’ the feedback from the periphery. The principle offers a general theoretical framework to functionally describe real-time neuronal and synaptic processing.