Recurrent Spiking Neural Networks for Continuous Motor Control: A Biologically Plausible Model of Motor Cortex
Authors:
Preprint
Abstract
Motor cortex acquires skilled movements through recurrent spiking dynamics and experience-dependent synaptic plasticity, yet how such circuits learn continuous motor control using biologically local signals remains unclear. Here, we developed a recurrent spiking neural network model of motor cortex that addresses both the computational and biological requirements of motor learning. We first analytically derived the recurrent firing-rate state directly from synaptic input and trained the network by implicit differentiation, thereby avoiding backpropagation through long spike trajectories. We then introduced State-Carry Initialization, which enables closed-loop motor control by continuously carrying the ongoing recurrent state across successive actions. Building on this formulation, we devised a learning rule termed Feedback-Spike Nudging (FSN), in which teaching signals are delivered as spikes through a feedback pathway. Synaptic changes depend only on presynaptic activity and the feedback-evoked change in postsynaptic firing, consistent with biological mechanisms of synaptic plasticity in motor cortex. The resulting networks learned continuous motor control through both supervised and reinforcement learning, which could be mapped respectively onto interactions of motor cortex with the cerebellum and basal ganglia. Moreover, the FSN learning rule aligned with experimental observations of structural spine plasticity in mouse motor cortex. Thus, our study lays a foundation for recurrent spiking models of motor cortex that are biologically grounded at both the functional and anatomical levels.Keywords
This publication has 52 references indexed in Scilit:
- Neural Sequences as an Optimal Dynamical Regime for the Readout of TimeNeuron, 2020
- Random synaptic feedback weights support error backpropagation for deep learningNature Communications, 2016
- Distinct Eligibility Traces for LTP and LTD in Cortical SynapsesNeuron, 2015
- Branch-specific dendritic Ca2+ spikes cause persistent synaptic plasticityNature, 2015
- Cerebellar networks with the cerebral cortex and basal gangliaTrends in Cognitive Sciences, 2013
- Error-backpropagation in temporally encoded networks of spiking neuronsNeurocomputing, 2002
- Learning of action through adaptive combination of motor primitivesNature, 2000
- What are the computations of the cerebellum, the basal ganglia and the cerebral cortex?Neural Networks, 1999
- A new cellular mechanism for coupling inputs arriving at different cortical layersNature, 1999
- Multiple paired forward and inverse models for motor controlNeural Networks, 1998