Sensorimotor control with thalamocortical-based spiking neural networks

Abstract

Large strides have been made in understanding the anatomy of the brain, however, how this anatomy contributes to complex computations - such as controlling movement - remains more elusive. Building cortical models that are representative of the brain’s anatomy is one way of closing this gap. An issue with most systems neuroscience models is that they lack neuronal diversity and cortically representative anatomy and connectivity. Here we built a liquid state machine (LSM) with cytoarchitecture and hierarchical connectivity inspired by the thalamocortical pathway to perform motor control in a ballistic reaching task. The LSM is a spiking neural network reservoir computing model. The reservoir is comprised of two network areas - loosely corresponding to sensory and motor areas. Each area contains Izhikevich models of pyramidal, large basket, and spiny stellate neuron populations which were laminarly organized to resemble cortical layers 2/3, 4, and 5. Physiological data of mammalian microcircuitry informed how these populations were interconnected within and between network areas. To condition the model on the input statistics, unsupervised learning was implemented via global spike-timing-dependent plasticity. Subsequently, a three-layer feed-forward neural network (decoder) was trained to translate the LSM activity into motor commands which move a point mass in 2D space. To map spiking activity to desired muscle activation patterns, the decoder weights and biases were adjusted with adaptive moment estimation optimization to match optimal muscle force trajectories generated by a linear quadratic Gaussian regulator. Comparing frequency vs. current (F/I) curves of simulated neurons to in-vivo neurons, we replicated physiologically realistic firing rates for multiple neuron types. The LSM's sensorimotor control performance was tested with ballistic reaches. Overall, the model was successful in generating motor commands that closely matched those of optimal trajectories. Unsupervised learning within the LSM prior to supervised learning reduced the overall network excitability, but reach performance was maintained. Beyond its ability to perform sensorimotor control, our biophysically-based neural network model provides a viable starting point to investigate sensorimotor control from various angles, such as population coding, neural manifolds, and the contribution of different neuron types to cortical computations.