Spiking Neural Network-Based Modeling of Human Motor Cortex for Robotic Limb Control

Abstract

Neural modelling of robotic limb control has become a paradigm shift field of research in neuroprosthetics, rehabilitation engineering, and brain-machine interfaces (BMIs), because the ultimate goal is to ensure that individuals with motor impairments regain natural standard motion. Traditional artificial neural network (ANN)-based of decoding systems though good at pattern recognition lack the biological accuracy, event-driven communication and time evolution of the neuronal firing in the cortex. In order to overcome these shortcomings, this paper introduces a biologically-motivated computational theory that consists of a biological model of the human motor cortex, a spiking neural network (SNNs), a types of neural network that is direct simulation of real neuron-membrane systems and represents their spike-timing behaviour. The offered architecture incorporates an entire processing pipeline, which receives neural activity of either electroencephalography (EEG), electrocorticography (ECoG), or intracortical microelectrode records, encodes neural activity to spike trains using sophisticated encoding algorithms, and maps the encodings into high-level movement intentions. Such notions are then converted into smooth and accurate motor trajectories to execute robotic limbs through an hierarchical motor cortex model based on an SNN that comprises of premotor, primary motor and corticospinal modules. By combining Spike-Timing-Dependent Plasticity (STDP), reward-modulated Spike-Timing-Dependent Plasticity (R-STDP), and surrogate-gradient backpropagation, a hybrid learning strategy can be used in order to achieve both unsupervised, reinforcement-based and supervised adaptation, providing biologically-plausible and computationally-efficient learning. Simulations performed on artificial intracortical spike data as well as real EEG data of motor imagery targeting movements evidenced that the SNN model is more accurate in movement prediction than ANN baselines, has a much lower latency to generate motor commands, and is less susceptible to neural noise and signal distortions. Moreover, the model is highly energy effective thus would be applicable in future neuromorphic hardware realisations. All in all, the results suggest the tremendous potential of the SNN -based motor cortex emulation to the next-generation intelligent prostheticics, and provides the new avenues towards adaptive, low-power, and high-fidelity BMIs that would support intuitive robotic limb control.