Research On Delayed Blinking Based On Real-time Biomimetic Cerebellar Model

The research of the cerebellum has never stopped,and in order to investigate the changes of neurons in the cerebellum during motor learning,the Pavlovian delayed blink conditioned reflex experimental paradigm was used to investigate the changes of neurons in the cerebellum during motor learning.The delayed blink conditioned reflex uses a pair of conditioned and unconditioned stimuli,and by repeatedly pairing the stimuli the animal learns to respond to the unconditioned stimulus when the conditioned stimulus signal is present.Using electrophysiological acquisition equipment to record the potential changes of Purkinje cells in the cerebellum after learning the experimental paradigm,the patterns of appearance of simple and complex spikes were counted.Finally,it was found that the complex spikes were mainly concentrated near the unconditioned stimulus signal,and it was found that the mice were able to predict the time of unconditioned stimulus appearance and respond in advance by training.In order to better investigate the generation of this phenomenon,this thesis uses a personal computer hardware platform with a graphics card to build a real-time simulation of a cerebellar model neural network to simulate the process of this experimental paradigm.The cerebellar model is an impulsive neural network with forward feedback constructed by imitating the neural circuits and spike firing mechanism of neurons on the cerebellar cortex.The model can better simulate the training results of the delayed blink reflex and can output the impulse signal before the guidance position as the training progresses.However,the real-time simulation of the cerebellar model neural network has problems such as small simulation scale,large fluctuations in simulation timing and inability to make full use of hardware performance.Therefore,based on this thesis,we propose an optimization scheme for parallel processing of many different computational modules.Based on the streaming feature of CUDA(Compute Unified Device Architecture),the model size is larger and real-time is maintained using the streaming scheme under the same hardware conditions.The cerebellar model in this experiment,like most of the cerebellar spiking neural network models,is coarse for cell potential calculation and cannot well represent the properties of Purkinje cells in electrophysiological recordings,so we want to introduce a more refined Purkinje cell model.In this thesis,a five-current model is used to ensure the realism of the simulated potentials with low computational complexity,and the sodium channels regulated by Markov chains can show more biological characteristics.The final model is able to simulate the shape of complex spikes and the characteristics of potential changes after complex spikes by fitting the model output to the electrophysiological recordings of Purkinje cells using a downhill simplex algorithm.The real-time simulation of the cerebellar model neural network has research implications for exploring the mechanism of cerebellar operation and is valuable for the study of nonlinear robot controllers.