Behavior Learning Of Caenorhabditis Elegans Based On Spiking Neural Network

Caenorhabditis elegans has the advantages of short life cycle,small neural structure,and rich behavior.It is a simple and ideal model organism.After decades of research,the connection map of all neurons in the C.elegans nervous system and the complete nervous system has been described in detail.However,many electrophysiological data of C.elegans nervous system are still missing,such as the excitability and inhibition of synapses,and connection strength.This is also one of the main obstacles to using its anatomical model for behavioral modeling and analysis.In this thesis,the impulse neuron and the biological nervous system architecture of C.elegans are used to construct the C.elegans impulse neural network model,and the improved genetic algorithm is used to optimize the network parameters to achieve light touch stress behavior and chemotaxis behavior,which can be used for studying the mechanism of C.elegans nervous system behavior.The specific work of this thesis is as follows:1.The common method of modeling the nervous system of C.elegans is to use feedforward artificial neural network models or local nervous system models related to nematode specific behaviors.These models do not conform to the biological nervous system in terms of neuron characteristics or network topology.And the modeling of the local neural network cannot reflect the cross-connectivity of the nematode nervous system.In addition,more and more biological studies have found evidence that nematode neurons can emit action potentials.Therefore,based on the biological nervous system of C.elegans,this thesis adopts the impulse neuron model to construct the C.elegans impulse neural network model.The connection relationship between neurons in the model is completely in accordance with the nervous system architecture of the real nematode,including 302 neurons and all synaptic connections.Dividing the two-way connection between neurons into two separate one-way connections,the model has a total of 5907 unknown neuron connection coefficients.The spiking neural network model contains a large number of feedback connections and long-and short-range connections.The topological structure is consistent with the C.elegans nervous system,and the neuron model conforms to the characteristics of the dynamic changes of biological neurons.Therefore,the network model constructed in this thesis has a higher similarity with the nervous system of real nematodes.2.Simulate the environmental stimulus received by C.elegans as the input sequence of the above-mentioned impulse neural network model,and simulate the motion generated by the environmental stimulus of C.elegans as the target output sequence of the impulse neural network model.The behavior of C.elegans can be represented by the input sequence and target output sequence of the spiking neural network model.Through the above methods,this thesis models the light touch stress behavior and chemotaxis behavior of C.elegans.The improved genetic algorithm is used to optimize the parameters of the C.elegans pulse neural network model,so that the actual output sequence of the network fits the target output sequence.Finally,the C.elegans spiking neural network model has the ability to simulate these two behaviors of the nematode.3.The above-mentioned spiking neural network model's learning process of nematode behavior can be understood as the parameter optimization process of the model.Since this spiking neural network model is not a feedforward neural network model with a hierarchical structure,the model parameters cannot be optimized by the gradient of the back propagation error according to the layer.The classic genetic algorithm is suitable for parameter optimization of any model.Because the crossover operation of genetic algorithm tends to destroy the discovered features during the parameter optimization process of the neural network model,5907 parameters of the model are optimized using genetic algorithm with continuously declining crossover probability.This article demonstrates the feasibility of the above parameter optimization method through experiments.