Research On The Multi-scale Cognitive Neural Model And Its Application

In the studies of artificial intelligence,the artificial neural network models,especially deep learning models,have fantastic performance in machine learning and the other fields,and were widely used in the engineering.However,the internal connection weights of such models are adjusted by the error back propagation algorithm,and their performance in small sample learning need to be improved,which are inconsistent with the basic laws of neuroscience and cognition.On the other hand,in the studies of cognitive science,most neural models and brain network models focus on the explanation of cognitive phenomena,so they are not suitable for machine learning tasks.Therefore,in this thesis,based on the latest research of cognitive science and neuroscience,also the studies of the activity patterns of the neural system in cognitive phenomena,we establish a multi-scale cognitive neural model that conforms the basic laws of cognitive science and neuroscience,and we further use this model for machine learning tasks.The main works of this thesis are as follow:First,a novel multi-scale cognitive model is established by analysing the structure of the neural system and cognitive behavior.The neural fields in the models consist of neurons with lateral interaction,the functions of different brain regions are simulated by different neural fields.The connection between the neurons in different fields obey the synaptic plasticity rules.Thus,the proposed model is capable of realizing memory and other cognitive tasks.Second,the dynamic properties of the model are studied.The model is a nonlinear dynamic model with multi-scale,and its dynamic properties are complex.In this thesis,the existence and stability of the stationary solutions of the model are analyzed by nonlinear analysis methods and tools such as the fixed point principle,and some important properties are obtained.so the feasibility and effectiveness of the model in cognitive and learning tasks are ensured.Third,the model is used to illustrate the process of visual classification,and a new supervised learning method is proposed for data classification.This method combines the scale adaptive algorithm based on cognitive results,and has good learning performance.It has good adaptability and validity in data classification.Numerical experiments show that the proposed classification method has better performance than some state-of-the-art classification methods in one-shot learning and small sample learning.Fourth,the model is used to illustrate the process of visual clustering,and a new unsupervised learning method is proposed for data clustering.This method adopts a similar way to supervised learning to get the clustering under the given cognitive scale.By expanding the scale,the clustering sequence is obtained.An optimal clustering result selection method based on the change of the distribution of the cluster is given.Numerical experiments show that the proposed clustering method can achieve good clustering results.This model conforms to the basic principles of neuroscience and cognitive science with strong interpretability,its learning processes conform to the cognitive processes,and can realize both supervised learning and unsupervised learning under the same framework.The proposed classification,clustering algorithms do not involve any optimization solution processes and manual parameter tuning,the parameters in the model have convincible biological basis,and can be adapted by the input.The proposed multi-scale cognitive neural model and its theoretical methods are important works in the fields of cognitive science and artificial intelligence.They provide new problems and objects for the research of applied mathematics in the intersecting fields,and have important theoretical significance and application value.