The Research On The Competitive Layer Model Of The Lotka-Volterra Neural Network And It's Applications

The competitive layer model (CLM) of neural network is a new structure of neuralnetwork. This model consists of several layers, the neurons in each layer connect toeach other, and the neurons in each column also connect to each other. The CLM isconsidered to have biological background, because that the similar competitive structurehad been found in visual nervous system. From view of mathematic, The CLM can bedescribed as an optimization problem expressed as an energy function. Then, the CLMcan be solved by designing a recurrent neural network (RNN).The CLM of the Lotka-Volterra (LV) RNN has rigorous theoretical foundation. Ithad been proven that the set of minimum points of the energy function equals to the setof stable attractors of the LV RNN. In addition, the CLM of the LV RNN has theproperty of feature binding, which can bind similar features into same layer. Theproperty of feature binding is considered to have extensive application prospect. So, thisthesis mainly researches the CLM of the LV RNN and its applications,including:1)researchs the property of feature binding and its applications;2) constructs new LVRNNs for solving some applications by extending the CLM of the LV RNN.The main results are as follows:(1)The property of feature binding of the CLM of the LV RNN is researched. Ageneral feature binding algorithm is given, a method of binding features into appointedlayer is proposed, and a method of binding dived features into appointed layer is alsopresented.(2)A method of detecting brain activity is proposed. The proposed method clustersthe time series of voxels in fMRI data by using the CLM of the LV RNN, and thenfinishes the task of detecting brain activity by binding active voxels into "active" layers.Experiments on synthetic and real fMRI data demonstrate the activated voxels can bedetected more accurately than some existing methods by the proposed method.(3) A method based on the CLM of the LV RNN is proposed for brain MR imagesegmentation. The method segments the brain MR image into background, CSF, GMand WM by using3steps which include small block image segmentation, region merging and region clustering. Compared with other three methods using simulations,the proposed method is shown to be more effective.(4)A method of extracting long contour is proposed. The method constructs acontour extractor which can bind collinear or co-circular edges into same layer. Themethod can extract a long contour by moving the contour extractor along the longcontour.(5)A method of region extraction is proposed. The method can extract a specifiedobject region through dealing with small block image one by one by using the CLM ofthe LV RNN. When dealing with each small block image, the object region can bebound into the "object region" layer via the method of binding dived features intoappointed layer. Experiments show that the proposed can exactly the object region fromcolor images.(6)The traveling salesman problem (TSP) can be described as an energy function.Base on the energy function, a new LV RNN is constructed for solving the TSP. Thefoundations of solving the TSP using the new LV RNN are obtained, including:1)anyone minimum point of the energy function is corresponding with a valid tour;2)eachinteresting trajectory are convergent;3)the set of minimum points of the energyfunction equals to the set of stable attractors of the LV RNN.(7) The shortest path (SP) can be described as an energy function. Base on theenergy function, a new LV RNN is constructed for solving the SP problem.Experiments show that the shortest path between any two points in the directed graphcan be obtained by using the constructed LV RNN.