Abstract: This paper consists of two parts: an application of amultilayer perceptron model to the detecting ofparticle motion and a method of expediting the study ofthe multilayer perceptron training procedure. In thefirst part, the motion of a single particle or multipleparticles is detected by identifying the reflectionsymmetry of two concatenated image frames. The accuracywill be higher than 90%, even though a certain amountof noise exists. In the second part, a convergent indexparameter formularized to evaluate the distance of thenetwork inner state from the Boltzmann distribution isput forward for the measurement of the trainingprocedure convergency. With the convergent index, itbecomes feasible to study the influence of the networkparameters (sizes of different layers, targetassigning, etc.) on the training procedure before thenetwork has reached a stable state; thus time is saved.This method is used in the designing of the network.!
展开▼
