Extreme Nonlinear Discriminant Analysis Network Inspired By ELM

The linear discriminant analysis methods cannot deal with the nonlinear problems,and nonlinearization is the main approach to solve it.Nonlinear discriminant method mainly include neural network and the nucleation methods;Neural network discriminant analysis method not only inherits the adaptivity,distributed storage,parallel processing and advantages of nonlinear mapping of neural network,but also inherited the slow training speed and ability of easily fall into local minimum defects;Kernel linear discriminant analysis method can obtain the global optimal analytical solution,but subject to number of hidden nodes(equal to the number of samples),and the computational cost increases with the data size.Inspired by extreme learning machine(ELM)random mapping,discriminant analysis method to neural network speed transformation,and implements a fast nonlinear discriminant analysis method,compared with the traditional gradient descent learning algorithm,It not only have fast learning speed,but also can obtain the global optimal solution of the characteristic.Compared with KLDA,hidden node parameters do not depend on the number of samples.Based on the above work,we proposed a novel discriminant analysis network with weighted random orthogonalization(O-ENDA).,As an important optimization method,orthogonalization is widely used in image recognition..For small scale of data with high dimension,we do the orthogonalization map locally.,on the one hand,It reduces the redundancy of data information classification,more uniform distribution of random weighting orthogonal And on the other hand,The complex feature extraction makes the examples independent linearly and diverse,in which improve the ability of generalization of models further by weighted random orthogonalization.