Random Weights Neural Networks Based On Conjugate Gradient Method

In recent years, the problem of large data analysis becomes a hot research in machine learning field. We can often see big data used in real life applications: such as Internet banking, credit card, network intrusion detection. So the research on big data problem has important theoretical significance and application value, we hope to study a feasibility algorithm that can deal with large data problem quickly and efficiently.Extreme Learning Machine( ELM) is a fast learning algorithm proposed by Professor Huang Guangbin and his collaborators. The purpose of this algorithm is to use the random mechanism to reduce the setting and selection of parameters, which greatly improve the learning speed and provide generalization ability.The ELM algorithm can be successfully used in many fields, but some shortcomings such as long training time and large data problems can not be processed in practice. So it is very important to speed up the training speed of ELM and to deal with the problem of large data.The main research work of this paper is that: for the long time of ELM training, it is unable to deal with the big data problem. Based on the existing frame of ELM, a new algorithm is proposed.In the experimental test, we found that elm training speed slow and not handle large data is mainly because of the ELM hidden nodes of the output matrix is by singular value decomposition for generalized inverse calculates. This method occupied much computer memory space and the operation process is complex. Conjugate gradient method as a kind of iterative method, uses less memory, has fast convergence speed, and the convergence process is stable. We hope that will be a combination of both, is proposed based on conjugate gradient randomly weighted neural network(CG-RWNNs):it first by equation method of hidden layer output matrix of pre treatment to treatment conditions, and conjugate gradient algorithm is used to calculate implied a value for the node. CG-RWNNs occupy less system resource and iterates fast. On the basis of ensuring, it can shorten the training time, and handle large data sample number. In last, The validity of the algorithm is proved by theoretical analysis.