Research Of Improved Extreme Learning Machine Based On Sequence Data

Compared with other neural network, standard Extreme Learning Machine(ELM)has less training time, good fitting effect, perfect generalization, excellent performance and simple algorithm. However, the limit of standard ELM is that belongs to the batch processing model, when the network model training, training data need to be processed simultaneously. In real life application, the training data is usually obtained in the form of bulk. Online Sequential Extreme Learning Machine(OS-ELM) can handle the situation that data obtained serialized, which the value of the output weights updated through continuous iteration. In this paper, the algorithm derivation and network model of ELM and OS-ELM were studied and analyzed. A modified Fast Sequence of Extreme Learning Machine(FS-ELM) was proposed in this paper.Through the study of algorithm formula in ELM and OS-ELM, the matrix region independent two parts can be superimposed with each other, when the orthogonal method was used to calculate the output weights. With the use of the two partial matrices which independent of the training data, the value of these parts will be constant superposed while the training data is obtained sequentially. And then the output weights was calculated by using the right output value calculation formula t when data acquisition is completed or need to calculate the value of the output weights. Experimental results show that in the initialization of the learning phase during training FS-ELM can accept any number of training data without affecting the accuracy of training and test impact. FS-ELM has a faster speed increase compared to OS-ELM in data training. And it ensures that the test accuracy is quite similar comparing with ELM and Online Sequence Extreme Learning Machine OS-ELM. In order to verify the speed and accuracy performance which FS-ELM possesses, a number of adequate comparative experiments on different scale datasets are conducted.ELM algorithm is simple in part because it requires fewer parameters manually set in. Wherein the number of hidden layer nodes need to be set manually based on the training data. At the same time, the numbers of hidden layer nodes in the network model impact the generalization ability. This paper presents the concept of the hidden layer nodes average impact factor in the sequence of obtaining training data while updating the number of hidden nodes, and gradually so that the total impact of the average output of the lowest weights, the network model is automatically achievegood generalization effect. Experiments show that the sequence of pruning algorithm proposed in this paper can automatically get a good number of hidden nodes, so that the network model has good fitting and generalization.In addition, since the sequence of the ultimate learning machine algorithm proposed in this paper is part of the matrix with respect to the training data is independent. Thus, the algorithm can be easily implemented in the proposed framework Map-Reduce based on Hadoop. With use of distributed Map-Reduce framework, the training model can handle larger amounts of training data which can not be achieved in a single machine. At the same time, the training efficiency of the network model can get better improved by making use of the ability to scale Hadoop platform.