| Recently, a variety of portable and wearable devices such as intelligent watches, intelligent bracelets etc., have attracted more and more people’s attention and affection. It is portable and wearable that these devices are. However, there are some intelligent electronic gadgets, which always suffer from the poor accuracy due to the limited computational storage resources. To address this issue, a novel Two-hidden-layer Extreme Learning Machine(TELM) is proposed in this thesis. Specifically, this algorithm adds one more hidden layer to the Single Hidden Layer Feedforward Neural Network(SLFN) that is the structure of the original Extreme Learning Machine(ELM). Therefore, it is a kind of Two-hidden-layer Feedforward Neural Network(TLFN) algorithm. In TELM, a novel method to obtain the parameters of the second hidden layer(connection weights between the first and second hidden layer and the bias of the second hidden layer) is introduced. Thus, make the actual hidden layer output for the second hidden layer closer to the expected hidden layer output for the second hidden layer. At the same time, it inherits the randomness of the original ELM for the first hidden layer(connection weights between the input weights and the first hidden layer and the bias for the first hidden layer). Experimental results show that TELM can consistently outperform the original ELM and some multilayer ELM variants, in terms of average accuracy and the number of hidden neurons, and also can be applied to the intelligent devices with limited storage.The structure of this thesis is as following:First of all, the background of the original extreme learning machine is briefly introduced.Secondly, this thesis explains some basic principles and concepts about the original extreme learning machine and also specifically analyses the advantages and disadvantages for the original extreme learning machine. Then, some improved algorithms and relevant applications of the original extreme learning machine, proposed by some experts and researchers, are explained briefly.Thirdly, a novel Two-hidden-layer Extreme Learning Machine(TELM) is proposed in this thesis, and some principles regarding this algorithm are illustrated in details. The algorithm can effectively improve both training accuracy and test accuracy in regression and classification problems, Moreover, it has better generalization performances. In TELM, the parameters for the first hidden layer are generated randomly just as the original extreme learning machine does. In this way, it can successfully maintain some merits of ELM and ensure the fast training speed. As for the parameters of the second hidden layer, a new method of parameter calculation is designed to obtain them. By doing so, it is easy to seek a better mapping between the input and the output so that it successfully improves the performance of the algorithm. To evaluate the performance of the proposed TELM, we select 3 optimizations to conduct regression experiments and 5 simple benchmark classification datasets, namely Vowel, Satellite, Segment, Optdigits as well as Pendigits, and a more complex classification dataset——MNIST dataset to conduct classification experiments. The experimental results demonstrate that TELM can achieve both better training accuracy and testing accuracy with a small number of hidden nodes and also prove the feasibility of this algorithm.Finally, TELM is applied to a wrist vein recognition system, which is a typical system with small storage. The experimental results show that TELM can not only get a better accuracy with fewer nodes and its training speed is also of the same order of magnitude as ELM’s and even slightly faster than the ELM’s. In addition, the results also demonstrate that TELM can provide an effective solution to portable devices with less storage resource. For this reason, these devices always hardly reach satisfying classification accuracy due to the limited storage capacity. |
