Font Size: a A A

Research And Application Of Deep Extreme Learning Machine

Posted on:2017-04-11Degree:MasterType:Thesis
Country:ChinaCandidate:J WeiFull Text:PDF
GTID:2308330503457298Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
In recent years, deep learning has become a new study in the field of machine learning. The multi-layer architecture contributes to layer-wise feature extraction for deep neural network to form higher-order feature representation; Greedy layer-wise unsupervised training can be regard as initialization settings of the supervised fine-tuning process, solving the problem which is difficult to obtain training labels in the big data Era. However, gradient-based optimization criterion is more used in the supervised training of deep neural network. In this learning framework, all hidden layer parameters need to be trained iteratively. Therefore, the deep structure is hard to achieve good generalization performance with fast learning speed.In view of the slow training speed and poor generalization performance of the gradient-based learning algorithm, a handy and efficient machine learning algorithms is proposed by Huang et al.——Extreme Learning Machine(ELM). In this strategy, according to any continuous probability distribution, the hidden layer output is randomly chosen by Single-Hidden Layer Feed-forward Neural Networks. In ELM, all the parameters of the networks needn’t tuned iteratively. In view of its fast learning speed, good generalization performance and the absence of local minima, ELM is widely used in all kinds of classification and regression problems. However, its shallow architecture makes it difficult to capture relevant higher-level abstractions in high dimensional data, which makes it difficult to be applied to expanding data in big data Era.Based on the above two algorithms, Huang et al. proposed a multi-layers Deep Extreme Learning Machine(DELM). Multi-layer neural network structure contributes to the extraction of the high-level abstract information; the ELM theory avoids iterative adjustment of the weights, improving the computational efficiency; semi-supervised layer-wise training mechanism helps to solve the problem in many practical problems that training labels are difficult to be obtained. Therefore, DELM can effectively solve these problems at the same time, which is high dimension, big data, the difficulty to get marked samples and form features.According to the Deep Extreme Learning Machine, this paper mainly does the following research:(1) In the robotic grasping recognition, the proposed multi-modal deep ELM method is used as training scheme of multi modal fusion model to deal with RGB and Depth information. The original pixel information in RGB modality and Depth modality can be effectively learned and fused. We perform the experimental validation on the recently developed publicly available Cornell grasping dataset.(2) According to the experimental results, an important merit of such MM-DELM method is that the training efficacy is highly improved. It can be seen that MM-DELM follows a similar convergence property of ELM, and the performances tend to be quite stable in a wide range of the parameters. So it provides the basis for the selection of parameters.(3) In the time series prediction, a new local prediction model of time series is proposed, in which the hybrid Euclidean distance is used as the similarity measurement to complete the reorganization of all of training samples, and the prediction model of recombination samples is derived by DELM.(4) In order to validate the accuracy of forecasting model, the five data sets selected from time-series database(TSDL) serve as the experimental data set, in which the generalization performance of DELM in one-step and multi-step forecasting can be verified.
Keywords/Search Tags:Deep Learning, Deep Extreme Learning Machine, robotic grasping recognition, multi-modal fusion, Time series prediction
PDF Full Text Request
Related items