Skip to main content

Über dieses Buch

This book contains some selected papers from the International Conference on Extreme Learning Machine 2016, which was held in Singapore, December 13-15, 2016. This conference will provide a forum for academics, researchers and engineers to share and exchange R&D experience on both theoretical studies and practical applications of the ELM technique and brain learning. Extreme Learning Machines (ELM) aims to break the barriers between the conventional artificial learning techniques and biological learning mechanism. ELM represents a suite of (machine or possibly biological) learning techniques in which hidden neurons need not be tuned. ELM learning theories show that very effective learning algorithms can be derived based on randomly generated hidden neurons (with almost any nonlinear piecewise activation functions), independent of training data and application environments. Increasingly, evidence from neuroscience suggests that similar principles apply in biological learning systems. ELM theories and algorithms argue that “random hidden neurons” capture an essential aspect of biological learning mechanisms as well as the intuitive sense that the efficiency of biological learning need not rely on computing power of neurons. ELM theories thus hint at possible reasons why the brain is more intelligent and effective than current computers. ELM offers significant advantages over conventional neural network learning algorithms such as fast learning speed, ease of implementation, and minimal need for human intervention. ELM also shows potential as a viable alternative technique for large‐scale computing and artificial intelligence.

This book covers theories, algorithms ad applications of ELM. It gives readers a glance of the most recent advances of ELM.



Earthen Archaeological Site Monitoring Data Analysis Using Kernel-based ELM and Non-uniform Sampling TFR

Known as an ancient civilization, there exists a large amount of earthen archaeological sites in China. Various types of environment monitoring data have been accumulated waiting to be analyzed for the aim of future protection. In this paper, a non-stationary data processing strategy is proposed for the better understanding of such monitoring data. The kernel-based extreme learning machine (ELM) is utilized to preprocess the original data and restore the missing parts. Then a new non-uniform sampling time-frequency representation (TFR) is proposed to analyze the non-stationary characteristic of restored data from a signal processing perspective. The test data is the real environment monitoring data of the burial pit at the Yang Mausoleum of the Han dynasty. The experimental result shows that the proposed scheme can extract different information from the original data.
Yue Qi, Mingzhe Zhu, Xinliang Zhang, Fei Fu

A Multi-valued Neuron ELM with Complex-Valued Inputs for System Identification Using FRA

In the paper a new kind of ELM network is presented, which uses a MVN (multivalued neuron) with complex weights and complex inputs and that seems to be particularly suitable for fault diagnosis and identification in the frequency domain with very simple structures, given their high generalization performance. The presented network has high potentiality with a very low number of neurons. The ELM architecture is then designed with general approach, following the philosophy of this class of neural techniques, and then applied to some specific example.
Francesco Grasso, Antonio Luchetta, Stefano Manetti

Quaternion Extreme Learning Machine

Quaternion signal processing has been an increasing popular research topic for its application in a wide range of fields, and extreme learning machine (ELM) is an emerging training strategy for the generalized single hidden layer feedforward neural networks. However, extreme learning machine could not fully explore its potentials in quaternion signal processing. To this end, this paper propose an quaternion ELM model, which retain the essential characters of the ELM such as the fast learning and universal approximation capability, while enjoying advantages originated from the quaternion algebra. Two simulation examples are provided to support our analysis and to exhibit the enhanced performance of the proposed model over ELM when dealing with the 3D and 4D signal processing problems.
Hui Lv, Huisheng Zhang

Robotic Grasp Stability Analysis Using Extreme Learning Machine

Recently, autonomous grasping of unknown objects is a fundamental requirement for robots performing manipulation tasks in real world environments. It is still considered as a challenging problem no matter how process we have made. It is significant that how the robot to judge the stability of grabbing object. In this paper, we analyze the data through process of grabbing 3 objects whether is successful or failed by constructing Global Alignment kernel with Extreme Learning Machine and Support Vector Machine. For comparative analysis, the Barrett hand’s finger angles and robot joint angles are also recorded. By processing obtained data in different ways, we have comparative results in various modes. Experiments denote the tactile results achieve better performance than the finger angle’s and robot joint angle’s.
Peng Bai, Huaping Liu, Fuchun Sun, Meng Gao

Extreme Learning Machine for Intent Classification of Web Data

Web search engines return a large amount of results for a user search query. Understanding the intent of these search queries can help us to narrow down the search results based on the type of information needed. In the research reported in this paper, we implemented machine learning algorithms to validate the accuracy of the classification of user search query. Broad categories of web query data are used from two different sources. Feature sets extracted solely from the web query are used to train the machine learning classifier. Classification results reveal that the performance of extreme learning machine (ELM) is much better when classifying user query intent than other machine learning classifiers.
Yogesh Parth, Wang Zhaoxia

Reinforcement Extreme Learning Machine for Mobile Robot Navigation

Obstacle avoidance is a very important problem for autonomous navigation of mobile robot. However, most of existing work regards the obstacle detection and control as separate problem. In this paper, we solve the joint learning problem of perception and control using the reinforcement learning framework. To address this problem, we propose an effective Reinforcement Extreme Learning Machine architecture, while maintaining ELM’s advantages of training efficiency. In this structure, the Extreme Learning Machine (ELM) is used as supervised laserscan classier for specified action. And then, the reward function we designed will give a reward to mobile robot according to the results of navigation. The Reinforcement Extreme Learning Machine is then conducted for updating the expected output weights for the final decision.
Hongjie Geng, Huaping Liu, Bowen Wang, Fuchun Sun

Detection of Cellular Spikes and Classification of Cells from Raw Nanoscale Biosensor Data

Nanoscale devices have provided promising endeavors for detecting crucial biomarkers such as DNA, proteins, and human cells at a finer scale. These biomarkers can improve prognosis by detecting dreadful disease such as cancer at an early stage than the current approaches. Analyzing raw data from these nanoscale devices for disease detection is tedious as the raw data suffers from noise. Furthermore, disease detection decisions are made based on manual or semi-automated analysis—which are time-consuming, monotonous and error-prone process. Recent trends show an unprecedented growth in the advancement of nanotechnology for medical diagnosis. These devices generate huge amount of raw data and analyzing raw data in order to classify biomarkers in a fully automated and robust way is a challenge. In this paper, we present an algorithm for identifying cellular spikes, we have adapted extreme learning machines and dynamic time warping for the classification of cancer in raw data collected from nanoscale biosensors, such as solid-state micropores. Our approach can classify cancer cells with an accuracy of 95.6%, and with a precision and recall of 85.7% and 80.0%, respectively.
Muhammad Rizwan, Abdul Hafeez, Ali R. Butt, Samir M. Iqbal

Hot News Click Rate Prediction Based on Extreme Learning Machine and Grey Verhulst Model

Click rate prediction of hot topics contributes to get event tendency, especially for sensitive news. However, click rate prediction is challenge due to inherent features of short-time series such as randomness, uncertainty, volatility and insufficiency of training samples. In this paper, a new hybrid click rate prediction method called Grey Verhulst—Extreme Learning Machine (GVELM) is proposed. Specifically, the raw short-time series data are filled into GV models to acquire stably initial prediction which have incorporated regular pattern of the historic data without noise. Then ELM is employed for prediction refinement for nonlinear space mapping. The experimental results show that the proposed method achieves better prediction accuracy compared with other five state-of-art algorithms.
Xu Jingting, Feng Jun, Sun Xia, Zhang Lei, Liu Xiaoning

Multiple Shadows Layered Cooperative Velocity Updating Particle Swarm Optimization

In real-time high dimensions optimization problem, how to quickly find the optimal solution and give timely response or decisive adjustment is very important. Inspired by space projection behavior, this paper suggests a new PSO variant, Multiple-shadows Layered Cooperative Velocity Updating Particle Swarm Optimization (ML-CVUPSO) that involves visual instructive projections among multiple shadows. According to several different views, the original problem can be divided into different relevant characteristic sub-problems after feature extraction. The ML-CVUPSO provides a flexible and feasible decomposed mechanism to simplify the high dimensions problem into a series of tractable sub-problems. The proposed variant is examined on several widely used benchmark functions, and the experimental results show that the proposed ML-CVUPSO algorithm improves the existing performance of other algorithms when dealing with the high dimension and multimodal problems.
Hongbo Wang, Kezhen Wang, Xuyan Tu

Short Term Prediction of Continuous Time Series Based on Extreme Learning Machine

Extreme Learning Machine (ELM) is a popular tool of machine learning, which has been used in many fields. Time series prediction is usually a complex problem without related parameters or features. In this paper, a prediction method for continuous time series based on the theory of extreme learning machines is proposed, which focus on short term prediction of continuous time series. Firstly, the ST-ELMpredicting model is constructed. Then the ways of training and predicting is analyzed. ST-ELM uses time series and predicted value to adjust itself. Mackey-Glass and Lorenz time series have been used as example for demonstration. It is showed this method can predict continuous time series timely and accurately without related parameters or features of time series.
Hongbo Wang, Peng Song, Chengyao Wang, Xuyan Tu

Learning Flow Characteristics Distributions with ELM for Distributed Denial of Service Detection and Mitigation

We present a methodology for modeling the distributions of network flow statistics for the specific purpose of network anomaly detection, in the form of Distributed Denial of Service attacks. The proposed methodology offers to model (using Extreme Learning Machines, ELM), at the IP subnetwork level (or all the way down to the single IP level, if computations allow), the usual distributions of certain network flow characteristics (or statistics), and then to use a One-Class classifier in the detection of abnormal joint flow statistics. The methodology makes use of the original ELM for its good performance to computational time ratio, but also because of the needs in this methodology to have simple update rules for making the model evolve in time, as new traffic and hosts come in.
Aapo Kalliola, Yoan Miche, Ian Oliver, Silke Holtmanns, Buse Atli, Amaury Lendasse, Kaj-Mikael Bjork, Anton Akusok, Tuomas Aura

Discovering Emergence and Bidding Behaviour in Competitive Electricity Market Using Agent-Based Simulation

The aim of this paper is to explore the implication of multi agent interaction, learning and competing in a repetitive trading environment. Using the complex systems paradigm, the study attempts to observe the behavior of the agents and the emergence phenomena resulting from the multi agent interaction. Using Q-learning, generator agents can rapidly learn the market mechanism and auction rules as they seek to maximize their revenue by modifying their bidding strategies. In this paper, we experiment with different pricing rule to observe its impact on agents’ behavior. The paper also describes the types of agents in each domain, together with the properties, relationships, processes and events associated with the agents. Emergence from this study includes collusion and capacity withholding to inflate price. The emergence is evidence that we can gain new knowledge from the Sciences of the Artificial.
Ly-Fie Sugianto, Zhigang Liao

Multi-kernel Transfer Extreme Learning Classification

In this paper, a novel transfer extreme learning machine (TELM) algorithm based on multi-kernel (MK) framework has been proposed for classification. In this case, the problem is transformed into a semi-supervised learning problem, which allows multi-kernel extreme learning machine (MK-TELM) classifiers to be trained for the data categorization. Compared with many popular algorithms, the proposed method, named as MK-TELM, shows its satisfactorily experimental results on the variety of data sets, which highlights the robustness and effectiveness for classification applications.
Xiaodong Li, Weijie Mao, Wei Jiang, Ye Yao

Chinese Text Sentiment Classification Based on Extreme Learning Machine

With the rapid growth of the Web text data, mining and analyzing these text data, especially the online review data posted by the users, can greatly help better understand the usersconsuming habits and public opinions, it also plays an important role in decision-making for the enterprises and the government. But in the process of vectoring text, many current Chinese text sentiment classifications treat words as atomic units, there is no notion of similarity between words. In order to solve this problem, this paper imports word embedding to capturing both the semantic and syntactic information of words from a large unlabeled corpus. In the section of experiment, we toke the noun, verb, and adjectives as candidate set, used \(\chi ^{2}\) statistic to reduce the number of dimensions. We mainly compared one-hot representation and word embedding as the expression of word to certain tasks, we also proposed the pooling method with word embedding to standardizing the vector, the ELM with kernels was adopted to analyze the text emotion tendentiousness. Finally the paper summarizes the current status, remaining challenges, and future directions in the field of sentiment classification.
Fangye Lin, Yuanlong Yu

Incremental ELMVIS for Unsupervised Learning

An incremental version of the ELMVIS+ method is proposed in this paper. It iteratively selects a few best fitting data samples from a large pool, and adds them to the model. The method keeps high speed of ELMVIS+ while allowing for much larger possible sample pools due to lower memory requirements. The extension is useful for reaching a better local optimum with greedy optimization of ELMVIS, and the data structure can be specified in semi-supervised optimization. The major new application of incremental ELMVIS is not to visualization, but to a general dataset processing. The method is capable of learning dependencies from non-organized unsupervised data—either reconstructing a shuffled dataset, or learning dependencies in complex high-dimensional space. The results are interesting and promising, although there is space for improvements.
Anton Akusok, Emil Eirola, Yoan Miche, Ian Oliver, Kaj-Mikael Björk, Andrey Gritsenko, Stephen Baek, Amaury Lendasse

Predicting Huntington’s Disease: Extreme Learning Machine with Missing Values

Problems with incomplete data and missing values are common and important in real-world machine learning scenarios, yet often underrepresented in the research field. Particularly data related to healthcare tends to feature missing values which must be handled properly, and ignoring any incomplete samples is not an acceptable solution. The Extreme Learning Machine has demonstrated excellent performance in a variety of machine learning tasks, including situations with missing values. In this paper, we present an application to predict the onset of Huntington’s disease several years in advance based on data from MRI brain scans. Experimental results show that such prediction is indeed realistic with reasonable accuracy, provided the missing values are handled with care. In particular, Multiple Imputation ELM achieves exceptional prediction accuracy.
Emil Eirola, Anton Akusok, Kaj-Mikael Björk, Hans Johnson, Amaury Lendasse

Deep-Learned and Hand-Crafted Features Fusion Network for Pedestrian Gender Recognition

In this paper, we propose an effective deep-learned and hand-crafted features fusion network (DHFFN) for pedestrian gender recognition. In the proposed DHFFN, the deep-learned and hand-crafted (i.e., HOG) features are extracted for the input image, followed by the feature fusion process that is to combine these two features together for fully exploring the merits from both deep-learned and HOG features. Extensive experiments on multiple public datasets have demonstrated that the proposed DHFFN method is superior to the state-of-the-art pedestrian gender recognition methods.
Lei Cai, Jianqing Zhu, Huanqiang Zeng, Jing Chen, Canhui Cai

Facial Landmark Detection via ELM Feature Selection and Improved SDM

Model initialization and feature extraction are crucial in supervised landmark detection. Mismatching caused by detector error and discrepant initialization is very common in these existing methods. To solve this problem, we have proposed a new method based on ELM feature selection and Improved Supervised Descent Method (ELMFS-iSDM), which also includes an automatic initialization model, for the robust facial landmark localization. In our new method, firstly, a fast detection will be processed to locate the eyes and mouth, and the initialization model will adapt to the real location according to fast facial points detection. Secondly, ELM based feature selection is adopted on our Improved Supervised Descent Method model to achieve a better performance. For each task, multiple features will be jointly learned by ELM feature selection and their weights will be calculated during training process. Experiments on four benchmark databases show that our method achieves state-of-the-art performance.
Peng Bian, Yi Jin, Jiuwen Cao

Online Sequential Extreme Learning Machine with Under-Sampling and Over-Sampling for Imbalanced Big Data Classification

In this paper, a novel method called online sequential extreme learning machine with under-sampling and over-sampling (OSELM-UO) for imbalanced Big data classification is proposed which combines the structures of under-sampling and over-sampling and applies online sequential extreme learning machine as its base model. The novel structure enables OSELM-UO performs well on both minority and majority classes and simultaneously overcomes the issues of information loss and overfitting. Moreover, when the dataset keeps growing, OSELM-UO can be applied without retraining all previous data. Experiments have been conducted for OSELM-UO and several imbalance learning methods over real-world datasets respectively under high imbalance ratio (IR) and large amount of samples and features. Through the analysis of the experimental results, OSELM-UO is shown to give the best results in various aspects.
Jie Du, Chi-Man Vong, Yajie Chang, Yang Jiao

An Automatic Identification System (AIS) Database for Maritime Trajectory Prediction and Data Mining

In recent years, maritime safety and efficiency become very important across the world. Automatic Identification System (AIS) tracks vessel movement by onboard transceiver and terrestrial and/or satellite base stations. The data collected by AIS contain broadcast kinematic information and static information. Both of them are useful for maritime anomaly detection and vessel route prediction which are key techniques in maritime intelligence. This paper is devoted to construct a standard AIS database for maritime trajectory learning, prediction and data mining. A path prediction method based on Extreme Learning Machine (ELM) is tested on this AIS database and the testing results show this database can be used as a standardized training resource for different trajectory prediction algorithms and other AIS data based mining applications.
Shangbo Mao, Enmei Tu, Guanghao Zhang, Lily Rachmawati, Eshan Rajabally, Guang-Bin Huang

Back Propagation Convex Extreme Learning Machine

Recently, extreme learning machine has greatly improved in training speed and learning effectiveness of feedforward neural network which includes one hidden layer. However, the random initialization of ELM model parameters can bring randomness and affect generalization ability. The paper proposed back propagation convex extreme learning machine (BP-CELM), in which the hidden layer parameters \( {\mathbf{(a}},\,{\mathbf{b}}) \) can be calculated by formulas. The convergence of BP-CELM is proved in the paper. Simulation results show that BP-CELM has higher training speed and better generalization performance than other randomized neural network algorithms.
Weidong Zou, Fenxi Yao, Baihai Zhang, Zixiao Guan

Data Fusion Using OPELM for Low-Cost Sensors in AUV

With mobility, security, intelligence and other advantages, autonomous underwater vehicle (AUV) becomes an indispensable instrument in the complex underwater environment. Owing to the independence of external signal (such as GPS) which is restricted or invalid in the water, inertial navigation system (INS) has become the most suitable navigation and positioning system for Underwater Vehicles. However, as the excessive reliance of sensor data, the precision of INS can be affected by sensor data especially heading angle data from low-cost sensor such as attitude and heading reference system (AHRS) and digital compass. Therefore, how to fuse low-cost sensor information to get more accurate data becomes the key to improve navigation accuracy. Based on the original Extreme Learning Machine (ELM) algorithm, the Optimally Pruned Extreme Learning Machine (OPELM) algorithm is presented as a more robust and general methodology in 2010, which make it possible to realize data fusion by using a more reliable network. In this paper, we proposed a method of data fusion which using Optimally-Pruned Extreme Learning Machine (OPELM) to improve the accuracy of heading angle from AHRS and digital compass. Our method has already been demonstrated by a range of real datasets, and it outperforms current available Kalman Filtering algorithms in efficiency.
Jia Guo, Bo He, Pengfei Lv, Tianhong Yan, Amaury Lendasse
Weitere Informationen

Premium Partner