Skip to main content
main-content

Über dieses Buch

The two volume set LNCS 6443 and LNCS 6444 constitutes the proceedings of the 17th International Conference on Neural Information Processing, ICONIP 2010, held in Sydney, Australia, in November 2010. The 146 regular session papers presented were carefully reviewed and selected from 470 submissions. The papers of part I are organized in topical sections on neurodynamics, computational neuroscience and cognitive science, data and text processing, adaptive algorithms, bio-inspired algorithms, and hierarchical methods. The second volume is structured in topical sections on brain computer interface, kernel methods, computational advance in bioinformatics, self-organizing maps and their applications, machine learning applications to image analysis, and applications.

Inhaltsverzeichnis

Frontmatter

Brain Computer Interface

Utilizing Fuzzy-SVM and a Subject Database to Reduce the Calibration Time of P300-Based BCI

Current Brain-Computer Interfaces (BCI) suffer the requirement of a subject-specific calibration process due to variations in EEG responses across different subjects. Additionally, the duration of the calibration process should be long enough to sufficiently sample high dimensional feature spaces. In this study, we proposed a method based on Fuzzy Support Vector Machines (Fuzzy-SVM) to address both issues for P300-based BCI. We conducted P300 speller experiments on 18 subjects, and formed a subject-database using a leave-one-out approach. By computing weight values for the data samples obtained from each subject, and by incorporating those values into the Fuzzy-SVM algorithm, we achieved to obtain an average accuracy of 80% with only 4 training letters. Conventional subject-specific calibration approach, on the other hand, needed 12 training letters to provide the same performance.

Sercan Taha Ahi, Natsue Yoshimura, Hiroyuki Kambara, Yasuharu Koike

Importance Weighted Extreme Energy Ratio for EEG Classification

Spatial filtering is important for EEG signal processing since raw scalp EEG potentials have a poor spatial resolution due to the volume conduction effect. Extreme energy ratio (EER) is a recently proposed feature extractor which exhibits good performance. However, the performance of EER will be degraded by some factors such as outliers and the time-variances between the training and test sessions. Unfortunately, these limitations are common in the practical brain-computer interface (BCI) applications. This paper proposes a new feature extraction method called importance-weighted EER (IWEER) by defining two kinds of weight termed

intra-trial importance

and

inter-trial importance

. These weights are defined with the density ratio theory and assigned to the data points and trials respectively to improve the estimation of covariance matrices. The spatial filters learned by the IWEER are both robust to the outliers and adaptive to the test samples. Compared to the previous EER, experimental results on nine subjects demonstrate the better classification ability of the IWEER method.

Wenting Tu, Shiliang Sun

Toward Automated Electrode Selection in the Electronic Depth Control Strategy for Multi-unit Recordings

Multi-electrode arrays contain an increasing number of electrodes. The manual selection of good quality signals among hundreds of electrodes becomes impracticable for experimental neuroscientists. This increases the need for an automated selection of electrodes containing good quality signals. To motivate the automated selection, three experimenters were asked to assign quality scores, taking one of four possible values, to recordings containing action potentials obtained from the monkey primary somatosensory cortex and the superior parietal lobule. Krippendorff’s alpha-reliability was then used to verify whether the scores, given by different experimenters, were in agreement. A Gaussian process classifier was used to automate the prediction of the signal quality using the scores of the different experimenters. Prediction accuracies of the Gaussian process classifier are about 80% when the quality scores of different experimenters are combined, through a median vote, to train the Gaussian process classifier. It was found that predictions based also on firing rate features are in closer agreement with the experimenters’ assignments than those based on the signal-to-noise ratio alone.

Gert Van Dijck, Ahmad Jezzini, Stanislav Herwik, Sebastian Kisban, Karsten Seidl, Oliver Paul, Patrick Ruther, Francesca Ugolotti Serventi, Leonardo Fogassi, Marc M. Van Hulle, Maria Alessandra Umiltà

Tensor Based Simultaneous Feature Extraction and Sample Weighting for EEG Classification

In this paper we propose a Multi-linear Principal Component Analysis (MPCA) which is a new feature extraction and sample weighting method for classification of EEG signals using tensor decomposition. The method has been successfully applied to Motor-Imagery Brain Computer Interface (MI-BCI) paradigm. The performance of the proposed approach has been compared with standard Common Spatial Pattern (CSP) as well with a combination of PCA and CSP methods. We have achieved an average accuracy improvement of two classes classification in a range from 4 to 14 percents.

Yoshikazu Washizawa, Hiroshi Higashi, Tomasz Rutkowski, Toshihisa Tanaka, Andrzej Cichocki

A Tongue-Machine Interface: Detection of Tongue Positions by Glossokinetic Potentials

Artifacts are electrical activities that are detected along the scalp by an electroencephalography (EEG) but that originate from non-cerebral origin, which often need to be eliminated before further processing of EEG signals. Glossokinetic potentials are artifacts related to tongue movements. In this paper we use these glossokinetic artifacts (instead of eliminating them) to automatically detect and classify tongue positions, which is important in developing a tongue-machine interface. We observe that with a specific selection of a few electrode positions, glossokinetic potentials show contralateral patterns, so that the magnitude of potentials is linearly proportional to the tongue positions flicking at the left to the right inside of cheek. We design a simple linear model based on principal component analysis (PCA) to translate glossokinetic potentials into tongue positions. Experiments on cursor control confirm the validity of our method for tongue position detection using glossokinetic potentials.

Yunjun Nam, Qibin Zhao, Andrzej Cichocki, Seungjin Choi

Practical Surface EMG Pattern Classification by Using a Selective Desensitization Neural Network

Real-time pattern classification of electromyogram (EMG) signals is significant and useful for developing prosthetic limbs. However, the existing approaches are not practical enough because of several limitations in their usage, such as the large amount of data required to train the classifier. Here, we introduce a method employing a selective desensitization neural network (SDNN) to solve this problem. The proposed approach can train the EMG classifier to perform various hand movements by using a few data samples, which provides a highly practical method for real-time EMG pattern classification.

Hiroshi Kawata, Fumihide Tanaka, Atsuo Suemitsu, Masahiko Morita

Reliability-Based Automatic Repeat reQuest with Error Potential-Based Error Correction for Improving P300 Speller Performance

The P300 speller allows users to select letters just by thoughts. However, due to the low signal-to-noise ratio of the P300 response, signal averaging is often performed, which improves the spelling accuracy but degrades the spelling speed. The authors have proposed

reliability-based automatic repeat request

(RB-ARQ) to ease this problem. RB-ARQ could be enhanced when it is combined with the error correction based on the error-related potentials. This paper presents how to combine both methods and how to optimize parameters to maximize the performance of the P300 speller. The result shows that the performance was improved by 40 percent on average.

Hiromu Takahashi, Tomohiro Yoshikawa, Takeshi Furuhashi

An Augmented-Reality Based Brain-Computer Interface for Robot Control

In this study we demonstrate how the combination of Augmented-Reality (AR) techniques and an asynchronous P300-based Brain-Computer Interface (BCI) can be used to control a robotic actuator by thought. We show results of an experimental study which required the users to move several objects placed on a desk by concentrating on a specific object. Competitive communication speed of up to 5.9 correct symbols per minute at a 100% accuracy level could be achieved for one subject using an asynchronous paradigm which enables the user to start communicating a command at an arbitrary time and thus mitigating the drawbacks of the standard cue based P300 protcols.

Alexander Lenhardt, Helge Ritter

Brain Computer Interfaces: A Recurrent Neural Network Approach

This paper explores the use of recurrent neural networks in the field of Brain Computer Interfaces(BCI). In particular it looks at a recurrent neural network, an echostate network and a CasPer neural network and attempts to use them to classify data from BCI competition III’s dataset IVa. In addition it proposes a new method, EchoCasPer, which uses the CasPer training scheme in a recurrent neural network. The results showed that temporal information existed within the BCI data to be made use of, but further pre-processing and parameter exploration was needed to reach competitive classification rates.

Gareth Oliver, Tom Gedeon

Research on Relationship between Saccade-Related EEG Signals and Selection of Electrode Position by Independent Component Analysis

Our goal is to develop a novel BCI based on an eye movements system employing EEG signals on-line. Most of the analysis on EEG signals has been performed using ensemble averaging approaches. However, in signal processing methods for BCI, raw EEG signals are analyzed.

In order to process raw EEG signals, we used independent component analysis(ICA).

Previous paper presented extraction rate of saccade-related EEG signals by five ICA algorithms and eight window size.

However, three ICA algorithms, the FastICA, the NG-FICA and the JADE algorithms, are based on 4th order statistic and AMUSE algorithm has an improved algorithm named the SOBI. Therefore, we must re-select ICA algorithms.

In this paper, Firstly, we add new algorithms; the SOBI and the MILCA. Using the Fast ICA, the JADE, the AMUSE, the SOBI, and the MILCA. The SOBI is an improved algorithm based on the AMUSE and uses at least two covariance matrices at different time steps. The MILCA use the independency based on mutual information. We extract saccade-related EEG signals and check extracting rates.

Secondly, we check relationship between window sizes of EEG signals to be analyzed and extracting rates.

Thirdly, we researched on relationship between Saccade-related EEG signals and selection of electrode position by ICA. In order to develop the BCI, it is important to use a few electrode. In previous studies, we analyzed EEG signals using by 19 electrodes. In this study, we checked various combination of electrode.

Arao Funase, Motoaki Mouri, Andrzej Cichocki, Ichi Takumi

Kernel Methods

Application of SVM-Based Filter Using LMS Learning Algorithm for Image Denoising

In this paper, a novel adaptive filter based on support vector machines (SVMs) that preserves image details and effectively suppresses impulsive noise is proposed. The filter employs an SVM impulse detector to judge whether an input pixel is noisy. If a noisy pixel is detected, a median filter is triggered to replace it. Otherwise, it stays unchanged. To improve the quality of the restored image, an adaptive LUM filter based on scalar quantization (SQ) is activated. The optimal weights of the adaptive LUM filter are obtained using the least mean square (LMS) learning algorithm. Experimental results demonstrate that the proposed scheme outperforms other decision-based median filters in terms of noise suppression and detail preservation.

Tzu-Chao Lin, Chien-Ting Yeh, Mu-Kun Liu

Tuning N-gram String Kernel SVMs via Meta Learning

Even though Support Vector Machines (SVMs) are capable of identifying patterns in high dimensional kernel spaces, their performance is determined by two main factors: SVM cost parameter and kernel parameters. This paper identifies a mechanism to extract meta features from string datasets, and derives a

n-gram

string kernel SVM optimization method. In the method, a

meta model

is trained over computed string meta-features for each dataset from a string dataset pool, learning algorithm parameters, and accuracy information to predict the optimal parameter combination for a given string classification task. In the experiments, the

n-gram

SVM were optimized using the proposed algorithm over four string datasets: spam, Reuters-21578, Network Application Detection and e-News Categorization. The experiment results revealed that the proposed algorithm was able to produce parameter combinations which yield good string classification accuracies for

n-gram

SVM on all string datasets.

Nuwan Gunasekara, Shaoning Pang, Nikola Kasabov

Bilinear Formulated Multiple Kernel Learning for Multi-class Classification Problem

In this paper, we propose a method of multiple kernel learning (MKL) to inherently deal with multi-class classification problems. The performances of kernel-based classification methods depend on the employed kernel functions, and it is difficult to predefine the optimal kernel. In the framework of MKL, multiple types of kernel functions are linearly integrated with optimizing the weights for the kernels. However, the multi-class problems are rarely incorporated in the formulation and the optimization is time-consuming. We formulate the multi-class MKL in a bilinear form and propose a scheme for computationally efficient optimization. The scheme makes the method favorably applicable to large-scaled samples in the real-world problems. In the experiments on multi-class classification using several datasets, the proposed method exhibits the favorable performance and low computation time compared to the previous methods.

Takumi Kobayashi, Nobuyuki Otsu

Feature Extraction Using Support Vector Machines

We discuss feature extraction by support vector machines (SVMs). Because the coefficient vector of the hyperplane is orthogonal to the hyperplane, the vector works as a projection vector. To obtain more projection vectors that are orthogonal to the already obtained projection vectors, we train the SVM in the complementary space of the space spanned by the already obtained projection vectors. This is done by modifying the kernel function. We demonstrate the validity of this method using two-class benchmark data sets.

Yasuyuki Tajiri, Ryosuke Yabuwaki, Takuya Kitamura, Shigeo Abe

Class Information Adapted Kernel for Support Vector Machine

This article presents a support vector machine (SVM) learning approach that adapts class information within the kernel computation. Experiments on fifteen publicly available datasets are conducted and the impact of proposed approach for varied settings are observed. It is noted that the new approach generally improves minority class prediction, depicting it as a well-suited scheme for imbalanced data. However, a SVM based customization is also developed that significantly improves prediction performance in terms of different measures. Overall, the proposed method holds promise with potential for future extensions.

Tasadduq Imam, Kevin Tickle

Gaze Pattern and Reading Comprehension

Does the way a person read influence the way they understand information or is it the other way around? In regard to reading of English text, just how much we can learn from a person’s gaze pattern? It is known that while reading, we inadvertently form rational connections between pieces of information we pick up from the text. That reflects in certain disruptions in the norms of reading paradigm and that gives us clues to our interest level in reading activities.

In this paper, we validate the above statement and then propose a novel method of detecting the level of engagement in reading based on a person’s gaze-pattern. We organised some experiments in reading tasks of over thirty participants and the experimental outputs are classified with Artificial Neural Networks with an approximately 80 percent accuracy. The design of this approach is simple and computationally feasible enough to be applied in a real-life system.

“Your eyes are the windows to your soul”

Tan Vo, B. Sumudu U. Mendis, Tom Gedeon

A Theoretical Framework for Multi-sphere Support Vector Data Description

In support vector data description (SVDD) a spherically shaped boundary around a normal data set is used to separate this set from abnormal data. The volume of this data description is minimized to reduce the chance of accepting abnormal data. However the SVDD does not guarantee that the single spherically shaped boundary can best describe the normal data set if there are some distinctive data distributions in this set. A better description is the use of multiple spheres, however there is currently no investigation available. In this paper, we propose a theoretical framework to multi-sphere SVDD in which an optimisation problem and an iterative algorithm are proposed to determine model parameters for multi-sphere SVDD to provide a better data description to the normal data set. We prove that the classification error will be reduced after each iteration in this learning process. Experimental results on 28 well-known data sets show that the proposed multi-sphere SVDD provides lower classification error rate comparing with the standard single-sphere SVDD.

Trung Le, Dat Tran, Wanli Ma, Dharmendra Sharma

Fast Implementation of String-Kernel-Based Support Vector Classifiers by GPU Computing

Text categorization is widely used in applications such as spam filtering, identification of document genre, authorship attribution, and automated essay grading. The rapid growth in the amount of text data gives rise to the urgent need for fast text classification algorithms. In this paper, we propose a GPU based SVM solver for large scale text datasets. Using Platt’s Sequential Minimal Optimization algorithm, we achieve a speedup of 5–40 times over LibSVM running on a high-end traditional processor. Prediction time based on the paralleled string kernel computing scheme shows 5–90 times faster performance than the CPU based implementation.

Yongquan Shi, Tao Ban, Shanqing Guo, Qiuliang Xu, Youki Kadobayashi

Model Generation and Classification

Classification of Imbalanced Data by Combining the Complementary Neural Network and SMOTE Algorithm

In classification, when the distribution of the training data among classes is uneven, the learning algorithm is generally dominated by the feature of the majority classes. The features in the minority classes are normally difficult to be fully recognized. In this paper, a method is proposed to enhance the classification accuracy for the minority classes. The proposed method combines Synthetic Minority Over-sampling Technique (SMOTE) and Complementary Neural Network (CMTNN) to handle the problem of classifying imbalanced data. In order to demonstrate that the proposed technique can assist classification of imbalanced data, several classification algorithms have been used. They are Artificial Neural Network (ANN), k-Nearest Neighbor (k-NN) and Support Vector Machine (SVM). The benchmark data sets with various ratios between the minority class and the majority class are obtained from the University of California Irvine (UCI) machine learning repository. The results show that the proposed combination techniques can improve the performance for the class imbalance problem.

Piyasak Jeatrakul, Kok Wai Wong, Chun Che Fung

Generalization Error of Faulty MLPs with Weight Decay Regularizer

Weight decay is a simple regularization method to improve the generalization ability of multilayered perceptrons (MLPs). Besides, the weight decay method can also improve the fault tolerance of MLPs. However, most existing generalization error results of using the weight decay method focus on fault-free MLPs only. For faulty MLPs, using a test set to study the generalization ability is not practice because there are huge number of possible faulty networks for a trained network. This paper develops a prediction error formula for predicting the performance of faulty MLPs. Our prediction error results allows us to select an appropriate model for MLPs under open node fault situation.

Chi Sing Leung, John Sum, Shue Kwan Mak

The Effect of Bottlenecks on Generalisation in Backpropagation Neural Networks

Many modifications have been proposed to improve back-propagation’s convergence time and generalisation capabilities. Typical techniques involve pruning of hidden neurons, adding noise to hidden neurons which do not learn, and reducing dataset size. In this paper, we wanted to compare these modifications’ performance in many situations, perhaps for which they were not designed. Seven famous UCI datasets were used. These datasets are different in dimension, size and number of outliers. After experiments, we find some modifications have excellent effect of decreasing network’s convergence time and improving generalisation capability while some modifications perform much the same as unmodified back-propagation. We also seek to find a combine of modifications which outperforms any single selected modification.

Xu Zang

Lagrange Programming Neural Networks for Compressive Sampling

Compressive sampling is a sampling technique for sparse signals. The advantage of compressive sampling is that signals are compactly represented by a few number of measured values. This paper adopts an analog neural network technique, Lagrange programming neural networks (LPNNs), to recover data in compressive sampling. We propose the LPNN dynamics to handle three sceneries, including the standard recovery of sparse signal, the recovery of non-sparse signal, and the noisy measurement values, in compressive sampling. Simulation examples demonstrate that our approach effectively recovers the signals from the measured values for both noise free and noisy environment.

Ping-Man Lam, Chi Sing Leung, John Sum, A. G. Constantinides

Input and Output Mapping Sensitive Auto-Associative Multilayer Perceptron for Computer Interface System Based on Image Processing of Laser Pointer Spot

In this paper, we propose a new auto-associative multilayer perceptron (AAMLP) that properly enhances the sensitivity of input and output (I/O) mapping by applying a high pass filter characteristic to the conventional error back propagation learning algorithm, through which small variation of input feature is successfully indicated. The proposed model aims to sensitively discriminate a data of one cluster with small different characteristics against another different cluster’s data. Objective function for the proposed neural network is modified by additionally considering an input and output sensitivity, in which the weight update rules are induced in the manner of minimizing the objective function by a gradient descent method. The proposed model is applied for a real application system to localize laser spots in a beam projected image, which can be utilized as a new computer interface system for dynamic interaction with audiences in presentation or meeting environment. Complexity of laser spot localization is very wide, therefore it is very simple in some cases, but it becomes very tough when the laser spot area has very slightly different characteristic compared with the corresponding area in a beam projected image. The proposed neural network model shows better performance by increasing the input-output mapping sensitivity than the conventional AAMLP.

Chanwoong Jung, Sang-Woo Ban, Sungmoon Jeong, Minho Lee

Improving Recurrent Neural Network Performance Using Transfer Entropy

Reservoir computing approaches have been successfully applied to a variety of tasks. An inherent problem of these approaches, is, however, their variation in performance due to fixed random initialisation of the reservoir. Self-organised approaches like intrinsic plasticity have been applied to improve reservoir quality, but do not take the task of the system into account. We present an approach to improve the hidden layer of recurrent neural networks, guided by the learning goal of the system. Our reservoir adaptation optimises the information transfer at each individual unit, dependent on properties of the information transfer between input and output of the system. Using synthetic data, we show that this reservoir adaptation improves the performance of offline echo state learning and Recursive Least Squares Online Learning.

Oliver Obst, Joschka Boedecker, Minoru Asada

Design of Artificial Neural Networks Using Differential Evolution Algorithm

The design of an Artificial Neural Network (ANN) is a difficult task for it depends on the human experience. Moreover it needs a process of testing and error to select which kind of a transfer function and which algorithm should be used to adjusting the synaptic weights in order to solve a specific problem. In the last years, bio-inspired algorithms have shown their power in different non-linear optimization problems. Due to their efficiency and adaptability, in this paper we explore a new methodology to automatically design an ANN based on the Differential Evolution (DE) algorithm. The proposed method is capable to find the topology, the synaptic weights and the transfer functions to solve a given pattern classification problems.

Beatriz A. Garro, Humberto Sossa, Roberto A. Vázquez

ESNs with One Dimensional Topography

In this paper the standard Echo State approach is combined with a topography, i.e. it is assigned with a position which implies certain constraints of the mutual connectivity between these neurons. The overall design of the network allows certain neurons to process new information earlier than others. As a consequence the connectivity of the trained output layer can be analyzed; conclusions can be drawn regarding which reservoir depth is sufficient to process the given task. In particular we look at connection strengths of different locations of the reservoir as a function of the test error which can be influenced by using ridge regression.

N. Michael Mayer, Matthew Browne, Horng Jason Wu

Computational Advance in Bioinformatics

iGAPK: Improved GAPK Algorithm for Regulatory DNA Motif Discovery

Computational DNA motif discovery is one of the major research areas in bioinformatics, which helps to understanding the mechanism of gene regulation. Recently, we have developed a GA-based motif discovery algorithm, named as GAPK, which addresses the use of some identified transcription factor binding sites extracted from orthologs for algorithm development. With our GAPK framework, technical improvements on background filtering, evolutionary computation or model refinement will contribute to achieving better performances. This paper aims to improve the GAPK framework by introducing a new fitness function, termed as relative model mismatch score (RMMS), which characterizes the conservation and rareness properties of DNA motifs simultaneously. Other technical contributions include a rule-based system for filtering background data and a “most one-in-out” (MOIO) strategy for motif model refinement. Comparative studies are carried out using eight benchmark datasets with original GAPK and two GA-based motif discovery algorithms, GAME and GALF-P. The results show that our improved GAPK method favorably outperforms others on the testing datasets.

Dianhui Wang, Xi Li

A Computer-Aided Detection System for Automatic Mammography Mass Identification

Automatic detection and identification of mammography masses is important for breast cancer diagnosis. However, it is challenging to differentiate masses from normal breast regions because they usually have low contrast and a poor boundary. In this study, we present a Computer-Aided Detection (CAD) system for automatic breast mass identification. A four-stage region-based procedure is adopted for processing the mammogram images, i.e. localization, segmentation, feature extraction, and feature selection and classification. The proposed CAD system is evaluated using selected mammogram images from the Mammographic Image Analysis Society (MIAS) database. The experimental results demonstrate that the proposed CAD system is able to identify mammography masses in an automated manner, and is useful as a decision support system for breast cancer diagnosis.

Hussein Samma, Chee Peng Lim, Ali Samma

Exploring Features and Classifiers to Classify MicroRNA Expression Profiles of Human Cancer

Recently, some non-coding small RNAs, known as microRNAs (miRNA), have drawn a lot of attention to identify their role in gene regulation and various biological processes. The miRNA profiles are surprisingly informative, reflecting the malignancy state of the tissues. In this paper, we attempt to explore extensive features and classifiers through a comparative study of the most promising feature selection methods and machine learning classifiers. Here we use the expression profile of 217 miRNAs from 186 samples, including multiple human cancers. Pearson’s and Spearman’s correlation coefficients, Euclidean distance, cosine coefficient, information gain, mutual information and signal to noise ratio have been used for feature selection. Backpropagation neural network, support vector machine, and

k

-nearest neighbor have been used for classification. Experimental results indicate that

k

-nearest neighbor with cosine coefficient produces the best result, 95.0% of recognition rate on the test data.

Kyung-Joong Kim, Sung-Bae Cho

SOMIX: Motifs Discovery in Gene Regulatory Sequences Using Self-Organizing Maps

We present a clustering algorithm called

S

elf-

o

rganizing

M

ap Neural Network with m

ix

ed signals discrimination (SOMIX), to discover binding sites in a set of regulatory regions. Our framework integrates a novel intra-node soft competitive procedure in each node model to achieve maximum discrimination of motif from background signals. The intra-node competition is based on an adaptive weighting technique on two different signal models: position specific scoring matrix and markov chain. Simulations on real and artificial datasets showed that, SOMIX could achieve significant performance improvement in terms of sensitivity and specificity over SOMBRERO, which is a well-known SOM based motif discovery tool. SOMIX has also been found promising comparing against other popular motif discovery tools.

Nung Kion Lee, Dianhui Wang

Microarray-Based Disease Classification Using Pathway Activities with Negatively Correlated Feature Sets

The vast amount of data on gene expression that is now available through high-throughput measurement of mRNA abundance has provided a new basis for disease diagnosis. Microarray-based classification of disease states is based on gene expression profiles of patients. A large number of methods have been proposed to identify diagnostic markers that can accurately discriminate between different classes of a disease. Using only a subset of genes in the pathway, such as so-called condition-responsive genes (CORGs), may not fully represent the two classification boundaries for Case and Control classes. Negatively correlated feature sets (NCFS) for identifying CORGs and inferring pathway activities are proposed in this study. Our two proposed methods (NCFS-i and NCFS-c) achieve higher accuracy in disease classification and can identify more phenotype-correlated genes in each pathway when comparing to several existing pathway activity inference methods.

Pitak Sootanan, Santitham Prom-on, Asawin Meechai, Jonathan H. Chan

Data Mining for Cybersecurity

A Malware Detection Algorithm Based on Multi-view Fusion

One of the major problems concerning information assurance is malicious code. In order to detect them, many existing run-time intrusion or malware detection techniques utilize information available in Application Programming Interface (API) call sequences to discriminate between benign and malicious processes. Although some great progresses have been made, the new research results of ensemble learning make it possible to design better malware detection algorithm. This paper present a novel approach of detecting malwares using API call sequences. Basing on the fact that the API call sequences of a software show local property when doing network, file IO and other operations, we first divide the API call sequences of a malware into seven subsequences, and then use each subsequence to build a classification model. After these building models are used to classify software, their outputs are combined by using BKS and the final fusion results will be used to label whether a software is malicious or not. Experiments show that our algorithm can detect known malware effectively.

Shanqing Guo, Qixia Yuan, Fengbo Lin, Fengyu Wang, Tao Ban

A Fast Kernel on Hierarchial Tree Structures and Its Application to Windows Application Behavior Analysis

System calls have been proved to be important evidence for analyzing the behavior of running applications. However, application behavior analyzers which investigate the majority of system calls usually suffer from severe system performance deterioration or frequent system crashes. In the presented study, a light weighted analyzer is approached by two avenues. On the one hand, the computation load to monitor the system calls are considerably reduced by limiting the target functions to two specific groups: file accesses and Windows Registry accesses. On the other hand, analytical accuracy is achieved by deep inspection into the string parameters of the function calls, where the proximity of the programs are evaluated by the newly proposed kernel functions. The efficacy of the proposed approach is evaluated on real world datasets with promising results reported.

Tao Ban, Ruo Ando, Youki Kadobayashi

Evolution of Information Retrieval in Cloud Computing by Redesigning Data Management Architecture from a Scalable Associative Computing Perspective

The new surge of interest in cloud computing is accompanied with the exponential growth of data sizes generated by digital media (images/audio/video), web authoring, scientific instruments, and physical simulations. Thus the question, how to effectively process these immense data sets is becoming increasingly urgent. Also, the opportunities for parallelization and distribution of data in clouds make storage and retrieval processes very complex, especially in facing with real-time data processing. Loosely-coupled associative computing techniques, which have so far not been considered, can provide the break through needed for cloud-based data management. Thus, a novel distributed data access scheme is introduced that enables data storage and retrieval by association, and thereby circumvents the partitioning issue experienced within referential data access mechanisms. In our model, data records are treated as patterns. As a result, data storage and retrieval can be performed using a distributed pattern recognition approach that is implemented through the integration of loosely-coupled computational networks, followed by a divide-and-distribute approach that allows distribution of these networks within the cloud dynamically.

Amir H. Basirat, Asad I. Khan

Factorizing Class Characteristics via Group MEBs Construction

Classic MEB (minimum enclosing ball) models characteristics of each class for classification by extracting core vectors through a (1 + 

ε

)-approximation problem solving. In this paper, we develop a new MEB system learning the core vectors set in a group manner, called group MEB (g-MEB). The g-MEB factorizes class characteristic in 3 aspects such as, reducing the sparseness in MEB by decomposing data space based on data distribution density, discriminating core vectors on class interaction hyperplanes, and enabling outliers detection to decrease noise affection. Experimental results show that the factorized core set from g-MEB delivers often apparently higher classification accuracies than the classic MEB.

Ye Chen, Shaoning Pang, Nikola Kasabov

A Hybrid Fuzzy-Genetic Colour Classification System with Best Colour Space Selection under Dynamically-Changing Illumination

This paper contributes in colour classification under dynamically changing illumination, extending further the capabilities of our previous works on Fuzzy Colour Contrast Fusion (FCCF), FCCF-Heuristic Assisted Genetic Algorithm (HAGA) for automatic colour classifier calibration and Variable Colour Depth (VCD). All the aforementioned algorithms were proven to accurately in real-time with a pie-slice technique. However, the pie-slice classifier is the accuracy-limiting factor in these systems. Although it is possible to address this problem by using a more complex shape for specifying the colour decision region, this would only increase the chances of overfitting. We propose a hybrid colour classification system that automatically searches for the best colour space for classifying any target colour. Moreover, this paper also investigates the general selection of training sets to get a better understanding of the generalisation capability of FCCF-HAGA. The experiments used a professional Munsell ColorChecker Chart with extreme illumination conditions where the colour channels start hitting their dynamic range limits.

Heesang Shin, Napoleon H. Reyes, Andre L. Barczak

Identifier Based Graph Neuron: A Light Weight Event Classification Scheme for WSN

Large-scale wireless sensor networks (WSNs) require significant resources for event recognition and classification. We present a light-weight event classification scheme, called Identifier based Graph Neuron (IGN). This scheme is based on highly distributed associative memory which enables the objects to memorize some of its internal critical states for a real time comparison with those induced by transient external conditions. The proposed approach not only conserves the power resources of sensor nodes but is also effectively scalable to large scale WSNs. In addition, our scheme overcomes the issue of

false-positive detection

-(which existing associated memory based solutions suffers from) and hence promises to deliver accurate results. We compare Identifier based Graph Neuron with two of the existing associated memory based event classification schemes and the results show that IGN correctly recognizes and classifies the incoming events in comparative amount of time and messages.

Nomica Imran, Asad Khan

Clustering Categorical Data Using an Extended Modularity Measure

Newman and Girvan [12] recently proposed an objective function for graph clustering called the Modularity function which allows automatic selection of the number of clusters. Empirically, higher values of the Modularity function have been shown to correlate well with good graph clustering. In this paper we propose an extended Modularity measure for categorical data clustering; first, we establish the connection with the Relational Analysis criterion. The proposed Modularity measure introduces an automatic weighting scheme which takes in consideration the profile of each data object. A modified Relational Analysis algorithm is then presented to search for the partitions maximizing the criterion. This algorithm deals linearly with large data set and allows natural clusters identification, i.e. doesn’t require fixing the number of clusters and size of each cluster. Experimental results indicate that the new algorithm is efficient and effective at finding both good clustering and the appropriate number of clusters across a variety of real-world data sets.

Lazhar Labiod, Nistor Grozavu, Younèns Bennani

A Morphological Associative Memory Employing a Reverse Recall

Recently the morphological associative memory proposed by Ritter attracts researcher’s attention. The model is superior to other models in terms of memory capacity and perfect recall rate. However the conventional MAM has a problem that the correct pattern cannot be recalled if a pattern has inclusive relation to other stored pattern. In this paper, as one of the solutions, an effective MAM employing a reverse recall is proposed. In the proposed method, candidate patterns of an input can be estimated by reverse recall from the kernel image recalled by a given inclusion input pattern, and then the plausible recall pattern can be determined by comparing the candidates with input pattern. We confirm the validity of the proposed method through hetero association experiments for twenty six alphabet patterns with inclusion patterns.

Hidetaka Harada, Tsutomu Miki

Analysis of Packet Traffics and Detection of Abnormal Traffics Using Pareto Learning Self Organizing Maps

Recently, the spread of the Internet makes familiar to the incident concerning the Internet, such as a DoS attack and a DDoS attack. Some methods which detect the abnormal traffics in the network using the information from headers and payloads of IP-packets transmitted in the networks are proposed. In this research, we propose a method of Pareto Learning SOM (Self Organizing Map) for IP packet flow analysis in which the occurrence rate is used for SOM computing. The flow of the packets can be visualized on the map and it can be used for detecting the abnormal flows of packets.

Hiroshi Dozono, Masanori Nakakuni, Takaru Kabashima, Shigeomi Hara

Log Analysis of Exploitation in Cloud Computing Environment Using Automated Reasoning

Recently server consolidation using virtualization leverages cloud computing. In cloud computing, we can apply centralized logging system using server consolidation. In this paper we propose a log analysis method in cloud computing environment using automated reasoning. On cloud computing providers, VM (virtual machine) monitoring is important to detect security incident. We discuss how to monitor VM, formatting and analyzing logs. Automated reasoning is more effective to retrieves information from large amount of log string. In proposed system, VM log is represented as clausal form and processed by FoL (First order Logic) theorem prover. We also present the numerical output of proposed system.

Ruo Ando, Kang Byung, Youki Kadobayashi

Self-organizing Maps and Their Applications

A Multidirectional Associative Memory Based on Self-organizing Incremental Neural Network

A multidirectional associative memory (AM) is proposed. It is constructed with three layer networks: an input layer, a memory layer, and an associate layer. The proposed method is able to realize many-to-many associations with no predefined conditions, and the association can be incrementally added to the network without destruction of old associations. Experiments show that the proposed AM works well for real tasks.

Hui Yu, Furao Shen, Osamu Hasegawa

Range Image Registration Using Particle Filter and Competitive Associative Nets

This paper presents a method using a particle filter (PF) and competitive associative nets (CAN2s) for range image registration to fuse 3D surfaces on range images taken from around an object by the laser range finder (LRF). The method uses the CAN2 for learning to provide a piecewise linear approximation of the LRF data involving various noise, and obtaining a coarse but fast pair-wise registration. The PF is used for reducing the cumulative error of the consecutive pair-wise registration. The effectiveness is shown by using the real LRF data of a rectangular box.

Shuichi Kurogi, Tomokazu Nagi, Takeshi Nishida

Rotation Invariant Categorization of Visual Objects Using Radon Transform and Self-Organizing Modules

The Radon transform in combination with self-organizing maps is used to build the rotation invariant systems for categorization of visual objects. The first system has one SOM per the Radon transform direction. The outputs from these directional SOMs that represent positions of the winners and related post-synaptic activities, form the input to the final categorizing SOM. Such a network delivers robust rotation invariant categorization of images rotated by angles up to around 12

o

. In the second network the angular Radon transform vectors are combined together and form the input to the categorizing SOM. This network can correctly categorized visual stimuli rotated by up to 30

o

. The rotation invariance can be improved by increasing the number of Radon transform angle, which has been equal to six in our initial experiments.

Andrew P. Papliński

Learning Topological Constraints in Self-Organizing Map

The Self-Organizing Map (SOM) is a popular algorithm to analyze the structure of a dataset. However, some topological constraints of the SOM are fixed before the learning and may not be relevant regarding to the data structure. In this paper we propose to improve the SOM performance with a new algorithm which learn the topological constraints of the map using data structure information. Experiments on artificial and real databases show that algorithm achieve better results than SOM. This is not the case with trivial topological constraint relaxation because of the high increase of the Topological error.

Guénaël Cabanes, Younès Bennani

Pseudo-network Growing for Gradual Interpretation of Input Patterns

In this paper, we propose a new information-theoretic method to interpret competitive learning. The method is called ”pseudo-network growing,” because a network re-grows gradually after learning, taking into account the importance of components. In particular, we try to apply the method to clarify the class structure of self-organizing maps. First, the importance of input units is computed, and then input units are gradually added, according to their importance. We can expect that the corresponding number of competitive units will be gradually increased, showing the main characteristics of network configurations and input patterns. We applied the method to the well-known Senate data with two distinct classes. By using the conventional SOM, explicit class boundaries could not be obtained, due to the inappropriate map size imposed in the experiment. However, with the pseudo-network growing, a clear boundary could be observed in the first growing stage, and gradually the detailed class structure could be reproduced.

Ryotaro Kamimura

The Adaptive Authentication System for Behavior Biometrics Using Pareto Learning Self Organizing Maps

In this paper, we propose an authentication system which can adapt to the temporal changes of the behavior biometrics with accustoming to the system. We proposed the multi-modal authentication system using Supervised Pareto learning Self Organizing Maps. In this paper, the adaptive authentication system with incremental learning which is applied as the feature of neural networks is developed.

Hiroshi Dozono, Masanori Nakakuni, Shinsuke Itou, Shigeomi Hara

Human Action Recognition by SOM Considering the Probability of Spatio-temporal Features

In this paper, an action recognition system was invented by proposing a compact 3D descriptor to represent action information, and employing self-organizing map (SOM) to learn and recognize actions. Histogram Of Gradient 3D (HOG3D) performed better among currently used descriptors for action recognition. However, the calculation of the descriptor is quite complex. Furthermore, it used a vector with 960 elements to describe one interest point. Therefore, we proposed a compact descriptor, which shortened the support region of interest points, combined symmetric bins after orientation quantization. In addition, the top value bin of quantized vector was kept instead of setting threshold experimentally. Comparing with HOG3D, our descriptor used 80 bins to describe a point, which reduced much computation complexity. The compact descriptor was used to learn and recognize actions considering the probability of local features in SOM, and the results showed that our system outperformed others both on KTH and Hollywood datasets.

Yanli Ji, Atsushi Shimada, Rin-ichiro Taniguchi

On Generalization Error of Self-Organizing Map

Self-organizing map is usually used for estimation of a low dimensional manifold in a high dimensional space. The main purpose of applying it is to extract the hidden structure from samples, hence it has not been clarified how accurate the estimation of the low dimensional manifold is. In this paper, in order to study the accuracy of the statistial estimation using the self-organizing map, we define the generalization error, and show its behavior experimentally. Based on experiments, it is shown that the learning curve of the self-organizing map is determined by the order that are smaller than dimensions of parameter. We consider that the topology of self-organizing map contributed to abatement of the order.

Fumiaki Saitoh, Sumio Watanabe

A Novel Approach for Sound Approaching Detection

The detection of approaching vehicles is a very important topic on the development of complementary traffic safety systems. However, the majority of the proposed approaches are very complex and not suitable for embedded applications. This paper proposes a new sound approaching detection algorithm specifically intended for hardware implementation. Experimental results show higher accuracy and earlier detection when comparing to other methods.

Hirofumi Tsuzuki, Mauricio Kugler, Susumu Kuroyanagi, Akira Iwata

Ground Penetrating Radar System with Integration of Mutimodal Information Based on Mutual Information among Multiple Self-Organizing Maps

We propose a ground penetrating radar system to integrate mutimodal information of space- and frequency- domain textural features in self-organization that is modulated by mutual information. We use the MuSOM (mutual-information-based self-organizing map) architecture we proposed previously, in which the mutual information among the data fed to multiple SOMs modulates the SOM dynamics. Experiments demonstrate that our system makes meaningful clusters of landmine features clearer than a conventional non-MuSOM system does.

Akira Hirose, Ayato Ejiri, Kunio Kitahara

Information-Theoretic Competitive and Cooperative Learning for Self-Organizing Maps

In this paper, we propose a new type of information-theoretic method for competitive learning based, upon mutual information between competitive units and input patterns. In addition, we extend this method to a case where cooperation between competitive units exists to realize self-organizing maps. In computational methods, free energy is introduced to simplify the computation of mutual information. We applied our method to two problems, namely, the Senate data and ionosphere data problems. In both, experimental results confirmed that better performance could be obtained in terms of quantization and topographic errors. We also found that the information-theoretic methods tended to produce more equi-probable distribution of competitive units.

Ryotaro Kamimura

Early Recognition Based on Co-occurrence of Gesture Patterns

We propose an approach to achieve early recognition of gesture patterns. We assume that there are two people who interact with a machine, a robot or something. In such a situation, a gesture of a person often has a relationship with a gesture of another person. We exploit such a relationship to realize early recognition of gesture patterns. Early recognition is a method to recognize sequential patterns at their beginning parts. Therefore, in the case of gesture recognition, we can get a recognition result of human gestures before the gestures have finished. Recent years, some approaches have been proposed. In this paper, we expand the application range of early recognition to multiple people based on the co-occurrence of gesture patterns. In our approach, we use Self-Organizing Map to represent gesture patterns of each person, and associative memory based approach learns the relationship between co-occurring gestures. In the experiments, we have found that our proposed method achieved the early recognition more accurately and earlier than the traditional approach.

Atsushi Shimada, Manabu Kawashima, Rin-ichiro Taniguchi

A Dynamically Reconfigurable Platform for Self-Organizing Neural Network Hardware

In this paper, we propose a dynamically reconfigurable platform for self-organizing neural network hardware. In the proposed platform, a hardware unit can be handled as a hardware object in object-oriented design. The hardware object is loaded into the FPGA’s virtual hardware circuit space, and accelerates the calculation of self-organizing neural networks. We design two types of the distance calculation, a winner-take-all and a rough-winner-take-all virtual hardware circuit as common parts of self-organizing neural networks. By combining them, we realize four types of self-organizing neural network. Experimental results show that the implemented self-organizing neural network hardware achieves about 100 times faster than the software implementation. Besides, the proposed platform can change its learning mode easily as well as the software implementation. Therefore, the proposed platform features both of the speed of hardware and the flexibility of software.

Hakaru Tamukoh, Masatoshi Sekine

Inversion of Many-to-one Mappings Using Self-Organising Maps

Bidirectionally trained neural networks would be very useful in many circumstances. Often, we have data available for a prediction problem, but prediction of properties for unknown or new situations is only part of the story. In many cases we know the effect we wish to achieve on the output, but what we do not know is how to modify the inputs to achieve this goal. A basic problem in this area is the inversion of many to one mappings. Our work is based on the popular backpropagation neural network to predict the GDP of developing countries. These networks are integrated with a Self-Organising Map to allow the inversion of many to one mappings.

Anne O. Mus

Self-Organizing Hidden Markov Models

The self-organizing mixture models (SOMMs) were proposed as an expectation-maximization (EM) algorithm that yields topology preserving maps of data based on probabilistic mixture models. Compared to self-organizing maps, the SOMM algorithm has a clear interpretation: it maximizes the sum of data log likelihood and a penalty term that enforces self-organization. The object of this paper is to extend the SOMM algorithm to deal with multivariate time series. The standard SOMM algorithm assumes that the data are independent and identically distributed samples. However, the i.i.d. assumption is clearly inappropriate for time series. In this paper we propose the extension of the SOMM algorithm for multivariate time series, which we call self-organizing hidden Markov models (SOHMMs), by assuming that the time series is generated by hidden Markov models (HMMs).

Nobuhiko Yamaguchi

An Image-Aided Diagnosis System for Dementia Classification Based on Multiple Features and Self-Organizing Map

Mild cognitive impairment (MCI) is considered as a transitional stage between normal aging and dementia. MCI has a high risk to convert into Alzheimer’s disease (AD). In the related research, the volumetric analysis of hippocampus is the most extensive study. However, the segmentation and identification of the hippocampus are highly complicated and time-consuming. Therefore, we designed a MRI-based classification framework to distinguish the patients of MCI and AD from normal individuals. First, volumetric features and shape features were extracted from MRI data. Afterward, Principle component analysis (PCA) was utilized to decrease the dimensions of feature space. Finally, a Self-organizing map classifier was trained for patient classification. By combining the volumetric features and shape features, the classification accuracy is reached to 86.76%, 66.67%, and 46.67% in AD, amnestic MCI (aMCI), and dysexecutive MCI (dMCI), respectively. In addition, with the help of PCA process, the classification result is improved to 93.63%, 73.33%, and 53.33% in AD, aMCI and dMCI, respectively.

Shih-Ting Yang, Jiann-Der Lee, Chung-Hsien Huang, Jiun-Jie Wang, Wen-Chuin Hsu, Yau-Yau Wai

Parallel Batch Training of the Self-Organizing Map Using OpenCL

The Self-Organizing Maps (SOMs) are popular artificial neural networks that are often used for data analyses through clustering and visualisation. SOM’s mathematical model is inherently parallel. However, many implementations have not successfully exploited its parallelism because previous attempts often required cluster-like infrastructures. This article presents the parallel implementation of SOMs, particularly the batch map variant using Graphics Processing Units (GPUs) through the use of Open Computing Language (OpenCL).

Masahiro Takatsuka, Michael Bui

Fast Kohonen Feature Map Associative Memory Using Area Representation for Sequential Analog Patterns

In this paper, we propose a Fast Kohonen Feature Map Associative Memory with Area Representation for Sequential Analog Patterns (FKFMAM-AR-SAP). This model is based on the conventional Improved Kohonen Feature Map Associative Memory with Area Representation for Sequential Analog Patterns (IKFMAM-AR-SAP). The proposed model can realize the one-to-many associations even when the first patterns are same in the plural sequential patterns. And, it has enough robustness for noisy input and damaged neurons. Moreover, the learning speed of the proposed model is faster than that of the conventional model. We carried out a series of computer experiments and confirmed the effectiveness of the proposed model.

Hiroki Midorikawa, Yuko Osana

Machine Learning Applications to Image Analysis

Facial Expression Based Automatic Album Creation

With simple cost effective imaging solutions being widely available these days, there has been an enormous rise in the number of images consumers have been taking. Due to this increase, searching, browsing and managing images in multi-media systems has become more complex. One solution to this problem is to divide images into albums for meaningful and effective browsing. We propose a novel automated, expression driven image album creation for consumer image management systems. The system groups images with faces having similar expressions into albums. Facial expressions of the subjects are grouped into albums by the Structural Similarity Index measure, which is based on the theory on how easily the human visual system can extract the shape information of a scene. We also propose a search by similar expression, in which the user can create albums by providing example facial expression images. A qualitative analysis of the performance of the system is presented on the basis of a user study.

Abhinav Dhall, Akshay Asthana, Roland Goecke

Age Classification Combining Contour and Texture Feature

Age classification based on computer vision has widespread applications. Most of previous works only utilize texture feature or use contour and texture feature separately. In this paper, we proposed an age classification system that integrate contour and texture information. Besides, we improve the traditional Local Binary Pattern(LBP) feature extraction method and get pure texture feature. Support Vector Machines with probabilistic output (SVM-PO) is used as individual classifiers. Then we use combination mechanism based on fuzzy integral to merge the output of different classifiers. The experiment results show pure texture feature outperforms other features and it can be well combined with contour feature.

Yan-Ming Tang, Bao-Liang Lu

A Salient Region Detector for GPU Using a Cellular Automata Architecture

The human visual cortex performs salient region detection, a process critical to the rapid understanding of a scene. This is performed on large arrays of locally interacting neurons that are slow to simulate sequentially. In this paper we describe and evaluate a novel, bio-inspired, cellular automata (CA) architecture for the determination of the salient regions within a scene. This parallel processing architecture is appropriate for implementation on a graphics processing unit (GPU). We compare the performance of this algorithm against that of CPU implemented salient region detectors. The CA algorithm is less subject to variation due to changing scale, viewpoint and illumination conditions. Also due to its GPU implementation, this algorithm is able to detect salient regions faster than the CPU implemented algorithms.

David Huw Jones, Adam Powell, Christos-Savvas Bouganis, Peter Y. K. Cheung

VG-RAM WNN Approach to Monocular Depth Perception

We have examined Virtual Generalizing Random Access Memory Weightless Neural Networks (VG-RAM WNN) as platform for depth map inference from static monocular images. For that, we have designed, implemented and compared the performance of VG-RAM WNN systems against that of depth estimation systems based on Markov Random Field (MRF) models. While not surpassing the performance of such systems, our results are consistent to theirs, and allow us to infer important features of the human visual cortex.

Hélio Perroni Filho, Alberto F. De Souza

Semi-supervised Classification by Local Coordination

Graph-based methods for semi-supervised learning use graph to smooth the labels of the points. However, most of them are transductive thus can’t give predictions for the unlabeled data outside the training set directly. In this paper, we propose an inductive graph-based algorithm that produces a classifier defined on the whole ambient space. A smooth nonlinear projection between the sample space and the label value space is achieved by local dimension reduction and coordination. The effectiveness of the proposed algorithm is demonstrated by the experiment.

Gelan Yang, Xue Xu, Gang Yang, Jianming Zhang

RANSAC Based Ellipse Detection with Application to Catadioptric Camera Calibration

In this paper, a simple method for ellipse detection is proposed and applied in central catadioptric camera calibration. It consists of two phases. Firstly it locates ellipse center candidates using center symmetry of ellipses, and the detected edge points are grouped into several subsets according to the center candidates. Then all the ellipses are fitted by performing RANSAC for each subset. We also present an approach for calibrating a central catadioptric camera based on the bounding ellipse of the catadioptric image. Using the proposed ellipse detection method, we can easily detect the bounding ellipse. As a result, a simple self-calibration can be realized, which can be used in some applications where high accuracy of the calibration is not required. Experiments show the proposed method is effective.

Fuqing Duan, Liang Wang, Ping Guo

Speed Up Image Annotation Based on LVQ Technique with Affinity Propagation Algorithm

For a support vector machine (SVM) classifier applied to image annotation, if too many training samples are used, the training speed might be very slow and also bring the problem of declining the classification accuracy. Learning vector quantization (LVQ) technique provides a framework to select some representative vectors which can be used to train the classifier instead of using original training data. A novel method which combines affinity propagation algorithm based LVQ technique and SVM classifier is proposed to annotate images. Experimental results demonstrate that proposed method has a better speed performance than that of SVM without applying LVQ.

Song Lin, Yao Yao, Ping Guo

Dictionary of Features in a Biologically Inspired Approach to Image Classification

We introduce new methods for creation of a dictionary of features for a biologically inspired model of visual object classification that is shown to handle the recognition of several object categories. We provide a new method for creation of this features dictionary using non-supervised cortex like methods. Different clustering approaches were experimented and improved performance is achieved on image centers which results in real time classification of images by HMAX model.

Sepehr Jalali, Joo Hwee Lim, Sim Heng Ong, Jo Yew Tham

A Highly Robust Approach Face Recognition Using Hausdorff-Trace Transformation

Face recognition research still face challenge in some specific domains such as pose, illumination and Expression. In this paper, we proposes a highly robust method for face recognition with variant illumination, scaling, rotation, blur, reflection and expression. Techniques introduced in this work are composed of two parts. The first one is the detection of facial features by using the concepts of Trace Transform and Fourier transform. Then, in the second part, the Hausdorff distance is employed to measure and determine of similarity between the models and tested images. Finally, our method is evaluated with experiments on the AR, ORL, Yale and XM2VTS face databases and compared with other related works (e.g. Eigen face and Hausdorff ARTMAP). The extensive experimental results show that the average of accuracy rate of face recognition with variant illumination, scaling, rotation, blur, reflection and difference emotions is higher than 88%.

Werasak Kurutach, Rerkchai Fooprateepsiri, Suronapee Phoomvuthisarn

Blind Image Tamper Detection Based on Multimodal Fusion

In this paper, we propose a novel feature processing approach based on fusion of noise and quantization residue features for detecting tampering or forgery in video sequences. The evaluation of proposed residue features – the noise residue features and the quantization features, their transformation in optimal feature subspace based on fisher linear discriminant features and canonical correlation analysis features, and their subsequent fusion for emulated copy-move tamper scenarios shows a significant improvement in tamper detection accuracy.

Girija Chetty, Monica Singh, Matthew White

Orientation Dependence of Surround Modulation in the Population Coding of Figure/Ground

Recent physiological studies have reported Border-Ownership (BO) selective cells that signal the direction of figure along a contour, which appear to be a basis for figure-ground segregation. Surround modulation has been proposed as an underlying neural mechanism of BO determination. The crucial question to the model is its orientation specificity: whether BO could be determined only from iso-orientation (with respect to the preferred orientation of the classical receptive field) that has been reported dominant in the modulation. We investigated computationally the dependence of surround modulation on the orientation characteristics during the determination of BO with natural images. The results showed that, even when modulation was limited to iso-orientation, population responses obtained through integration were unchanged while the responses of individual cells were varied, indicating a dominant role of iso-orientation suppression and an effectiveness of population coding in BO determination.

Keiichi Kondo, Ko Sakai

Increased Robustness against Background Noise: Pattern Recognition by a Neocognitron

The

neocognitron

is a hierarchical multi-layered neural network capable of robust visual pattern recognition. It has been demonstrated that recent versions of the neocognitron exhibit excellent performance for recognizing handwritten digits. When characters are written on a noisy background, however, recognition rate was not always satisfactory. This paper proposes several modifications, by which the neocognitrons can be much more robust against background noise.

Kunihiko Fukushima

Improving the Performance of Facial Expression Recognition Using Dynamic, Subtle and Regional Features

Human facial expression is a complex process characterized of dynamic, subtle and regional emotional features. State-of-the-art approaches on facial expression recognition (FER) have not fully utilized this kind of features to improve the recognition performance. This paper proposes an approach to overcome this limitation using patch-based ‘salient’ Gabor features. A set of 3D patches are extracted to represent the subtle and regional features, and then inputted into patch matching operations for capturing the dynamic features. Experimental results show a significant performance improvement of the proposed approach due to the use of the dynamic features

.

Performance comparison with pervious work also confirms that the proposed approach achieves the highest CRR reported to date on the JAFFE database and a top-level performance on the Cohn-Kanade (CK) database.

Ligang Zhang, Dian Tjondronegoro

Identity Retrieval in Biometric Access Control Systems Using Multimedia Fusion

In this paper, we propose a novel multimedia sensor fusion approach based on heterogeneous sensors for biometric access control applications. The proposed fusion technique uses multiple acoustic and visual sensors for extracting dominant biometric cues, and combines them with non-dominant cues. The performance evaluation of the proposed fusion protocol and a novel cascaded authentication approach using a 3D stereovision database shows a significant improvement in performance and robustness.

Girija Chetty, Renuka Biswas, Julian Goodwin

Improvement of Reuse of Classifiers in CBIR Using SVM Active Learning

In content-based image retrieval, relevance feedback is often adopted as the method of interactions to grasp user’s query concept. However, since this method tasks the user, a small amount of relevance feedback is desirable. For this purpose, Nakajima

et al.

have proposed a method in which classifiers learned by using relevance feedback are reused. In this paper, we improve the criterion for reuse of classifiers so that retrieval becomes more accurate and quick. Experimental results show that our method performs much better than the conventional methods.

Masaaki Tekawa, Motonobu Hattori

Realizing Hand-Based Biometrics Based on Visible and Infrared Imagery

This paper describes a hand-based biometric system by using visible and infrared imagery. We develop an acquisition device which could capture both color and infrared hand images. We modify an ordinary web camera to capture the hand vein that normally requires specialized infrared sensor. Our design is simple and low-cost, and we do not need additional installation of special apparatus. The device can capture the epidermal and subcutaneous features from the hand simultaneously. In specific, we acquire four independent, yet complementary features namely palm print, knuckle print, palm vein, and finger vein, from the hand for recognition. As a low-resolution sensor is deployed in this study, the images quality may be slightly poorer than those acquired using high resolution scanner or CCD camera. The line and ridge patterns on the hand may not appear clear. Therefore, we propose a pre-processing technique to enhance the contrast and sharpness of the images so that the dominant print and line features can be highlighted and become disguisable from the background. After that, we use a simple feature extractor called Directional Coding to obtain useful representation of the hand modalities. The hand features are fused using Support Vector Machine (SVM). The fusion of these features yields promising result for practical multi-modal biometrics system.

Goh Kah Ong Michael, Tee Connie, Teo Chuan Chin, Neo Han Foon, Andrew Teoh Beng Jin

Visual Object Detection by Specifying the Scale and Rotation Transformations

We here propose a simple but highly potential algorithm to detect a model object’s position on an input image by determining the initially unknown transformational states of the model object, in particular, size and 2D-rotation. In this algorithm, a single feature is extracted around or at the center of the input image through 2D-Gabor wavelet transformation, in order to find not only the most likely relative size and rotation to the model object, but also the most appropriate positional region on the input image for detecting the correct relative transformational states. We also show the reliable function on the face images of different persons, or of different appearance in the same person.

Yasuomi D. Sato, Jenia Jitsev, Christoph von der Malsburg

Multi-view Gender Classification Using Hierarchical Classifiers Structure

In this paper, we propose a hierarchical classifier structure for gender classification based on facial images by reducing the complexity of the original problem. In the proposed framework, we first train a classifier, which will properly divide the input images into several groups. For each group, we train a gender classifier, which is called expert. These experts can be any commonly used classifiers, such as Support Vector Machine (SVM) and neural network. The symmetrical characteristic of human face is utilized to further reduce the complexity. Moreover, we adopt soft assignment instead of hard one when dividing the input data, which can reduce the error introduced by the division. Experimental results demonstrate that our framework significantly improves the performance.

Tian-Xiang Wu, Bao-Liang Lu

Partial Extraction of Edge Filters by Cumulant-Based ICA under Highly Overcomplete Model

It has been well known that ICA can extract edge filters from natural scenes. However, it has been also known that the existing cumulant-based ICA can not extract edge filters. It suggests that the simple ICA model is insufficient for explaining the properties of natural scenes. In this paper, we propose a highly overcomplete model for natural scenes. Besides, we show that the 4-th order covariance has a positive constant lower bound under this model. Then, a new cumulant-based ICA algorithm is proposed by utilizing this lower bound. Numerical experiments show that this cumulant-based algorithm can extract edge filters.

Yoshitatsu Matsuda, Kazunori Yamaguchi

Random Projection Tree and Multiview Embedding for Large-Scale Image Retrieval

Image retrieval on large-scale datasets is challenging. Current indexing schemes, such as k-d tree, suffer from the “curse of dimensionality”. In addition, there is no principled approach to integrate various features that measure multiple views of images, such as color histogram and edge directional histogram. We propose a novel retrieval system that tackles these two problems simultaneously. First, we use random projection trees to index data whose complexity only depends on the low intrinsic dimension of a dataset. Second, we apply a probabilistic multiview embedding algorithm to unify different features. Experiments on MSRA large-scale dataset demonstrate the efficiency and effectiveness of the proposed approach.

Bo Xie, Yang Mu, Mingli Song, Dacheng Tao

Online Gesture Recognition for User Interface on Accelerometer Built-in Mobile Phones

Recently, several smart phones are equipped with a 3D-accelerometer that can be used for gesture-based user interface (UI). In order to utilize the gesture UI for the real-time systems with various users, the diversity robust algorithm, yet having low training/recognition complexity, is required. Meantime, dynamic time warping (DTW) has shown good performance on the simple time-series pattern recognition problems. Since DTW is based on the template matching, its processing time and accuracy depend on the number of templates and their quality, respectively. In this paper, an optimized method for online gesture UI of mobile devices is proposed which is based on the DTW and modified k-means clustering algorithm. The templates, estimated by using the modified clustering algorithm, can preserve the time varying attribute while contain diversities of the given training patterns. The proposed method was validated on 20 types of gestures which are designed for the mobile contents browsing. The experimental results showed that the proposed method is suitable to the online mobile UI.

BongWhan Choe, Jun-Ki Min, Sung-Bae Cho

Constructing Sparse KFDA Using Pre-image Reconstruction

Kernel Fisher Discriminant Analysis (KFDA) improves greatly the classification accuracy of FDA via using kernel trick. However, the final solution of KFDA is expressed as an expansion of all training examples, which seriously undermines the classification efficiency, especially in real-time applications. This paper proposes a novel framework to construct sparse KFDA using pre-image reconstruction. The proposed method (PR-KFDA) appends greedily the pre-image of the residual between the current approximate model and the original KFDA model in feature space with the local optimal Fisher coefficients to acquire sparse representation of KFDA solution. Experimental results show that PR-KFDA can reduce the solution of KFDA effectively while maintaining comparable test accuracy.

Qing Zhang, Jianwu Li

Applications

Learning Basis Representations of Inverse Dynamics Models for Real-Time Adaptive Control

In this paper, we propose a novel approach for adaptive control of robotic manipulators. Our approach uses a representation of inverse dynamics models learned from a varied set of training data with multiple conditions obtained from a robot. Since the representation contains various inverse dynamics models for the multiple conditions, adjusting a linear coefficient vector of the representation efficiently provides real-time adaptive control for unknown conditions rather than solving a high-dimensional learning problem. Using this approach for adaptive control of a trajectory-tracking problem with an anthropomorphic manipulator in simulations demonstrated the feasibility of the approach.

Yasuhito Horiguchi, Takamitsu Matsubara, Masatsugu Kidode

Feel Like an Insect: A Bio-Inspired Tactile Sensor System

Insects use their antennae (feelers) as near range sensors for orientation, object localization and communication. This paper presents an approach for an active tactile sensor system. This includes a new type of hardware construction as well as a software implementation for interpreting the sensor readings. The discussed tactile sensor is able to detect an obstacle and its location in 3D space. Furthermore the material properties of the obstacles are classified by use of neural networks.

Sven Hellbach, André Frank Krause, Volker Dürr

Spectral Domain Noise Suppression in Dual-Sensor Hyperspectral Imagery Using Gaussian Processes

The use of hyperspectral data is limited, in part, by increased spectral noise, particularly at the extremes of the wavelength ranges sensed by scanners. We apply Gaussian Processes (GPs) as a preprocessing step prior to extracting mineralogical information from the image using automated feature extraction. GPs are a probabilistic machine learning technique that we use for suppressing noise in the spectral domain. The results demonstrate that this approach leads to large reductions in the amount of noise, leading to major improvements in our ability to automatically quantify the abundance of iron and clay minerals in hyperspectral data acquired from vertical mine faces.

Arman Melkumyan, Richard J. Murphy

A High Order Neural Network to Solve Crossbar Switch Problem

High-order neural networks can be considered as an expansion of Hopfield neural networks, and have stronger approximation property and faster convergence rate. However, in practice high order network is seldom to be used to solve combinatorial optimization problem. In this paper crossbar switch problem, which is an NP-complete problem, is used as an example to demonstrate how to use high order discrete Hopfield neural network to solve engineering optimization problems. The construction method of energy function and the neural computing algorithm are presented. It is also discussed the method how to speed the convergence and escape from local minima. Experimental results show that high order network has a quick convergence speed, and outperforms the traditional discrete Hopfield network.

Yuxin Ding, Li Dong, Ling Wang, Guohua Wu

Identification of Liquid State of Scrap in Electric Arc Furnace by the Use of Computational Intelligence Methods

A constant aspiration to optimize electric arc steelmaking process causes an increase of the use of advanced analytical methods for the process support. Optimization of the production processes lead to real benefits, which are, for example, lower costs of production. More often computational intelligence methods are used for this purpose. In this paper authors present three methods used for identification of liquid state of scrap in electric arc furnace using analysis of signals of the current of furnace electrodes.

Marcin Blachnik, Tadeusz Wieczorek, Krystian Mączka, Grzegorz Kopeć

Simulating Wheat Yield in New South Wales of Australia Using Interpolation and Neural Networks

Accurate modeling of wheat production in advance provides wheat growers, traders, and governmental agencies with a great advantage in planning the distribution of wheat production. The conventional approach in dealing with such prediction is based on time series analysis through statistical or intelligent means. These time-series based methods are not concerned about the factors that cause the sequence of the events. In this paper, we treat the historical wheat data in New South Wales over 130 years as non-temporal collection of mappings between wheat yield and both wheat plantation area and rainfall through data expansion by 2D interpolation. Neural networks are then used to define a dynamic system using these mappings to achieve modeling wheat yield with respect to both the plantation area and rainfall. No similar study has been reported in the world in this field. Our results demonstrate that a four-layer multilayer perceptron model is capable of producing accurate modeling for wheat yield.

William W. Guo, Lily D. Li, Greg Whymark

Investment Appraisal under Uncertainty – A Fuzzy Real Options Approach

The main purpose of this paper is to propose a fuzzy approach for investment project valuation in uncertain environments from the aspect of real options. The traditional approaches to project valuation are based on discounted cash flows (DCF) analysis which provides measures like net present value (NPV) and internal rate of return (IRR). However, DCF-based approaches exhibit two major pitfalls. One is that DCF parameters such as cash flows cannot be estimated precisely in the uncertain decision making environments. The other one is that the values of managerial flexibilities in investment projects cannot be exactly revealed through DCF analysis. Both of them would entail improper results on strategic investment projects valuation. Therefore, this paper proposes a fuzzy binomial approach that can be used in project valuation under uncertainty. The proposed approach also reveals the value of flexibilities embedded in the project. Furthermore, this paper provides a method to compute the mean value of a project’s fuzzy expanded NPV that represents the entire value of project. Finally, we use the approach to practically evaluate a project.

Shu-Hsien Liao, Shiu-Hwei Ho

Developing a Robust Prediction Interval Based Criterion for Neural Network Model Selection

This paper studies how an optimal Neural Network (NN) can be selected that is later used for constructing the highest quality delta-based Prediction Intervals (PIs). It is argued that traditional assessment criteria, including RMSE, MAPE, BIC, and AIC, are not the most appropriate tools for selecting NNs from a PI-based perspective. A new NN model selection criterion is proposed using the specific features of the delta method. Using two synthetic and real case studies, it is demonstrated that this criterion outperforms all traditional model selection criteria in terms of picking the most appropriate NN. NNs selected using this criterion generate high quality PIs evaluated by their length and coverage probability.

Abbas Khosravi, Saeid Nahavandi, Doug Creighton

Backmatter

Weitere Informationen