Advances in Natural Computation

Wavelet neural networks(WNN) are a class of neural networks consisting of wavelets. A novel learning method based on immune genetic algorithm(IGA) for continuous wavelet neural networks is presented in this paper. Through adopting multi-encoding, this algorithm can optimize the structure and the parameters of WNN in the same training process. Simulation results show that WNN with novel algorithm has a comparatively simple structure and enhance the probability for global optimization. The study also indicates that the proposed method has the potential to solve a wide range of neural network construction and training problems in a systematic and robust way.

The MARN has the same structure as the RBF network and has the ability to grow and prune the hidden neurons to realize a minimal network structure. Several algorithms have been used to training the network. This paper proposes the use of Unscented Kalman Filter (UKF) for training the MRAN parameters i.e. centers, radii and weights of all the hidden neurons. In our simulation, we implemented the MRAN trained with UKF and the MRAN trained with EKF for states estimation. It is shown that the MRAN trained with UKF is superior than the MRAN trained with EKF.

An improved neural networks online learning scheme is proposed to speed up the learning process in cerebellar model articulation controllers(CMAC). The improved learning approach is to use the learned times of the addressed hypercubes as the credibility (confidence) of the learned values in the early learning stage, and the updating data for addressed hypercubes is proportional to the inverse of the exponent of learned times, in the later stage the updating data for addressed hypercubes is proportional to the inverse of learned times. With this idea, the learning speed can indeed be improved.

A new smooth support vector regression based on

ε

-insensitive logistic loss function, shortly L

ε

-SSVR, was proposed in this paper, which is similar to SSVR, but without adding any heuristic smoothing parameters and with robust absolute loss. Taking advantage of L

ε

-SSVR, one can now consider SVM as linear programming, and efficiently solve large-scale regression problems without any optimization packages. Details of this algorithm and its implementation were presented in this paper. Simulation results for both artificial and real data show remarkable improvement of generalization performance and training time.

In this study, two methods for the electrocardiogram (ECG) QRS waves detection were presented and compared. One hand, a modified approach of the linear approximation distance thresholding (LADT) algorithm was studied and the features of the ECG were gained for the later work. The other hand, Mexican-hat wavelet transform was adopted to detect the character points of ECG. A part of the features of the ECG were used to train the RBF network, and then all of them were used to examine the performance of the network. The algorithms were tested with ECG signals of MIT-BIH, and compared with other tests, the result shows that the detection ability of the Mexican-hat wavelet transform is very good for its quality of time-frequency representation and the ECG character points was represented by the local extremes of the transformed signals and the correct rate of QRS detection rises up to 99.9%. Also, the classification performance with its result is so good that the correct rate with the trained wave is 100%, and untrained wave is 86.6%.

A new hybrid learning algorithm combining the extended Kalman filter (EKF) and particle filter is presented. The new algorithm is firstly applied to train diagonal recurrent neural network (DRNN). The EKF is used to train DRNN and particle filter applies the resampling algorithm to optimize the particles, namely DRNNs, with the relative network weights. These methods make the training shorter and DRNN convergent more quickly. Simulation results of the nonlinear dynamical identification verify the validity of the new algorithm.

Based on rolling optimization method and on-line learning strategies, a novel weighted least squares support vector machines (WLS-SVM) are proposed for nonlinear system identification in this paper. The good robust property of the novel approach enhances the generalization ability of LS-SVM method, and a real world nonlinear time-variant system is presented to test the feasibility and the potential utility of the proposed method.

Some sufficient conditions for the globally exponential stability of the equilibrium point of neural networks with multiple time varying delays are developed, and the estimation of the exponential convergence rate is presented. The obtained criteria are dependent on time delay, and consist of all the information on the neural networks. The effects of time delay and number of connection matrices of the neural networks on the exponential convergence rate are analyzed, which can give a clear insight into the relation between the exponential convergence rate and the parameters of the neural networks. Two numerical examples are used to demonstrate the effectiveness of the obtained the results.

The

k

-nearest neighbor rule is one of the most attractive pattern classification algorithms. In practice, the value of

k

is usually determined by the cross-validation method. In this work, we propose a new method that locally determines the number of nearest neighbors based on the concept of statistical confidence. We define the confidence associated with decisions that are made by the majority rule from a finite number of observations and use it as a criterion to determine the number of nearest neighbors needed. The new algorithm is tested on several real-world datasets and yields results comparable to those obtained by the

k

-nearest neighbor rule. In contrast to the

k

-nearest neighbor rule that uses a fixed number of nearest neighbors throughout the feature space, our method locally adjusts the number of neighbors until a satisfactory level of confidence is reached. In addition, the statistical confidence provides a natural way to balance the trade-off between the reject rate and the error rate by excluding patterns that have low confidence levels.

The fuzzy self-organizing map neural network using kernel principal component analysis is presented and a hybrid-learning algorithm (KPCA-FSOM) divided into two stages is proposed to train this network. The first stage, the KPCA algorithm is applied to extract the features of nonlinear data. The second stage, combining both the fuzzy theory and locally-weight distortion index to extend SOM basic algorithm, the fuzzy SOM algorithm is presented to train the SOM network with features gained. A real life application of KPCA-FSOM algorithm in classifying data of acrylonitrile reactor is provided. The experimental results show this algorithm can obtain better clustering and network after training can more effectively monitor yields

.

An evolved recurrent neural network is proposed which automates the design of the network architecture and the connection weights using a new evolutionary learning algorithm. This new algorithm is based on a cooperative system of evolutionary algorithm (EA) and particle swarm optimisation (PSO), and is thus called REAPSO. In REAPSO, the network architecture is adaptively adjusted by PSO, and then EA is employed to evolve the connection weights with this network architecture, and this process is alternated until the best neural network is accepted or the maximum number of generations has been reached. In addition, the strategy of EAC and ET are proposed to maintain the behavioral link between a parent and its offspring, which improves the efficiency of evolving recurrent neural networks. A recurrent neural network is evolved by REAPSO and applied to the state estimation of the CSTR System. The performance of REAPSO is compared to TDRB, GA, PSO and HGAPSO in these recurrent networks design problems, demonstrating its superiority.

Locally Linear Embedding (LLE) is an efficient nonlinear algorithm for mapping high-dimensional data to a low-dimensional observed space. However, the algorithm is sensitive to several parameters that should be set artificially, and the resulting maps may be invalid in case of noises. In this paper, the original LLE algorithm is improved by introducing the self-organizing features of a novel SOM model we proposed recently called DGSOM to overcome these shortages. In the improved algorithm, nearest neighbors are selected automatically according to the topology connections derived from DGSOM. The proposed algorithm can also estimate the intrinsic dimensionality of the manifold and eliminate noises simultaneously. All these advantages are illustrated with abundant experiments and simulations.

In many neural network applications, the selection of best training set to represent the entire sample space is one of the most important problems. Active learning algorithms in the literature for neural networks are not appropriate for Probabilistic Neural Networks (PNN). In this paper, a new active learning method is proposed for PNN. The method was applied to several benchmark problems.

A novel methodology to determine the optimum number of centers and the network parameters simultaneously based on Particle Swarm Optimization (PSO) algorithm with matrix encoding is proposed in this paper. For tackling structure matching problem, a random structure updating rule is employed for determining the current structure at each epoch. The effectiveness of the method is illustrated through the nonlinear system identification problem.

Self-Organizing Maps (SOM) is a powerful tool for clustering and discovering patterns in data. Competitive learning in the SOM training process focuses on finding a neuron that is most similar to that of an input vector. Since an update of a neuron only benefits part of the feature map, it can be thought of as a local optimization problem. The ability to move away from a local optimization model into a global optimization model requires the use of game theory techniques to analyze overall

quality

of the SOM. A new algorithm GTSOM is introduced to take into account cluster quality measurements and dynamically modify learning rates to ensure improved quality through successive iterations.

Due to the excellent performance of the HMM (Hidden Markov Model) in pattern recognition, it has been widely used in voice recognition, text recognition. In recent years, the HMM has also been applied to the intrusion detection. The intrusion detection method based on the HMM is more efficient than other methods. The HMM based intrusion detection method is composed by two processes: one is the HMM process; the other is the hard decision process, which is based on the profile database. Because of the dynamical behavior of system calls, the hard decision process based on the profile database cannot be efficient to detect novel intrusions. On the other hand, the profile database will consume many computer resources. For these reasons, the combined detection method was provided in this paper. The neural network is a kind of artificial intelligence tools and is combined with the HMM to make soft decision. In the implementation, radial basis function model is used, because of its simplicity and its flexibility to adapt pattern changes. With the soft decision based on the neural network, the robustness and accurate rate of detection model network, the robustness and accurate rate of detection model are greatly improved. The efficiency of this method has been evaluated by the data set originated from Hunan Technology University.

This paper describes two methods on how to generate different neural networks in an ensemble. One is based on negative correlation learning. The other is based on cross-validation with negative correlation learning, i.e., bagging with negative correlation learning. In negative correlation learning, all individual networks are trained simultaneously on the same training set. In bagging with negative correlation learning, different individual networks are trained on the different sampled data set with replacement from the training set. The performance and correct response sets are compared between two learning methods. The purpose of this paper is to find how to design more effective neural network ensembles.

New algorithm was devised to speed up the convergence of backpropagation networks and the Bayesian Information Criterion was presented to obtain the optimal network structure. Nonlinear neural network problem can be partitioned into the nonlinear part in the weights of the hidden layers and the linear part in the weights of the output layer. We proposed the algorithm for speeding up the convergence by employing the conjugate gradient method for the nonlinear part and the Kalman filter algorithm for the linear part. From simulation experiments with daily data on the stock prices in the Thai market, it was found that the algorithm and the Bayesian Information Criterion could perform satisfactorily.

Webpage categorization has turned out to be an important topic in recent years. In a webpage, text is usually the main content, so that

auto text categorization

(ATC) becomes the key technique to such a task. For Chinese text categorization as well as Chinese webpage categorization, one of the basic and urgent problems is the construction of a good benchmark corpus. In this study, a machine learning approach is presented to refine a corpus for Chinese webpage categorization, where the AdaBoost algorithm is adopted to identify outliers in the corpus. The standard

k nearest neighbor

(kNN) algorithm under a

vector space model

(VSM) is adopted to construct a webpage categorization system. Simulation results as well as manual investigation of the identified outliers reveal that the presented method works well.

Kernel self-organizing map has been recently studied by Fyfe and his colleagues [1]. This paper investigates the use of a novel bio-kernel function for the kernel self-organizing map. For verification, the application of the proposed new kernel self-organizing map to HIV drug resistance classification using mutation patterns in protease sequences is presented. The original self-organizing map together with the distributed encoding method was compared. It has been found that the use of the kernel self-organizing map with the novel bio-kernel function leads to better classification and faster convergence rate ...

In this paper a new learning algorithm, Lever Training Machine (LTM), is presented for binary classification. LTM is a supervised learning algorithm and its main idea is inspired from a physics principle: Lever Principle. Figuratively, LTM involves rolling a hyper-plane around the convex hull of the target training set, and using the equilibrium position of the hyper-plane to define a decision surfaces. In theory, the optimal goal of LTM is to maximize the correct rejection rate. If the distribution of target set is convex, a set of such decision surfaces can be trained for exact discrimination without false alarm. Two mathematic experiments and the practical application of face detection confirm that LTM is an effective learning algorithm.

Based on conventional natural gradient algorithm (NGA) and equivariant adaptive separation via independence algorithm (EASI), a novel sign algorithm for on-line blind separation of independent sources is presented. A sign operator for the adaptation of the separation model is obtained from the derivation of a generalized dynamic separation model. A variable step-size sign algorithm rooted in NGA is also derived to better match the dynamics of the input signals and unmixing matrix. The proposed algorithms are appealing in practice due to their computational simplicity. Experimental results verify the superior convergence performance over conventional NGA and EASI algorithm in both stationary and non-stationary environments.

Linear Discriminant Analysis (LDA) is one of the most popular linear projection techniques for feature extraction. The major drawback of this method is that it may encounter the small sample size problem in practice. In this paper, we present a novel LDA approach for high-dimensional data. Instead of direct dimension reduction using PCA as the first step, the high-dimensional data are mapped into a relatively lower dimensional similarity space, and then the LDA technique is applied. The preliminary experimental results on the ORL face database verify the effectiveness of the proposed approach.

To solve the bottleneck of memory in current prediction of protein secondary structure program, a chip training algorithm for a Distributed Neural Networks based on multi-agents is proposed in this paper. This algorithm evolves the global optimum by competition from a group of neural network agents by processing different groups of sample chips. The experimental results demonstrate that this method can effectively improve the convergent speed, has good expansibility, and can be applied to the prediction of protein secondary structure of middle and large size of amino-acid sequence.

The problem of construction of smoothing curve is actually regression problem. How to use SVR to solve the problem of curve smoothing reconstruction in reverse engineering is discussed in this paper. A modified support vector regression model is proposed. Numerical result shows that the smoothness of curves fitted by modified method is better than by the standard SVR, when there are some bad measure points in the data.

Support vector regression (SVR) is a powerful tool to solve regression problem, this paper proposes a fast Sequential Minimal Optimization (SMO) algorithm for training support vector regression (SVR), firstly gives a analytical solution to the size two quadratic programming (QP) problem, then proposes a new heuristic method to select the working set which leads to algorithm’s faster convergence. The simulation results indicate that the proposed SMO algorithm can reduce the training time of SVR, and the performance of proposed SMO algorithm is better than that of original SMO algorithm.

In most of the clustering algorithms the number of clusters must be given in advance. However it’s hard to do so without prior knowledge. The RPCL algorithm solves the problem by delearning the rival(the 2nd winner) every step, but its performance is too sensitive to the delearning rate. Moreover, when the clusters are not well separated, RPCL’s performance is poor. In this paper We propose a RPFCL algorithm by associating a Fuzzy Inference System to the RPCL algorithm to tune the delearning rate. Experimental results show that RPFCL outperforms RPCL both in clustering speed and in achieving correct number of clusters.

For improving coding efficiency, a new predictive vector quantization (VQ) method was proposed in this paper. Two codebooks with different dimensionalities and different size were employed in our algorithm. The defined blocks are first classified based on variance. For smooth areas, the current processing vectors are sampled into even column vectors and odd column vectors. The even column vectors are encoded with the lower-dimensional and smaller size codebook. The odd ones are predicted using the decoded pixels from intra-blocks and inter-blocks at the decoder. For edge areas, the current processing vectors are encoded with traditional codebook to maintain the image quality. An efficient method for codebook design was also presented to improve the quality of the resulted codebook. The experimental comparisons with the other methods show good performance of our algorithm.

In this paper, we propose an improved fuzzy RBF network which dynamically adjusts the rate of learning by applying the Delta-bar-Delta algorithm in order to improve the learning performance of fuzzy RBF networks. The proposed learning algorithm, which combines the fuzzy C-Means algorithm with the generalized delta learning method, improves its learning performance by dynamically adjusting the rate of learning. The adjustment of learning rate is achieved by self-generating middle-layered nodes and applying the Delta-bar-Delta algorithm to the generalized delta learning method for the learning of middle and output layers. To evaluate the learning performance of the proposed RBF network, we used 40 identifiers extracted from a container image as the training data. Our experimental results show that the proposed method consumes less training time and improves the convergence of learning, compared to the conventional ART2-based RBF network and fuzzy RBF network.

In this paper, we propose a universal approach to study dynamical behaviors of various neural networks with time-varying delays. A universal model is proposed, which includes most of the existing models as special cases. An effective approach, which was first proposed in [1] , to investigate global stability is given, too. It is pointed out that the approach proposed in the paper [1] applies to the systems with time-varying delays, too.

How crucial is the long-distance connections in small-world networks produced by the semi-random SW strategy? In this paper, we attempted to investigate some related questions by constructing a semi-random small-world network through only randomly adding ’long-range lattice distance connections’ to a regular network. The modified network model is compared with the most used NW small-world network. It can be found that, by using the new modified small-worldify algorithm, one can obtain a better clustered small-world network with similar average path length. Further more, we numerically found that, for a dynamical network on typical coupling scheme, the synchronizability of the small-world network formed by our procedure is no better than that of the small-world network formed by NW’s algorithm, although the two classes of network constructed at the same constructing prices and having similar average path length. These results further confirmed that, the random coupling in some sense the best candidate for such nonlocal coupling in the semi-random strategy. Main results are confirmed by extensive numerical simulations.

In this paper, a novel mathematical model of neuron-Double Synaptic Weight Neuron (DSWN)

1

is presented. The DSWN can simulate many kinds of neuron architectures, including Radial-Basis-Function (RBF), Hyper Sausage and Hyper Ellipsoid models, etc. Moreover, this new model has been implemented in the new CASSANN-II neurocomputer that can be used to form various types of neural networks with multiple mathematical models of neurons. The flexibility of the DSWN has also been described in constructing neural networks. Based on the theory of Biomimetic Pattern Recognition (BPR) and high-dimensional space covering, a recognition system of omni directionally oriented rigid objects on the horizontal surface and a face recognition system had been implemented on CASSANN-II neurocomputer. In these two special cases, the result showed DSWN neural network had great potential in pattern recognition.

In order to explore the search mechanism of chaotic neural network(CNN), this paper first investigates the time evolutions of four chaotic noise models, namely Logistic map, Circle map, Henon map, and a Special Two-Dimension (2-D) Discrete Chaotic System. Second, based on the CNN proposed by Y. He, we obtain three alternate CNN through replacing the chaotic noise source (Logistic map) with Circle map, Henon map, and a Special 2-D Discrete Chaotic System. Third, We apply all of them to TSP with 4-city and TSP with 10-city, respectively. The time evolutions of energy functions and outputs of typical neurons for each model are obtained in terms of TSP with 4-city. The rate of global optimization(GM) for TSP with 10-city are shown in tables by changing chaotic noise scaling parameter

γ

and decreasing speed parameter

β

. Finally, the features and effectiveness of four models are discussed and evaluated according to the simulation results. We confirm that the chaotic noise with the symmetry structure property of reverse bifurcation is necessary for chaotic neural network to search efficiently, and the performance of the CNN may depend on the nature of the chaotic noise.

This paper presents a novel learning model in qubit neuron according to quantum circuit and describes the influence to learning with gradient descent by changing the number of neurons. The first approach is to reduce the number of neurons in the output layer for the conventional technique. The second is to present a novel model, which has a 3-qubit neuron including a work qubit in the input layer. For the number of neurons in the output layer, the convergence rate and the average iteration for learning are examined. Experimental results are presented in order to show that the present method is effective in the convergence rate and the average iteration for learning.

The min-max modular neural network with Gaussian zero-crossing function (M

3

-GZC) has locally tuned response characteristic and emergent incremental learning ability, but it suffers from quadratic complexity in storage space and response time. Redundant Sample pruning and redundant structure pruning can be considered to overcome these weaknesses. This paper aims at the latter; it analyzes the properties of receptive field in M

3

-GZC network, and then proposes a strategy for pruning redundant modules. Experiments on both structure pruning and integrated with sample pruning are performed. The results show that our algorithm reduces both the size of the network and the response time notably while not changing the decision boundaries.

A general procedure for combining binary classifiers for multiclass classification problems with one-against-one decomposition policy is presented in this paper. Two existing schemes, namely the min-max combination and the most-winning combination, may be regarded as its two special cases. We show that the accuracy of the combination procedure will increase and time complexity will decrease as its main parameter increases under a proposed selection algorithm. The experiments verify our main results, and our theoretical analysis gives a valuable criterion for choosing different schemes of combining binary classifiers.

Humans use many different kinds of information from different sensory organs in motion tasks. It is important in human sensing to extract useful information and effectively use the multiple kinds of information. From the viewpoint of a computational theory, we approach the integration mechanism of human sensory and motor information. In this study, the modular structure of auto-encoder is introduced to extract the intrinsic properties about a recognized object that are contained commonly in multiple kind of information. After the learning, the relaxation method using the learned model can solve the transformation between the integrated kinds of information. This model was applied to the problem how a locomotive robot decides a leg’s height to climb over an obstacle from the visual information.

Learning problem has three distinct phases, that is, model representation, learning criterion (target function) and implementation algorithm. This paper focuses on the close relation between the selection of learning criterion for committee machine and network approximation and competitive adaptation. By minimizing the KL deviation between posterior distributions, we give a general posterior modular architecture and the corresponding learning criterion form, which reflects remarkable adaptation and scalability. Besides this, we point out, from the generalized KL deviation defined on finite measure manifold in information geometry theory, that the proposed learning criterion reduces to so-called Mahalanobis deviation of which ordinary mean square error approximation is a special case, when each module is assumed Gaussian.

The radial basis function (RBF) neural networks have been widely used for approximation and learning due to its structural simplicity. However, there exist two difficulties in using traditional RBF networks: How to select the optimal number of intermediate layer nodes and centers of these nodes? This paper proposes a novel ART2/RBF hybrid neural networks to solve the two problems. Using the ART2 neural networks to select the optimal number of intermediate layer nodes and centers of these nodes at the same time and further get the RBF network model. Comparing with the traditional RBF networks, the ART2/RBF networks have the optimal number of intermediate layer nodes , optimal centers of these nodes and less error.

This paper deals complex number procedure neural networks and its learning algorithm. The conception and mathematic description of complex number procedure neurons are proposed based on traditional complex number neuron and procedure neuron. Feed-forward complex number neural networks model are considered. Grads-descent learning algorithm is deduced according to the supervising learning, and its learning procedure consists of two parallel procedures, the real part and imaginary part. An application example is given which show that the complex procedure neural network is suitable for signal processing problem.

Urban traffic system is a complex system in a random way, it is necessary to optimize traffic control signals to cope with so many urban traffic problems. A multi-layer chaotic neural networks involving feedback (ML-CNN) was developed based on Hopfield networks and chaos theory, it was effectively used in dealing with the optimization of urban traffic signal timing. Also an energy function on the network and an equation on the average delay per vehicle for optimal computation were developed. Simulation research was carried out at the intersection in Jiangmen city in China, and which indicates that urban traffic signal timing’s optimization by using ML-CNN could reduce 25.1% of the average delay per vehicle at intersection by using the conventional timing methods. The ML-CNN could also be used in other fields.

The problem of direct adaptive neural control for a class of nonlinear systems with an unknown gain sign and nonlinear uncertainty is discussed in this paper. Based on the principle of sliding mode control and the approximation capability of multilayer neural networks (MNNs), and using Nussbaum-type function, a novel design scheme of direct adaptive neural control is proposed. By adopting the adaptive compensation term of the upper bound function of the sum of residual and approximation error, the closed-loop control system is shown to be globally stable, with tracking error converging to zero. Simulation results show the effectiveness of the proposed approach.

The performance of Self Organizing Map (SOM) is always influenced by learn methods. The resultant quality of the topological formation of the SOM is also highly dependent onto the learning rate and the neighborhood function. In literature, there are plenty of studies to find a proper method to improve the quality of SOM. However, a new term “stiffness factor” has been proposed and was used in SOM training in this paper. The effect of the stiffness factor has also been tested with a real-world problem and got positive influence.

This paper proposed a novel self-adaptive wavelet network model for Regression Analysis. The structure of this network is distinguished from those of the present models. It has four layers. This model not only can overcome the structural redundancy which the present wavelet network cannot do, but also can solve the complicated problems respectively. Thus, generalization performance has been greatly improved; moreover, rapid learning can be realized. Some experiments on regression analysis are presented for illustration. Compared with the existing results, the model reaches a hundredfold improvement in speed and its generalization performance has been greatly improved.

Reliable prediction of sales can improve the quality of business strategy. In this research, fuzzy logic and artificial neural network are integrated into the fuzzy back-propagation network (FBPN) for printed circuit board industry. The fuzzy back propagation network is constructed to incorporate production-control expert judgments in enhancing the model’s performance. Parameters chosen as inputs to the FBPN are no longer considered as of equal importance, but some sales managers and production control experts are requested to express their opinions about the importance of each input parameter in predicting the sales with linguistic terms, which can be converted into pre-specified fuzzy numbers, aggregated and corresponding input parameters when fed into the FBPN. The proposed system is evaluated through the real life data provided by a printed circuit board company. Model evaluation results for research indicate that the Fuzzy back-propagation outperforms the other three different forecasting models in MAPE.

This paper presents an evolutionary artificial neural network (EANN) to the prediction of the BF hot metal silicon content. The pareto differential evolution (PDE) algorithm is used to optimize the connection weights and the network’s architecture (number of hidden nodes) simultaneously to improve the prediction precision. The application results show that the prediction of hot metal silicon content is successful. Data, used in this paper, were collected from No.1 BF at Laiwu Iron and Steel Group Co..

This paper presents a simulation of a biological olfactory neural system with a KIII set, which is a high-dimensional chaotic neural network. The KIII set differs from conventional artificial neural networks by use of chaotic attractors for memory locations that are accessed by, chaotic trajectories. It was designed to simulate the patterns of action potentials and EEG waveforms observed in electrophysioloical experiments, and has proved its utility as a model for biological intelligence in pattern classification. An application on recognition of handwritten numerals is presented here, in which the classification performance of the KIII network under different noise levels was investigated.

This paper incorporates robustness into neural network modeling and proposes a novel two-phase robustness analysis approach for determining the optimal feedforward neural network (FNN) architecture in terms of Hellinger distance of probability density function (PDF) of error distribution. The proposed approach is illustrated with an example in this paper.

The problem of stochastic robust stability analysis for Markovian jump neural networks with time delay has been investigated via stochastic stability theory. The neural network under consideration is subject to norm-bounded stochastic nonlinear perturbation. The sufficient conditions for robust stability of Markovian jumping stochastic neural networks with time delay have been developed for all admissible perturbations. All the results are given in terms of linear matrix inequalities.

With the changing of the stimulus frequency, there are a lot of firing dynamics behaviors of interspike intervals (ISIs), such as quasi-periodic, bursting, period-chaotic, chaotic, periodic and the bifurcations of the chaotic attractor appear alternatively in Hodgkin-Huxley (H-H) neuron model. The chaotic behavior is realized over a wide range of frequency and is visualized by using ISIs, and many kinds of abrupt undergoing changes of the ISIs are observed in deferent frequency regions, such as boundary crisis, interior crisis and merging crisis displaying alternately along with the changes changes of external signal frequency, too. And there are many periodic windows and fractal structures in ISIs dynamics behaviors. The saddle node bifurcation resulted collapses of chaos to period-12 orbit in dynamics of ISIs is identified.

An algorithm of Dynamic Decay Adjustment Radial Basis Function (RBF-DDA) neural networks is presented. It can adaptively get the number of the hidden layer nodes and the center values of data. It resolve the problem of deciding RBF parameters randomly and generalization ability of RBF is improved. When is applied to the system of image pattern recognition, the experimental results show that the recognition rate of the improved RBF neural network still achieves 97.4% even under stronger disturbance. It verifies the good performance of improved algorithm.

The problem of global exponential stability of cellular neural networks with time-varying delays is discussed by employing a method of delay differential inequality. A simple sufficient condition is given for global exponential stability of the cellular neural networks with time-varying delays. The result obtained here improves some results in the previous works.

Stochastic resonance (SR) effect has been discovered in non-dynamical threshold systems such as sensory systems. This paper presents a network simulating basic structure of a sensory system to study SR. The neuronal network consists of two layers of the Hodgkin-Huxley (HH) neurons. Compared with single HH model, subthreshold stimulating signals do not modulate output signal-noise ratio, thus a fixed level of noise from circumstance can induce SR for the various stimulating signals. Numeric experimental results also show that noises do not always deteriorate the capability of the detection of suprathreshold input signals.

A spiking neural network (SNN) model trained with spiking-timing-dependent-plasticity (STDP) is proposed to perform a 2D co-ordinate transformation of the polar representation of an arm position to a Cartesian representation in order to create a virtual image map of a haptic input. The position of the haptic input is used to train the SNN using STDP such that after learning the SNN can perform the co-ordinate transformation to generate a representation of the haptic input with the same co-ordinates as a visual image. This principle can be applied to complex co-ordinate transformations in artificial intelligent systems to process biological stimuli.

This work presents a model of minimal time-continuous target-cell specific use-dependent short-term synaptic plasticity (STP) observed in the pyramidal cells that can account for both short-term depression and facilitation. In general it provides a concise and portable description that is useful for predicting synaptic responses to more complex patterns of simulation, for studies relating to circuit dynamics and for equating dynamic properties across different synaptic pathways between or within preparations. This model allows computation of postsynaptic responses by either facilitation or depression in the synapse thus exhibiting characteristics of dynamic synapses as that found during short-term synaptic plasticity, for any arbitrary pre-synaptic spike train in the presence of realistic background synaptic noise. Thus it allows us to see specific effect of the spike train on a neuronal lattice both small-scale and large-scale, so as to reveal the short-term plastic behavior in neurons.

To realize mutual association function, we propose a hippoca- mpus-neocortex model with multi-layered chaotic neural network (

MCNN

). The model is based on Ito

etal

.’s hippocampus-cortex model (2000), which is able to recall temporal patterns, and form long-term memory. The

MCNN

consists of plural chaotic neural networks (

CNN

s), whose each

CNN

layer is a classical association model proposed by Aihara.

MCNN

realizes mutual association using incremental and relational learning between layers, and it is introduced into

CA

3 of hippocampus. This chaotic hippocampus-neocortex model intends to retrieve relative multiple time series patterns which are stored (experienced) before when one common pattern is represented. Computer simulations verified the efficiency of proposed model.

In most neural network models, neurons are viewed as the only computational units, while the synapses are treated as passive scalar parameters (weights). It has, however, long been recognized that biological synapses can exhibit rich temporal dynamics. These dynamics may have important consequences for computing and learning in biological neural systems. This paper proposes a novel stochastic model of single neuron with synaptic dynamics, which is characterized by several stochastic differential equations. From this model, we obtain the evolution equation of their density function. Furthermore, we give an approach to cut the evolution equation of the high dimensional function down to the evolution equation of one dimension function.

Spiking Neural Networks (SNNs) use inter-spike time coding to process input data. In this paper, a new learning algorithm for SNNs that uses the inter-spike times within a spike train is introduced. The learning algorithm utilizes the spatio-temporal pattern produced by the spike train input mapping unit and adjusts synaptic weights during learning. The approach was applied to classification problems.

Several new conditions for exponential convergence of DNN were proposed in this paper. These conditions guarantee the existence and uniqueness of equilibrium of DNN with certain different activation functions. To demonstrate the differences and features of the new criteria, some remarks are presented.

This paper proposes a neural network for saddle point problems (SPP) by an approximation approach. It first proves both the existence and the convergence property of approximate solutions, and then shows that the proposed network is globally exponentially stable and the solution of (SPP) is approximated. Simulation results are given to demonstrate further the effectiveness of the proposed network.

Implementing of intersection operation and union operation in fuzzy reasoning is explored by three Integrate-And-Fire (IAF) neurons, with two neurons as inputs and the other one as output. We prove that if parameter values of the neurons are set appropriately for intersection operation, firing rate of the output neuron is equal to or is lower than the lower one of two input neurons. We also prove that if parameter values of the neurons are set appropriately for union operation, the firing rate of the output neuron is equal to or is higher than the higher one of the two input neurons. The characteristic of intersection operation and union operation implemented by IAF neurons is discussed.

Spike trains are treated as exact time dependent stepwise functions called response functions. Five variables defined at sequential moments with equal interval are introduced to characterize features of response function; and these features can reflect temporal patterns of spike train. These variables have obvious geometric meaning in expressing the response and reasonable coding meaning in describing spike train since the well known ’firing rate’ is among them. The dissimilarity or distance between spike trains can be simply defined by means of these variables. The reconstruction of spike train with these variables demonstrates that information carried by spikes is preserved. If spikes of neuron ensemble are taken as a spatial sequence in each time bins, spatial patterns of spikes can also be quantified with a group of variables similar to temporal ones.

In this paper we propose a new stochastic nonlinear evolution model that is used to describe activity of neuronal population, we obtain dynamic image of evolution on the average number density in three-dimensioned space along with time, which is used to describe neural synchronization motion. This paper takes into account not only the impact of noise in phase dynamics but also the impact of noise in amplitude dynamics. We analyze how the initial condition and intensity of noise impact on the dynamic evolution of neural coding when the neurons spontaneously interact. The numerical result indicates that the noise acting on the amplitude influences the width of number density distributing around the limit circle of amplitude and the peak value of average number density, but the change of noise intensity cannot make the amplitude to participate in the coding of neural population. The numerical results also indicate that noise acting on the amplitude does not affect phase dynamics.

Till now, the problem of neural coding remains a puzzle. The intrinsic information carried in irregular neural spikes sequence is not known yet. But solution of the problem will have direct influence on the study of neural information mechanism. In this paper, coding mechanism of the neural spike sequence, which is caused by input stimuli of various frequencies, is investigated based on analysis of H-H equation with the method of nonlinear dynamics. The signals of external stimuli – those continuously varying physical or chemical signals – are transformed into frequency signals of potential in many sense organs of biological system, and then the frequency signals are transformed into irregular neural coding. This paper analyzes in detail the neuron response of stimuli with various periods and finds the possible rule of coding.

We study the synchronization phenomena in a pair of Hindmarsh-Rose (HR) neurons with chemical coupling. We find that excitatory synaptic coupling pushes two neurons towards antisynchrony, and weak or moderate inhibitory synaptic coupling pushes two neurons towards antisynchrony too, but sufficiently strong inhibitory synaptic coupling pushes two neurons towards synchronized periodic oscillations without spikes. And synchronization patterns can’t be changed even if the intrinsic frequency of individual cell is changed by modulating external input current. Investigating the effect of synapse on ISIs bifurcation structures shows that whether excitatory synapse or inhibitory synapse, both remarkably influence ISIs structures. That is, the chemical coupling between neurons wholly distorts the neuronal information.

This paper deals with the estimation of the linear and the nonlinear quantile regressions using the idea of support vector machine. Accordingly, the optimization problem is transformed into the Lagrangian dual problem, which is easier to solve. In particular, for the nonlinear quantile regression the idea of kernel function is introduced, which allows us to perform operations in the input space rather than the high dimensional feature space. Experimental results are then presented which illustrate the performance of the proposed method.

A doubly regularized likelihood estimating procedure is introduced for the heteroscedastic censored regression. The proposed procedure provides the estimates of both the conditional mean and the variance of the response variables, which are obtained by two stepwise iterative fashion. The generalized cross validation function and the generalized approximate cross validation function are used alternately to estimate tuning parameters in each step. Experimental results are then presented which indicate the performance of the proposed estimating procedure.

The Support vector machines derive the class decision hyper planes from a few, selected prototypes, the support vectors (SVs) according to the principle of structure risk minimization, so they have good generalization ability. We proposed a new prototype selection method based on support vectors for nearest neighbor rules. It selects prototypes only from support vectors. During classification, for unknown example, it can be classified into the same class as the nearest neighbor in feature space among all the prototypes. Computational results show that our method can obtain higher reduction rate and accuracy than popular condensing or editing instance reduction method.

The kernel minimum squared error estimation (KMSE) model can be viewed as a general framework that includes kernel Fisher discriminant analysis (KFDA), least squares support vector machine (LS-SVM), and kernel ridge regression (KRR) as its particular cases. For continuous real output the equivalence of KMSE and LS-SVM is shown in this paper. We apply standard methods for computing prediction intervals in nonlinear regression to KMSE model. The simulation results show that LS-SVM has better performance in terms of the prediction intervals and mean squared error(MSE). The experiment on a real date set indicates that KMSE compares favorably with other method.

The type of kernel function has a great important influence on the performance of support vector machines (SVMs); however, there is no theoretical guidance to choose a good kernel. To solve classification problem, Amari presented a method of modifying kernel based on information geometry theory. In the paper, we first review the classical formulation of regression problem, then propose an approach to constructing the kernel function in support vector regression machines from information-geometrical viewpoint, and point out its difference with the method that Amari used in support vector classification machines. Finally some simulation results show the effectiveness of the proposed method.

In recent years, support vector machines (SVMs) have become a popular tool for pattern recognition and machine learning. Training a SVM involves solving a constrained quadratic programming problem, which requires large memory and enormous amounts of training time for large-scale problems. In contrast, the SVM decision function is fully determined by a small subset of the training data, called support vectors. Therefore, it is desirable to remove from the training set the data that is irrelevant to the final decision function. In this paper we propose two new methods that select a subset of data for SVM training. Using real-world datasets, we compare the effectiveness of the proposed data selection strategies in terms of their ability to reduce the training set size while maintaining the generalization performance of the resulting SVM classifiers. Our experimental results show that a significant amount of training data can be removed by our proposed methods without degrading the performance of the resulting SVM classifiers.

Regularized Least-Squares Classification (RLSC) can be regarded as a kind of 2 layers neural network using regularized square loss function and kernel trick. Poggio and Smale recently reformulated it in the framework of the mathematical foundations of learning and called it a key algorithm of learning theory. The generalization performance of RLSC depends heavily on the setting of its kernel and hyper parameters. Therefore we presented a novel two-step approach for optimal parameters selection: firstly the optimal kernel parameters are selected by maximizing kernel target alignment, and then the optimal hyper-parameter is determined via minimizing RLSC’s leave-one-out bound. Compared with traditional grid search, our method needs no independent validation set. We worked on IDA’s benchmark datasets using Gaussian kernel, the results demonstrate that our method is feasible and time efficient.

A new strategy of modelling of chaotic systems is presented. First, more information is acquired utilizing the reconstructed embedding phase space. Then, based on the Recurrent Least Squares Support Vector Machines (RLS-SVM), modelling of the chaotic system is realized. We use the power spectrum and dynamic invariants involving the Lyapunov exponents and the correlation dimension as criterions, and then apply our method to the Chua‘s circuit time series. The simulation of dynamic invariants between the origin and generated time series shows that the proposed method can capture the dynamics of the chaotic time series effectively.

Based on the wavelet decomposition and reproducing kernel Hilbert space (RKHS), a novel notion of least squares wavelet support vector machine (LS-WSVM) with universal reproducing wavelet kernels is proposed for approximating arbitrary nonlinear functions. The good reproducing property of wavelet kernel function enhances the generalization ability of LS-WSVM method and some experimental results are presented to illustrate the feasibility of the proposed method.

A new Fuzzy Support Vector Machines (

λ

—FSVMs) based on

λ

—cut is proposed in this paper. The proposed learning machines combine the membership of fuzzy set with support vector machines. The

λ

—cut set is introduced to distinguish the training samples set in term of the importance of the data. The more important sets are selected as new training sets to construct the fuzzy support vector machines. The benchmark two-class problems and multi-class problems datasets are used to test the effectiveness and validness of

λ

—FSVMs. The experiment results indicate that

λ

—FSVMs not only has higher precision but also solves the overfitting problem of the support vector machines more effectively.

Kernels are employed in Support Vector Machines (SVM) to map the nonlinear model into a higher dimensional feature space where the linear learning is adopted. The characteristic of kernels has a great impact on learning and predictive results of SVM. Good characteristic for fitting may not represents good characteristic for generalization. After the research on two kinds of typical kernels—global kernel (polynomial kernel) and local kernel (RBF kernel), a new kind of SVM modeling method based on mixtures of kernels is proposed. Through the implementation in Lithopone calcination process, it demonstrates the good performance of the proposed method compared to single kernel.

Support Vector Machine (SVM) has been applied in many classification systems successfully. However, it is restricted to work well on the small sample sets. This paper presents a novel parallel reduced support vector machine. The proposed algorithm consists of three parts: firstly dividing the training samples into some grids; then training sample subset through density clustering; and finally classifying the samples. After clustering the positive samples and negative samples, this algorithm picks out such samples that locate on the edge of clusters as reduced sample subset. Then, we sum up these reduced sample subsets as reduced sample set. These reduced samples are then used to find the support vectors and the optimal classifying hyperplane by support vector machine. Additionally, it also improves classification precision by reducing the percentage of counterexamples in kernel object

ε

-area. Experiment results show that not only efficiency but also classification precision are improved, compared with other algorithms.

Support vector machines (SVMs) have been successfully used in solving nonlinear regression and times series problems. However, the application of SVMs for reliability prediction is not widely explored. Traditionally, the recurrent neural networks are trained by the back-propagation algorithms. In the study, SVM learning algorithms are applied to the recurrent neural networks to predict system reliability. In addition, the parameter selection of SVM model is provided by Genetic Algorithms (GAs). A numerical example in an existing literature is used to compare the prediction performance. Empirical results indicate that the proposed model performs better than the other existing approaches.

A modified sequential minimal optimization (SMO) algorithm for support vector machine (SVM) regression is proposed based on Shevade’s SMO-1 algorithm. The main improvement is that a modified heuristics method is used in this modified SMO algorithm to choose the first Lagrange multiplier when optimizing the Lagrange multipliers corresponding to the non-boundary examples. To illustrate the validity of the proposed modified SMO algorithm, a benchmark dataset and a practical application in predicting the melt index of high-pressure low-density polyethylene (HP-LDPE) are used; the results demonstrate that this modified SMO algorithm is faster in most cases with the same parameters setting and more likely to obtain the better generalization performance than Shevade’s SMO-1 algorithm.

This paper proposes a method of automatic gait recognition using Fourier descriptors and independent component analysis (ICA) for the purpose of human identification at a distance. Firstly, a simple background generation algorithm is introduced to subtract the moving figures accurately and to obtain binary human silhouettes. Secondly, these silhouettes are described with Fourier descriptors and converted into associated one-dimension signals. Then ICA is applied to get the independent components of the signals. For reducing the computational cost, a fast and robust fixed-point algorithm for calculating ICs is adopted and a criterion how to select ICs is put forward. Lastly, the nearest neighbor (NN), support vector machine (SVM) and backpropagation neural network (BPNN) classifiers are chosen for recognition and this method is tested on the small UMD gait database and the NLPR gait database. Experimental results show that our method has encouraging recognition accuracy.

Ordinal regression is complementary to the standard machine learning tasks of classification and metric regression which goal is to predict variables of ordinal scale. However, every input must be exactly assigned to one of these classes without any uncertainty in standard ordinal regression models. Based on structural risk minimization (SRM) principle, a new support vector learning technique for ordinal regression is proposed, which is able to deal with training data with uncertainty. Firstly, the meaning of the uncertainty is defined. Based on this meaning of uncertainty, two algorithms have been derived. This technique extends the application horizon of ordinal regression greatly. Moreover, the problem about early warning of food security in China is solved by our algorithm.

This paper presents briefly an incremental learning method based on SVM for online sketchy shape recognition. It can collect all classified results corrected by user and select some important samples as the retraining data according to their distance to the hyper-plane of the SVM-classifier. The classifier can then do incremental learning quickly on the newly added samples, and the retrained classifier can be adaptive to the user’s drawing styles. Experiment shows the effectiveness of the proposed method.

This short paper proposes a face recognition scheme, wherein features called eigenspectra are extracted successively by the fast Fourier transform (FFT) and the principle component analysis (PCA) and classification results are obtained by a classifier called kernel-based nonlinear representor (KNR). Its effectiveness is shown by experimental results on the Olivetti Research Laboratory (ORL) face database.

In this paper, a neural network model and its associate learning rule are developed for sequential blind extraction in the case that the number of observable mixed signals is less than the one of sources. This approach is also suitable for the case in which the mixed matrix is nonsingular. Using this approach, all separable sources can be extracted one by one. The solvability analysis of the problem is also presented, and the new solvable condition is weaker than existing solvable conditions in some literatures.

Accurate employee turnover prediction plays an important role in providing early information for unanticipated turnover. A novel classification technique, support vector machines (SVMs), has been successfully employed in many fields to deal with classification problems. However, the application of SVMs for employee voluntary turnover prediction has not been widely explored. Therefore, this investigation attempts to examine the feasibility of SVMs in predicting employee turnover. Besides, two other tradition regression models, Logistic and Probability models are used to compare the prediction accuracy with the SVM model. Subsequently, a numerical example of employee voluntary turnover data from a middle motor marketing enterprise in central Taiwan is used to compare the performance of three models. Empirical results reveal that the SVM model outperforms the logit and probit models in predicting the employee turnover based on job performance. Consequently, the SVM model is a promising alternative for predicting employee turnover in human resource management.

As feature extraction techniques, Kernel Principal Component Analysis (KPCA) and Independent Component Analysis (ICA) can both be considered as generalization of Principal Component Analysis (PCA), which has been used for palmprint recognition and gained satisfactory results [3], therefore it is natural to wonder the performances of KPCA and ICA on this issue. In this paper, palmprint recognition using the KPCA and ICA methods is developed and compared with the PCA method. Based on the experimental results, some useful conclusions are drawn, which fits into the scene for a better picture about considering these unsupervised subspace classifiers for palmprint recognition.

In order to get good hyperparameters of SVM, user needs to conduct extensive cross-validation such as leave-one-out (

LOO

) cross-validation. Alpha seeding is often used to reduce the cost of SVM training. Compared with the existing schemes of alpha seeding, a new efficient alpha seeding method is proposed. Through some examples, its good performance has been proved. Interpretation from both geometrical and mathematical view is also given.

In this paper we study location of multiple acoustic sources by blind source separation (BSS) method, which based on canonical correlation analysis (CCA). The receiving array is a sparse array. This array is composed of three separated subarrays. From the receiving data set, we can obtain the separate components by CCA. After a simple correlation, time difference can be obtained, and then compute the direction of arrival (DOA) of different acoustic sources. The coordinate of different acoustic sources can be obtained at last. The important contribution of this new location method is that it can reduce the effect of inter-sensor spacing and other factors. Simulation result confirms the validity and practicality of the proposed approach. Results of location are more accurate and stable based on this new method.

An approach for short-term load forecasting by combining self-organizing map(SOM) and support vector machine(SVM) is proposed in this paper. First, historical load data of same type are clustered using SOM, and then daily 48-point load values are vertically predicted respectively based on SVM. In clustering, factors such as date type, weather conditions and time delay are considered. In addition, influences of kernel function and SVM parameters on load forecasting are discussed and performance of SOM-SVM is compared with pure SVM. It is shown that normal smoothing technique in preprocessing is not suitable to be used in vertical forecasting. Finally, the approach is tested by data from EUNITE network, and results show that the approach runs with high speed and good accuracy.

Many problems in pattern classifications involve some form of dimensionality reduction. ISOMAP is a representative nonlinear dimensionality reduction algorithm, which can discover low dimensional manifolds from high dimensional data. To speed ISOMAP and decrease the dependency to the neighborhood size, we propose an improved algorithm. It can automatically select a proper neighborhood size and an appropriate landmark set according to a stress function. A multi-class classifier with high efficiency is obtained through combining the improved ISOMAP with SVM. Experiments show that the classifier presented is effective in fingerprint classifications.

A variable structure control (VSC) scheme for linear MIMO systems based on support vector machine (SVM) is developed. By analyzing the characters of linear MIMO system, a VSC scheme based on Exponent Reaching Law is adopted to track desired trajectory. Then one input of the system is trained as the output of SVM, while sliding mode function, differences and other inputs of the system are trained as the inputs of SVM. So one VSC input of the black-box system could be obtained directly by trained SVM after other inputs of the system are selected manually, and recognition of system parameters is avoided. A linear MIMO system is used to prove the scheme, and simulation results show that this scheme has high identification precision and quick training speed.

FastICA is now a popular algorithm for independent component analysis (ICA) based on negentropy. However the convergence of FastICA has not been comprehensively studied. This paper provides the global convergence analysis of FastICA and some practical considerations on algorithmic implementations. The exhaustive equilibria are obtained from the iteration first. Then the global convergence property is given on the 2-channel system with cubic nonlinearity function, and the results can also be generalized to the multi-channel system. In addition, two practical considerations, e.g. the convergence threshold for demixing matrix and independence restriction for sources, are evaluated and the influence on the separation solutions is illustrated respectively.

In this paper, a support vector machine (SVM) with linear kernel function based nonparametric model identification and dynamic matrix control (SVM_DMC) technique is presented. First, a step response model involving manipulated variables is obtained via system identification by SVM with linear kernel function according to random test data or manufacturing data. Second, an explicit control law of a receding horizon quadric objective is gotten through the predictive control mechanism. Final, the approach is illustrated by a simulation of a system with dead time delay. The results show that SVM_DMC technique has good performance in predictive control with good capability in keeping reference trajectory.

It is well-known that the major task of the SVM approach lies in the selection of its kernel. The quality of kernel will determine the quality of SVM classifier directly. However, the best choice of a kernel for a given problem is still an open research issue. This paper presents a novel method which learns SVM kernel by transforming it into a standard semi-definite programming (SDP) problem and then solves this SDP problem using various existing methods. Experimental results are presented to prove that SVM with the kernel learned by our proposed method outperforms that with a single common kernel in terms of generalization power.

Based on KKT condition and Lagrangian multiplier method a weighted SVM regression model and its on-line training algorithm are developed. Standard SVM regression model processes every sample equally with the same error requirement, which is not suitable in the case that different sample has different contribution to the construction of the regression model. In the new weighted model, every training sample is given a weight coefficient to reflect the difference among samples. Moreover, standard online training algorithm couldn’t remove redundant samples effectively. A new method is presented to remove the redundant samples. Simulation with a benchmark problem shows that the new algorithm can quickly and accurately approximate nonlinear and time-varying functions with less computer memory needed.

An online gradient method for BP neural networks is presented and discussed. The input training examples are permuted stochastically in each cycle of iteration. A monotonicity and a weak convergence of deterministic nature for the method are proved.

In this paper, a new constructive algorithm for wavelet neural networks (WNN) is proposed. Employing the time-frequency localization property of wavelet, the wavelet network is constructed from the low resolution to the high resolution. At each resolution, a new wavelet is initialized as a member of wavelet frames. The input weight freezing technique is used and the Levenberg-Marquardt (LM) algorithm, a quasi-Newton method, is used to train the new wavelet in the WNN. After training, the new wavelet will be added to the wavelet network if the reduction of the residual error between the desired output and WNN output is greater than a threshold. The proposed algorithm is suitable to situations when the wavelet library is very large. The simulations demonstrate the effectiveness of the proposed approach.

In 1984 Hopfield showed that the time evolution of a symmetric Hopfield neural networks are a motion in state space that seeks out minima in the energy function (i.e., equilibrium point set of Hopfield neural networks). Because high-order Hopfield neural networks have more extensive applications than Hopfield neural networks, and have been discussed on the convergence of the networks. In practice, a neural network is often subject to environmental noise. It is therefore useful and interesting to find out whether the high-order neural network system still approacher some limit set under stochastic perturbation. In this paper, we will give a number of useful bounds for the noise intensity under which the stochastic high-order neural network will approach its limit set. Our result cancels the requirement of symmetry of the connection weight matrix and includes the classic result on Hopfield neural networks, which is a special case of stochastic high-order Hopfield neural networks. In the end, A example is given to verify the effective of our results.

As a special type of Self-Organizing Maps, the Dynamic Cell Structures (DCS) network has topology-preserving adaptive learning capabilities that can, in theory, respond and learn to abstract from a much wider variety of complex data manifolds. However, the highly complex learning algorithm and non-linearity behind the dynamic learning pattern pose serious challenge to validating the prediction performance of DCS and impede its spread in control applications, safety-critical systems in particular.

In this paper, we improve the performance of DCS networks by providing confidence measures on DCS predictions. We present the validity index, an estimated confidence interval associated with each DCS output, as a reliability-like measure of the network’s prediction performance. Our experiments using artificial data and a case study on a flight control application demonstrate an effective validation scheme of DCS networks to achieve better prediction performance with quantified confidence measures.

In this study, we propose this new algorithm that generates score function in ICA (Independent Component Analysis) using entropy theory. To generate score function, estimation of probability density function about original signals are certainly necessary and density function should be differentiated. Therefore, we used kernel density estimation method in order to derive differential equation of score function by original signals. After changing the formula to convolution form to increase speed of density estimation, we used FFT algorithm which calculates convolution faster. Proposed score function generation method reduces estimation error, it is density difference of recovered signals and original signals. Also, we insert constraint which is able to information maximization using smoothing parameters. In the result of computer simulation, we estimate density function more similar to original signals compared with Extended Infomax algorithm and Fixed Point ICA in blind source separation problem and get improved performance at the SNR (Signal to Noise Ratio) between recovered signals and original signals.

A new criterion on exponential stability of the equilibrium point for a class of discrete cellular neural networks (CNNs) with delay is established by Lyapunov-Krasovskii function methods. The obtained result shows a relation between the delayed time and the corresponding parameters of the network. A numerical example is given to illustrate the efficiency of the proposed approach.

A new type of local connection neural network is proposed in this paper. There is a called K-type activation function in its hidden layer so as to have less computation compared with other local connection neural network. First the structure and algorithm of the proposed network are given. Then the function of network and its properties are analyzed theoretically. The proposed network can be used in the function approximation and modeling. Finally, numerical applications are used to verify the advantages of proposed network compared with other local connection neural networks.

In this paper, we review a visual approach and propose it for analysing computer-network activity, which is based on the use of unsupervised connectionist neural network models and does not rely on any previous knowledge of the data being analysed. The presented Intrusion Detection System (IDS) is used as a method to investigate the traffic which travels along the analysed network, detecting SNMP (Simple Network Management Protocol) anomalous traffic patterns. In this paper we have focused our attention on the study of anomalous situations generated by a MIB (Management Information Base) information transfer.

In this paper, based on the theory of nonseparable wavelet, a novel nonseparable wavelet model has been proposed. The structure of the model is distinguished from that of wavelet network (RBF structure). It is a four-layer structure, which helps overcome the structural redundancy. In the process of the training of the network, in the light of the characteristics of nonseparable wavelet, a novel method of setting the initial value of weight has been proposed. It can overcome the shortcoming of gradient descent methodology that it makes the convergence of the network slow. Some experiments with the novel model for function learning will be shown. Comparing with the present wavelet networks, BP network, the results in this paper show that the speed and generalization performance of the novel model have been greatly improved.

This study proposes a novel method that can recognize illusory contour figures by using a neural network model referenced on the mechanism of feature extraction found in a visual cortex domain. A common factor in all such illusory contour figures, such as the Kanizsa triangle is the perception of a surface occluding part of a background, i.e. illusory contours are always accompanied by illusory surfaces. In this paper, we propose a neural network model that predicts the shape of illusory surfaces based on features of the visual cortex domain. This model employs an important two-stage process of the Induced Stimuli Extraction System (ISES) and Illusory Surfaces Perception System (ISPS). The former system extracts the induced stimuli for the perception of illusory surfaces, and the latter forms the illusory surfaces from the induced stimuli. The proposed model is demonstrated on a variety of Kanizsa-type illusory contour displays. The results of the experiment shows that the proposed model is successful not only in extracting the induced stimuli for the perception of illusory contours, but also in perceiving the illusory surface figures from the induced stimuli.

In this paper, in aid of ANN ensemble, a supervised online learning strategy continuously achieves omnidirectional information accumulation for 3D object cognition from 2D view sequence. The notion of similarity is introduced to solve the paradox between information simplicity and accuracy. Images are segmented into homogeneous region for training, correspondent to distinct model views characteristic of neighboring generalization. Real-time techniques are adopted to expand knowledge until satisfactory. The insert into joint model views is only needed in case of impartibility. Simulation experiment has achieved encouraging results, and proved the approach effective and feasible.

The traditional ITS have considered the learners as a knowledge receiver. The recent development of teachable agent make it possible to provide the learner with an active role as a knowledge constructor and to take initiatives to persist in learning. In order to make an adaptive teachable agent that responds intelligently for individual learner, it should reflect the individual differences in the level of cognition and motivation, and its ongoing changes. For the purpose of developing individualized teachable agent, it is proposed to a student model based on the correlation among three dimensions: individual differences, learner responses, and learning outcome. A correlation analysis among the log data, questionnaire scores, and learning measurements was conducted. We delineated the relationships among three dimensions, learner responses (mouse-click pattern, duration & frequency at particular task, individual choice etc), individual characteristics (metacognitive awareness, self-efficacy, learning goal, and performance goal), and learning outcomes (interest and comprehension) during interacting with the teachable agent. The results suggest that certain type of learner responses or the combination of the responses would be useful indices to predict the learners’ individual characteristics and ongoing learning outcome.

The purpose of this study is to investigate the application of sample entropy (SampEn) measures to electrophysiological studies of single and dual tasking performance. The complexity of short-duration (~s) epochs of EEG data were analysed using SampEn along with the surrogate technique. Individual tasks consisted of an auditory discrimination task and two motor tasks of varying difficulty. Dual task conditions were combinations of one auditory and one motor task. EEG entropies were significantly lower in dual tasks compared to that in the single tasks. The results of this study have demonstrated that entropy measurements can be a useful alternative and nonlinear approach to analyzing short duration EEG signals on a time scale of seconds.

Chaos theory was applied to analysis of the time series of plethysmograms under various human physio-psychological conditions. It found that the largest Lyapunov exponent could be used to characterize physio-psychological status. A visual representation method based on constellation graphs was developed to indexing temporal changes in the largest Lyapunov exponent. Changes of constellation angles were found to clearly characterizing variations of physio-psychological status in a series of experiments.

In order to build a human-computer interface that is sensitive to a user’s expressed emotion, we propose a neural network based emotion estimation algorithm using heart rate variability (HRV) and galvanic skin response (GSR). In this study, a video clip method was used to elicit basic emotions from subjects while electrocardiogram (ECG) and GSR signals were measured. These signals reflect the influence of emotion on the autonomic nervous system (ANS). The extracted features that are emotion-specific characteristics from those signals are applied to an artificial neural network in order to recognize emotions from new signal collections. Results show that the proposed method is able to accurately distinguish a user’s emotion.

From the last decade, modeling of cognitive agents have drawn great attention and provide a new paradigm for addressing fundamental questions in cognitive science. In this paper, a logical model for reasoning about cognitive agent’s three attitudes

Belief

,

Capability

and

Promise

is proposed. A formalization is provided based on the modal logic to specify and analyze dependencies between the three attitudes. By adopting a set of constraints that describe how the three attitudes are related to each other, we can draw a number of properties of the model. To show the potential applications of the model, we apply the BCP model to a decision-making example in trading agent competition for supply chain management(TAC SCM). The logical model proposed here provides a rigorous semantic basis for modeling cognitive agent and reasoning about multi-agent interactions.

This paper presents a framework for expressing free-hand sketching in 3D for conceptual design input. In the framework, sketch outlines will be recognized as formal rigid shapes first. Then under a group of gestures and DFAs’(deterministic finite automata) control, the framework can express user’s free sketching intents freely. Based on this framework, we implemented a sketch-based 3D prototype system supporting conceptual designs. User can easily and rapidly create 3D objects such as hexahedron, sphere, cone, extrusion, swept body, revolved body, lofted body and their assemblies by sketching and gestures.

The block-based DCT image compression methods usually result in discontinuities called blocking artifacts at the boundaries of blocks due to the coarse quantization. A measurement of blocking artifacts based on Human Vision System (HVS) is proposed. This method separates the blocking effects from the original edges in the image by an adaptive edge detection based on local activity and luminance masking. The blocking artifacts in the non-edge area and on the edge are calculated separately. The weighted sum is regarded as the evaluation result. Simulation results show that the proposed measurement is robust for different kind of images, and has the general performance of image quality evaluation metric.

This study simulates a lexical decision task in Korean by using a feed forward neural network model with a back propagation learning rule. Reaction time is substituted by a entropy value called ‘semantic stress’. The model demonstrates frequency effect, lexical status effect and non-word legality effect, suggesting that lexical decision is made within a structure of orthographic and semantic features. The test implies that the orthographic and semantic features can be automatically applied to lexical information process.

The purpose of this study is to examine the Neuroanatomical areas related with onomatopoeia and phainomime word recognition. Using the block-designed fMRI, whole-brain images (N=11) were acquired during lexical decisions. We examined how the lexical information initiatesbrain activation during visual word recognition. The onomatopoeic word recognition activated the bilateral occipital lobes and superior mid-temporal-gyrus, whereas the phainomime words recognition activated left SMA and bilateral cerebellum as well as bilateral occipital lobes. Regions more activated for the phainomime word than onomatopoeia included left SMA and bilateral cerebellum. Regions more activated for the onomatopoeia than phainomime word included left superior and mid-temporal gyri. The word recognition for onomatopoeia plus phainomime word showed activation on bilateral middle and superior temporal gyrus, right supramarginal gyrus, left middle temporal gyrus, left middle occipital gyrus, and right occipital gyrus. This is the first fMRI research to analyze onomatopoeia and phainomime word.

This paper investigates the cooperative aspects of selective attention in which primary (or bottom-up) information is dynamically integrated by the secondary (top-down or context) information from different channels, and in which the secondary information provides a criterion of what should be many target candidates We present a computational model of selective attention that implements these cooperative behaviors. Simulation results, obtained using still and video images, are presented showing the interesting properties of the model that are not captured by only competitive aspects of selective attention.

Selective attention works throughout the whole process of vision information processing. Existing attention models concentrate on its role in feature extraction in initial stage, but ignore role of attention in other stages. In this paper, we extend attention to middle stage, especially in guiding perceptual grouping. Selective attention functions in two aspects. One is to select the most salient primitive as grouping seed. The other is to organize groups and decide their pop-out sequence. Compared with traditional attention models, our model judges primitive salience according to global properties rather than local ones. And focus of attention shifts in unit of perceptual object rather than spatial region. These two improvements boost the model’s grouping quality and more fit to high stage of vision information processing. Experiments and quantitative analysis testify our model’s good performance in certain class of images.

In this work we present the computational algorithm that combines perceptual and cognitive information during the visual search for object features. The algorithm is initially driven purely by the bottom-up information but during the recognition process it becomes more constrained by the top-down information. Furthermore, we propose a concrete model for integrating information from successive saccades and demonstrate the necessity of using two coordinate systems for measuring feature locations. During the search process, across saccades, the network uses an object-based coordinate system, while during a fixation the network uses the retinal coordinate system that is tied to the location of the fixation point. The only information that the network stores during saccadic exploration is the identity of the features on which it has fixated and their locations with respect to the object-centered system.

This paper deals with agent based decision support system for patient’s right diagnosis and treatment under emergency circumstance. The well known reinforcement learning is utilized for modeling emergency healthcare system. Also designed is a novel interpretation of Markov decision process providing clear mathematical formulation to connect reinforcement learning as well as to express integrated agent system. Computational issues are also discussed with the corresponding solution procedure.

This paper aims to tackle two fundamental problems faced by multiple object recognition systems: invariance and transformation estimation. A neural normalization approach is adopted, which allows for the subsequent incorporation of invariant features. Two new approaches are introduced: dynamic inputs (DI) and attraction force analysis (AFA). The DI concept refers to a cloud of inputs that is allowed to change its configuration in order to latch onto objects thus creating object-based reference frames. AFA is used in order to provide clouds with transformation estimations thus maximizing the efficiency with which they can latch onto objects. AFA analyzes the length and angular properties of the correspondences that are found between stored-patterns and the information conveyed by clouds. The solution provides significant invariance and useful estimations pertaining to translation, scale, rotation and combinations of these. The estimations provided are also considerably resistant to other factors such as deformation, noise, occlusion and clutter.

Although the basic sparse coding model has been quite successful at explaining the receptive fields of simple cells in V1, it ignores an important constrain: perception task. We put forward a novel sparse coding model, called task-oriented sparse coding (

TOSC

) model, combining the discriminability constrain supervised by classification task, besides the sparseness criteria. Simulation experiments are performed using real images including class of scene and class of building. The results show that

TOSC

can organize some significant receptive fields with distinct topological structure which will favor the classification task. Moreover, the coefficients of

TOSC

notablely improve the classification accuracy, from the 53.5% of pixel-based model to 86.7%, in the case of none distinct damage on the performance of reconstruction error and sparseness.

TOSC

model, complementing the feedback sparse coding model, is more consistent with biological mechanism, and shows good potential in the feature extraction for pattern classification.

This paper proposes a Gabor-based PCA method using Whiten Cosine Similarity Measure (WCSM) for Face Recognition from One training Sample per Person. Gabor wavelet representation of face images first derives desirable features, which is robust to the variations due to illumination, facial expression changes. PCA is then employed to reduce the dimensionality of the Gabor features. Whiten Cosine Similarity Measure is finally proposed for classification to integrate the virtues of the whiten translation and the cosine similarity measure. The effectiveness and robustness of the proposed method are successfully tested on CAS-PEAL dataset using one training sample per person, which contains 6609 frontal images of 1040 subjects. The performance enhancement power of the Gabor-based PCA feature and WCSM is shown in term of comparative performance against PCA feature, Mahalanobis distance and Euclidean distance. In particular, the proposed method achieves much higher accuracy than the standard Eigenface technique in our large-scale experiment.

Word sense disambiguation plays an important role in natural language processing, such as information retrieval, text summarization, machine translation etc. This paper proposes a corpus-based Chinese word sense disambiguation approach using HowNet. The method is based on the co-occurrence frequency between the relatives (such as synonym, antonymy, meronymy) of target word and each word in the context. Further, domains have been used to characterize the senses of polysemous word. To our knowledge, this is the first time a Chinese word sense disambiguation method using domain knowledge is reported. The accuracy is 73.2% at present. The experimental result shows that the method is very promising for Chinese word sense disambiguation.

Humans have the ability to flexibly adjust their information processing strategy according to situational characteristics. However, such ability has been largely overlooked in computational modeling research in high-order human cognition, particularly in learning. The present work introduces frameworks of cognitive models of human learning that take contextual factors into account. The framework assumes that human learning processes are not strictly error minimization, but optimization of knowledge. A simulation study was conducted and showed that the present framework successfully replicated observed psychological phenomena.

The method of lexical chains is the first time introduced to generate summaries from Chinese texts. The algorithm which computes lexical chains based on the HowNet knowledge database is modified to improve the performance and suit Chinese summarization. Moreover, the construction rules of lexical chains are extended, and relationship among more lexical items is used. The algorithm constructs lexical chains first, and then strong chains are identified and significant sentences are extracted from the text to generate the summary. Evaluation results show that the performance of the system has a notable improvement both in precision and recall compared to the original system.

General linear model (GLM) is a most popularly method of functional magnetic imaging (fMRI) data analysis. The key of this model is how to constitute the design-matrix to model the interesting effects better and separate noises. In this paper, the new general linear convolution model is proposed by introducing dynamic characteristic function as hemodynamic response function for the processing of the fMRI data. The method is implemented by a new dynamic function convolving with stimulus pattern as design-matrix to detect brain active signal. The efficiency of the new method is confirmed by its application into the real-fMRI data. Finally, real- fMRI tests showed that the excited areas evoked by a visual stimuli are mainly in the region of the primary visual cortex.

This paper presents a novel approach toward high precision biology species categorization which is mainly based on KNN algorithm. KNN has been successfully used in natural language processing (NLP). Our work extends the learning method for biological data. We view the DNA or RNA sequences of certain species as special natural language texts. The approach for constructing composition vectors of DNA and RNA sequences is described. A learning method based on KNN algorithm is proposed. An experimental system for biology species categorization is implemented. Forty three different bacteria organisms selected randomly from EMBL are used for evaluation purpose. And the preliminary experiments show promising results on precision.

Emerging patterns (EP) represent a class of interaction structures and have recently been proposed as a tool for data mining. Especially, EP have been applied to the production of new types of classifiers during classification in data mining. Traditional clustering and pattern mining algorithms are inadequate for handling the analysis of high dimensional gene expression data or the analysis of multi-source data based on the same variables (e.g. genes), and the experimental results are not easy to understand. In this paper, a simple scheme for using EP to improve the performance of classification procedures in multi-source data is proposed. Also, patterns that make multi-source data easy to understand are obtained as experimental results. A new method for producing EP based on observations (e.g. samples in microarray data) in the search of classification patterns and the use of detected patterns for the classification of variables in multi-source data are presented.

Combination of kernel PLS (KPLS) and kernel SVD (KSVD) with minimum-squared-error (MSE) criteria has created new machine learning methods for cancer classification and has been successfully applied to seven publicly available cancer datasets. Besides the high accuracy of the new methods, very fast training speed is also obtained because the matrix inversion in the original MSE procedure is avoided. Although the KPLS-MSE and the KSVD-MSE methods have equivalent accuracies, the KPLS achieves the same results using significantly less but more qualitative components.

Biometric based person identity authentication is gaining more and more attention. It has been proved that combining multi-biometric modalities enables to achieve better performance than single modality. This paper fused Face and fingerprint (for one identity, face and fingerprint are from the really same person) for person identity authentication, and Support Vector Machine (SVM) is adopted as the fusion strategy. Performances of three SVMs based on three different kernel functions (Polynomial, Radial Based Function and Hyperbolic Tangent) are given out and analyzed in detail. Three different protocols are defined and operated on different data sets. In order to enhance the ability to bear face with bigger pose angle, a client specific SVM classifier is brought forward. Experiment results proved that it can improve the fusion authentication accuracy, and consequently expand the allowable range of face turning degree to some extend in fusion system also.

In this paper, a new multi-modality medical image fusion algorithm based on pulse-coupled neural networks (PCNN) is presented. Firstly a multi-scale decomposition on each source image is performed, and then the PCNN is used to combine these decomposition coefficients. Finally an inverse multi-scale transform is taken upon the new fused coefficients to reconstruct fusion image. The new algorithm utilizes the global feature of source images because the PCNN has the global couple and pulse synchronization characteristics. Series of experiments are performed about multi-modality medical images fusion such as CT/MRI, CT/SPECT, MRI/PET, etc. The experimental results show that the new algorithm is very effective and provides a good performance in fusing multi-modality medical images.

In this paper, we demonstrate that the machine learning approach of

rule extraction from a trained neural network

can be successfully applied to SARS-coronavirus cleavage site analysis. The extracted rules predict cleavage sites better than consensus patterns. Empirical experiments are also shown.

A new method is proposed for generating

if-then

rules to predict peptide binding to class I MHC proteins, from the amino acid sequence of any protein with known binders and non-binders. In this paper, we present an approach based on artificial neural networks (ANN) and knowledge-based genetic algorithm (KBGA) to predict the binding of peptides to MHC class I molecules. Our method includes rule extraction from a trained neural network and then enhancing the extracted rules by genetic evolution. Experimental results show that the method could generate new rules for MHC class I binding peptides prediction.

The problem of determining optimal decision model is a difficult combinatorial task in the fields of pattern classification, machine learning, and especially bioinformatics. Recently, support vector machine (SVM) has shown a higher performance than conventional learning methods in many applications. This paper proposes a new kernel function for support vector machine (SVM) and its learning method that results in fast convergence and good classification performance. The new kernel function is created by combining a set of kernel functions. A new learning method based on evolution algorithm (EA) is proposed to obtain the optimal decision model consisting of an optimal set of features as well as an optimal set of the parameters for combined kernel function. The experiments on clinical datasets such as stomach cancer, colon cancer, and leukemia datasets data sets indicates that the combined kernel function shows higher and more stable classification performance than other kernel functions.

Pixel values of contrast enhanced computed tomography (CE-CT) images are randomly changed. Also, the middle liver part has a problem to segregate the liver structure because of similar gray-level values of neighboring organs in the abdomen. In this paper, an automatic liver segmentation method using histogram processing is proposed for overcoming randomness of CE-CT images and removing other abdominal organs. Forty CE-CT slices of ten patients were selected to evaluate the proposed method. As the evaluation measure, the normalized average area and area error rate were used. From the results of experiments, liver segmentation using histogram process has similar performance as the manual method by medical doctor.

An improved adaptive RBF neural network is proposed to realize the continuous classification of left and right hand motor imagery tasks. Leader-follower clustering is used to initialize the centers and variances of hidden layer neurons, which matches the time-variant input features. Based on the features of multichannel EEG complexity and field power, the time courses of two evaluating indexes i.e. classification accuracy and mutual information (MI) are calculated to obtain the maximum with 87.14% and 0.53bit respectively. The results show that the improved algorithm can provide the flexible initial centers of RBF neural network and could be considered for the continuous classification of mental tasks for BCI (Brain Computer Interface) application.

This paper investigates the similarity of two sequences, one of the main issues for fragments clustering and classification when sequencing the genomes of microbial communities directly sampled from natural environment. In this paper, we use the relative entropy as a criterion of similarity of two sequences and discuss its characteristics in DNA sequences. A method for evaluating the relative entropy is presented and applied to the comparison between two sequences. With combination of the relative entropy and the length of variables defined in this paper, the similarity of sequences is easily obtained. The SOM and PCA are applied to cluster subsequences from different genomes. Computer simulations verify that the method works well.

At Windber Research Institute we have started research programs that use artificial neural networks (ANNs) in the study of breast cancer in order to identify heterogeneous data predictors of patient disease stages. As an initial effort, we have chosen matrix metalloproteinases (MMPs) as potential biomarker predictors. MMPs have been implicated in the early and late stage development of breast cancer. However, it is unclear whether these proteins hold predictive power for breast disease diagnosis, and we are not aware of any exploratory modeling efforts that address the question. Here we report the development of ANN models employing plasma levels of these proteins for breast disease predictions.

In whole genome shotgun sequencing when DNA fragments are derived from thousands of microorganisms in the environment sample, traditional alignment methods are impractical to use because of their high computation complexity. In this paper, we take the divergence vector which is consist of Kullback-Leibler divergences of different word lengths as the feature vector. Based on this, we use BP neural network to identify whether two fragments are from the same microorganism and obtain the similarity between fragments. Finally, we develop a new novel method to cluster DNA fragments from different microorganisms into different groups. Experiments show that it performs well.

In this paper,we constructed a data set of rice proteins with known locations from SWISS-PROT,using the Support Vector Machine to predicte the type of a given rice protein by incorporating sequence information with physics chemistry property of amino acid. Results are assessed through 5-fold cross-validation tests.

After the Chinese National Neuroinformatics Working Group was formed in 2001, neuroinformatics research has progressed rapidly in China. This paper reviews the history of neuroinformatics in China, reports current researches and discusses recent trends of neuroinformatics in China.

The Australian National Neuroscience Facility (NNF) has been established to provide Australian neuroscientists with access to networks of laboratories offering neuroscience consultancy, technical expertise and state-of-the-art equipment. The facility is fostering neuroscience excellence, combining science, technology, innovation, investment, creativity and the opportunity to advance our understanding and treatment of the brain and mind. Within the NNF a Neuroscience Informatics platform has been established with the objective of enhancing both the national neuroscience research capability, as well as the commercialisation opportunities for the Australian health and biotechnology industries. The Platform has developed a NeuroGrid facility consisting of computational resources and Grid middleware, internet accessible neuroimage databases, and standardised neuroimage analysis tools. A customised NeuroGrid portal is currently under development. It is envisaged that the NeuroGrid facility and software tools will provide the basis for application of Grid computing technologies to other areas of neuroscience research.

The Biomedical Imaging Lab in Singapore has been involved in neuroinformatics research for more than a decade. We are focused on clinical neuroinformatics, developing suitable models, tools, and databases. We report here our work on construction of anatomical, vascular, and functional brain atlases as well as development of atlas-assisted neuroscience education, research, and clinical applications. We also present future research activities.

There is a global trend to bring together research resources of the brain in the hope that these collaborations will provide critical information to the understanding of the brain as a system and its functions. Japan, among several countries, is committed to actively participating in this process with the hope that millions of people will greatly benefit from this activity. Currently, we are formulating plans and strategies in order to carry out this objective.. This paper will discuss perspectives of the Japanese Neuroinformatics Node.

In this paper, we propose a novel approach called the gradual noisy chaotic neural network (G-NCNN) to find a collision-free time slot schedule in a time division multiple access (TDMA) frame in packet radio network (PRN). In order to find a minimal average time delay of the network, we aim to find an optimal schedule which has the minimum frame length and provides the maximum channel utilization. The proposed two-phase neural network approach uses two different energy functions, with which the G-NCNN finds the minimal TDMA frame length in the first phase and the NCNN maximizes the node transmissions in the second phase. Numerical examples and comparisons with the previous methods show that the proposed method finds better solutions than previous algorithms. Furthermore, in order to show the difference between the proposed method and the hybrid method of the Hopfield neural network and genetic algorithms, we perform a paired t-test between two of them and show that G-NCNN can make significantly improvements.

This paper presents a fast online support vector machine (FOSVM) algorithm for variable-step CDMA power control. The FOSVM algorithm distinguishes new added samples and constructs current training sample set using K.K.T. condition in order to reduce the size of training samples. As a result, the training speed is effectively increased. We classify the received signals into two classes with FOSVM algorithm, then according to the output label of FOSVM and the distance from the data points to the SIR decision boundary, variable-step power control command is determined. Simulation results illustrate that the algorithm has a fast training speed and less support vectors. Its convergence performance is better than the fixed-step power control algorithm.

This study addresses a new blind channel equalization method using fourth-order cumulants of channel inputs and a three-layer neural network equalizer. The proposed algorithm is robust with respect to the existence of heavy Gaussian noise in a channel and does not require the minimum-phase characteristic of the channel. The transmitted signals at the receiver are over-sampled to ensure the channel described by a full-column rank matrix. It changes a single-input/single-output (SISO) finite-impulse response (FIR) channel to a single-input/multi-output (SIMO) channel. Based on the properties of the fourth-order cumulants of the over-sampled channel inputs, the iterative algorithm is derived to estimate the deconvolution matrix which makes the overall transfer matrix transparent, i.e., it can be reduced to the identity matrix by simple reordering and scaling. By using this estimated deconvolution matrix, which is the inverse of the over-sampled unknown channel, a three-layer neural network equalizer is implemented at the receiver. In simulation studies, the stochastic version of the proposed algorithm is tested with three-ray multi-path channels for on-line operation, and its performance is compared with a method based on conventional second-order statistics. Relatively good results, with fast convergence speed, are achieved, even when the transmitted symbols are significantly corrupted with Gaussian noise.

A new equalization method for a wireless ATM communication channel using a simplified version of the complex bilinear recurrent neural network (S-CBLRNN) is proposed in this paper. The S-BLRNN is then applied to the equalization of a wireless ATM channel for 8PSK and 16QAM. The results show that the proposed S-CBLRNN converges about 40 % faster than the CBLRNN and gives very favorable results in both of the MSE and SER criteria over the other equalizers.

Recently, interacting neural network has been studied out coming a novel result that the two neural networks can synchronize to a stationary weight state with the same initial inputs. In this paper, a simple but novel interacting neural network based authentication scheme is proposed, which can provide a full dynamic and security remote user authentication over a completely insecure communication channel.

MANET (Mobile Ad Hoc Network) is a collection of wireless mobile nodes forming a temporary computer communication network without the aid of any established infrastructure or centralized administration. MANET is characterized by both highly dynamic network topology and limited energy. This makes the efficiency of MANET depending not only on its control protocol, but also on its topology management and energy management. Clustering Strategy can improve the flexibility and scalability in network management. With graph theory model and genetic annealing hybrid optimization algorithm, this paper proposes a new clustering strategy named GASA (Genetic Algorithm Simulated Annealing). Simulation indicates that this strategy can with lower clustering cost and obtain dynamic balance of topology and load inside the whole network, so as to prolong the network lifetime.

Support Vector Machine(SVM) has become a popular tool for learning with large amounts of high dimensional data, but sometimes we prefer to incremental learning algorithms to handle very vast data for training SVM is very costly in time and memory consumption or because the data available are obtained at different intervals. For its outstanding power to summarize the data space in a concise way, incremental SVM framework is designed to deal with large-scale learning problems. This paper proposes a gradual algorithm for training SVM to incremental learning in a dividable way, taking the possible impact of new training data to history data each other into account. Training data are divided and combined in a crossed way to collect support vectors, and being divided into smaller sets makes it easier to decreases the computation complexity and the gradual process can be trained in a parallel way. The experiment results on test dataset show that the classification accuracy using proposed incremental algorithm is superior to that using batch SVM model, the parallel training method is effective to decrease the training time consumption.

The feature selection and weighting are two important parts of automatic text classification. In this paper we give a new method based on concept attributes. We use the

DEF

Terms of the Chinese word to extract concept attributes, and a Concept Tree (C-Tree) to give these attributes proper weighs considering their positions in the C-Tree, as this information describe the expression powers of the attributes. If these attributes are too weak to sustain the main meanings of the words, they will be deserted and the original word will be reserved. Otherwise, the attributes are selected in stead of the original words. Our main research purpose is to make a balance between concept features and word ones by set a shielded level as the threshold of the feature selection after weighting these features. According to the experiment results, we conclude that we can get enough information from the combined feature set for classification and efficiently reduce the useless features and the noises. In our experiment, the feature dimension is reduced to a much smaller space and the category precise is much better than the word selection methods. By choose different shielded levels, we finally select a best one when the average category precise is up to 93.7%. From the results, we find an extra finding that the precise differences between categories are smaller when we use combined features.

This paper applied the Data Envelopment Analysis (DEA) method to evaluate the input-output efficiency of constructing the enterprise information. Through projecting the inefficacy DEA unit to make the DEA effective in the unit, the data from projection can be used to train the BP network. In the pure BP network model, some flaws exist in BP network model such as slow speed in convergence and easily plunging into the local minima. On the contrary, artificial immune model has a few advantages such as antibody diversity inheritance mechanism and cell-chosen mechanism, which have been applied in this research. In this research, the BP network has been designed, and the IA-BP network model established. By taking the enterprise information application level, enterprise human resources state and information benefit index as the inputs, and the enterprise investing as the output, this model carries out the network training, until to get the satisfied investment decision method. Basing on this model, the enterprise can realize the maximized return on investment. This model not only constructs a new viable method to effectively use the research data, but also overcomes the drawbacks of non-linear description in the traditional information investment decision. The application results show that the model can satisfy the requirements of enterprise information, and provide the best decision method for enterprises as well.

As argued in this paper, a decision support system based on data mining and knowledge discovery is an important factor in improving productivity and yield. The proposed decision support system consists of a neural network model and an inference system based on fuzzy logic. First, the product results are predicted by the neural network model constructed by the quality index of the products that represent the quality of the etching process. And the quality indexes are classified according to and expert’s knowledge. Finally, the product conditions are estimated by the fuzzy inference system using the rules extracted from the classified patterns. We employed data mining and intelligent techniques to find the best condition for the etching process. The proposed decision support system is efficient and easy to be implemented for process management based on an expert’s knowledge.

Reticular Activating system which has a form of small neural networks in the brain is closely related system with the automatic nervous system. It takes charge of the function that distinguishes between memorizing one and the others, accepts the only selected information and discards the unnecessary things.In this paper, we propose Reticular Activating system which has functions of selective reaction, learning and inference. This system consists of Knowledge acquisition, selection , storing and retrieving part. Reticular Activating layer is connected to Meta knowledge in the high level of this system and takes part in Data Selection. We applied this system to the problem of analyzing the customer’s tastes.

In this paper, we describe glove-based information retrieval method and input device for wearable computers. We suggest an easy and effective alphanumeric input algorithm using gloves and conduct efficiency test. The key to the development of the proposed device is the use of unique operator-to-key mapping method, key-to-symbol mapping method and simple algorithm. We list and discuss traditional algorithm and method using a glove, then describe an improved newly proposed algorithm using gloves. The efficiency test was conducted and the results were compared with other glove based device and algorithm for wearable computers.

In this paper, we provide information an artificial intelligence agent for a smart home and discuss a context model for implementation in an efficient smart home. An artificial intelligence agent in a smart home learns about the occupants and the smart environment, and predicts the appliance service that they will want. We propose the SVM (Support Vector Machine) for the learning and prediction aspects of the artificial intelligence agent. The experiment was done using three methods. Each of these three methods applies a higher importance to a different set of context data, out of the data related to the occupant, home environment, and the characteristics of the home appliances. Excellent results were seen when the experiment applied a higher importance to the data related to the characteristics of the home appliances.

In this paper, a dynamical model for data clustering is proposed. This approach employs a network consisting of interacting elements with each representing an attribute vector of input data and receiving attractions from other elements within a certain region. Those attractions, determined by a predefined similarity measure, drive the elements to converge to their corresponding cluster center. With this model, neither the number of data clusters nor the initial guessing of cluster centers is required. Computer simulations for clustering of real images and Iris data set are performed. The results obtained so far are very promising.

Complete synchronization of

N

coupled systems with symmetric configurations is studied in this paper. The main idea of the synchronization stability criterion is based on stability analysis of zero solution of linearized dynamical systems. By rigorous theoretical analysis, a general synchronization stability criteria is derived for

N

coupled systems with the first state variable diffusive coupling. This criterion is convenient for us to explore the synchronization of a class of coupled dynamical systems. Finally, the famous Lorenz system and Hindmarsh-Rose(HR) neuron are used to test our theoretical analysis.

Cellular automata(CA) is not only a discrete dynamical system with infinite dimensions, but also an important computational model. How simple can a CA be and yet support interesting and complicated behavior. There are many unsolved problems in the theory of CA, which appeal many researchers to focus their attentions on the field, especially subclass of CA – linear CA. These studies cover the topological properties, chaotical properties, invertibility, attractors and the classification of linear CA etc.. This is a survey of known results and open questions of D-dimensional linear CA over ℤ

m

.

This paper considers a design framework of a computational experiment in finance. The examination of relationships between statistics used for economic forecasts evaluation and profitability of investment decisions reveals that only the ‘degree of improvement over efficient prediction’ shows robust links with profitability. If profits are not observable, this measure is proposed as an evaluation criterion for an economic prediction. Combined with directional accuracy, it could be used in an estimation technique for economic behavior, as an alternative to conventional least squares. Model discovery and performance surface optimization with genetic algorithm demonstrate profitability improvement with an inconclusive effect on statistical criteria.

Neural networks have advantages of the high tolerance to noisy data as well as the ability to classify patterns having not been trained. While being applied in data mining, the time required to induce models from large data sets are one of the most important considerations. In this paper, we introduce a query-based learning scheme to improve neural networks’ performance in data mining. Results show that the proposed algorithm can significantly reduce the training set cardinality. Additionally, the quality of training results can be also ensured. Our future work is to apply this concept to other data mining schemes and applications.

One of the most significant factors influencing the liquidity of financial markets is the amount of currency in circulation. Even the central bank is responsible for the distribution of the currency it could not assess the demand for the currency as it is influenced by the non-banking sector. Therefore the amount of currency in circulation have to be forecasted. This paper introduces feedforward structured neural network model and discusses its applicability to the forecasting of the currency in circulation. The forecasting performance of the new neural network model is compared with an ARIMA model. The results indicates that the performance of the neural network model is slightly better and that both models might be applied at least as supportive tools for the liquidity forecasting.

This study investigates the effectiveness of a hybrid approach with the time delay neural networks (TDNNs) and the genetic algorithms (GAs) in detecting temporal patterns for stock market prediction tasks. Since TDNN is a multi-layer, feed-forward network whose hidden neurons and output neurons are replicated across time, it has one more estimate of time delays in addition to a number of control variables of the artificial neural network (ANN) design. To estimate these many aspects of the TDNN design, a general method based on trial and error along with various heuristics or statistical techniques is proposed. However, for the reason that determining time delays or network architectural factors in a stand-alone mode doesn’t guarantee the illuminating improvement of the performance for building the TDNN models, we apply GAs to support optimization of time delays and network architectural factors simultaneously for the TDNN model. The results show that the accuracy of the integrated approach proposed for this study is higher than that of the standard TDNN and the recurrent neural networks (RNNs).

In this paper we firstly analysis the chaotic characters of three sets of the financial time series (Hang Sheng Index (HIS), Shanghai Stock Index and US gold price) based on the phase space reconstruction. But when we adopt the feedforward neural networks to predict those time series, we found this method run short of a criterion in selecting the training set, so we present a new method: using correlation dimension (CD) as the criterion. By the experiments, the method is proved effective.

Clustering of incomplete data using a neural network and the Gradient-Based Fuzzy c-Means (GBFCM) is proposed in this paper. The proposed algorithm is applied to the Iris data to evaluate its performance. When compared with the existing Optimal Completion Strategy FCM (OCSFCM), the proposed algorithm shows 18%-20% improvement of performance over the OCSFCM.

Artificial Neural Network (ANN) is widely used in the business to get on forecasting, but is often low performance for noisy data. Many techniques have been developed to improve ANN outcomes such as adding more algorithms, feature selection and feature weighting in input variables and modification of input case using instance selection. This paper proposes a Euclidean distance matrix approach to instance selection in ANN for financial forecasting. This approach optimizes a selection task for relevant instance. In addition, the technique improves prediction performance. In this research, ANN is applied to solve problems in forecasting a demand for corporate insurance. This research has compared the performance of forecasting a demand for corporate insurance through two types of ANN models; ANN and ISANN (ANN using Instance Selection supported by Euclidean distance metrics). Using ISANN to forecast a demand for corporate insurance is the most outstanding.

L-system is a grammar-like formalism introduced to simulate the development of organisms. The L-system grammar can be viewed as a sort of genetic information that will be used to generate a specific structure. However, throughout development, the string (genetic information) that will effectively be used to ‘draw’ the phenotype of an individual is a result of the derivation of the L-system grammar. This work investigates the effect of applying a genetic programming approach to evolve derived L-systems instead of evolving the Lsystem grammar. The crossing over of plants from different species results in hybrid plants resembling a ‘Frankstree’, i.e. plants resultant from phenotypically different parents that present unusual body structures.

This paper proposes a novel constructive learning algorithm for a competitive neural network. The proposed algorithm is developed by taking ideas from the immune system and demonstrates robustness for data clustering in the initial experiments reported here for three benchmark problems. Comparisons with results from the literature are also provided. To automatically segment the resultant neurons at the output, a tool from graph theory was used with promising results. A brief sensitivity analysis of the algorithm was performed in order to investigate the influence of the main user-defined parameters on the learning speed and accuracy of the results presented. General discussions and avenues for future works are also provided.