Search

Recent years have witnessed an incredibly increasing interest in the topic of incremental learning. Unlike conventional machine learning situations, data flow targeted by incremental learning becomes available continuously over time. Accordingly, it is desirable to be able to abandon the traditional assumption of the availability of representative training data during the training period to develop decision boundaries. Under scenarios of continuous data flow, the challenge is how to transform the vast amount of stream raw data into information and knowledge representation, and accumulate experience over time to support future decision-making process. In this paper, we propose a general adaptive incremental learning framework named ADAIN that is capable of learning from continuous raw data, accumulating experience over time, and using such knowledge to improve future learning and prediction performance. Detailed system level architecture and design strategies are presented in this paper. Simulation results over several real-world data sets are used to validate the effectiveness of this method.

In this paper, the exponential synchronization for a class of complex networks with finite distributed delays coupling is studied via periodically intermittent control. Some novel and useful criteria are derived by utilizing a different technique compared with some correspondingly previous results. As a special case, some sufficient conditions ensuring the exponential synchronization for a class of coupled neural networks with distributed delays are obtained. Furthermore, a feasible region of the control parameters is derived for the realization of exponential synchronization. It is worth noting that the synchronized state in this paper is not an isolated node but a non-decoupled state, in which the inner coupling matrix and the degree of the nodes play a central role. Additionally, the traditional assumptions on control width, non-control width, and discrete delays are removed in our results. Finally, some numerical simulations are given to demonstrate the effectiveness of the proposed control method.

Genetic regulatory networks can be described by nonlinear differential equations with time delays. In this paper, we study both locally and globally delay-independent stability of genetic regulatory networks, taking messenger ribonucleic acid alternative splicing into consideration. Based on nonnegative matrix theory, we first develop necessary and sufficient conditions for locally delay-independent stability of genetic regulatory networks with multiple time delays. Compared to the previous results, these conditions are easy to verify. Then we develop sufficient conditions for global delay-independent stability for genetic regulatory networks. Compared to the previous results, this sufficient condition is less conservative. To illustrate theorems developed in this paper, we analyze delay-independent stability of two genetic regulatory networks: a real-life repressilatory network with three genes and three proteins, and a synthetic gene regulatory network with five genes and seven proteins. The simulation results show that the theorems developed in this paper can effectively determine the delay-independent stability of genetic regulatory networks.

Epipolar geometry, the cornerstone of perspective stereo vision, has been studied extensively since the advent of computer vision. Establishing such a geometric constraint is of primary importance, as it allows the recovery of the 3-D structure of scenes. Estimating the epipolar constraints of nonperspective stereo is difficult, they can no longer be defined because of the complexity of the sensor geometry. This paper will show that these limitations are, to some extent, a consequence of the static image frames commonly used in vision. The conventional frame-based approach suffers from a lack of the dynamics present in natural scenes. We introduce the use of neuromorphic event-based--rather than frame-based--vision sensors for perspective stereo vision. This type of sensor uses the dimension of time as the main conveyor of information. In this paper, we present a model for asynchronous event-based vision, which is then used to derive a general new concept of epipolar geometry linked to the temporal activation of pixels. Practical experiments demonstrate the validity of the approach, solving the problem of estimating the fundamental matrix applied, in a first stage, to classic perspective vision and then to more general cameras. Furthermore, this paper shows that the properties of event-based vision sensors allow the exploration of not-yet-defined geometric relationships, finally, we provide a definition of general epipolar geometry deployable to almost any visual sensor.

Feedforward neural network is one of the most commonly used function approximation techniques and has been applied to a wide variety of problems arising from various disciplines. However, neural networks are black-box models having multiple challenges/difficulties associated with training and generalization. This paper initially looks into the internal behavior of neural networks and develops a detailed interpretation of the neural network functional geometry. Based on this geometrical interpretation, a new set of variables describing neural networks is proposed as a more effective and geometrically interpretable alternative to the traditional set of network weights and biases. Then, this paper develops a new formulation for neural networks with respect to the newly defined variables; this reformulated neural network (ReNN) is equivalent to the common feedforward neural network but has a less complex error response surface. To demonstrate the learning ability of ReNN, in this paper, two training methods involving a derivative-based (a variation of backpropagation) and a derivative-free optimization algorithms are employed. Moreover, a new measure of regularization on the basis of the developed geometrical interpretation is proposed to evaluate and improve the generalization ability of neural networks. The value of the proposed geometrical interpretation, the ReNN approach, and the new regularization measure are demonstrated across multiple test problems. Results show that ReNN can be trained more effectively and efficiently compared to the common neural networks and the proposed regularization measure is an effective indicator of how a network would perform in terms of generalization.

In this paper, we analyze the convergence of two general classes of optimization algorithms for regularized kernel methods with convex loss function and quadratic norm regularization. The first methodology is a new class of algorithms based on fixed-point iterations that are well-suited for a parallel implementation and can be used with any convex loss function. The second methodology is based on coordinate descent, and generalizes some techniques previously proposed for linear support vector machines. It exploits the structure of additively separable loss functions to compute solutions of line searches in closed form. The two methodologies are both very easy to implement. In this paper, we also show how to remove non-differentiability of the objective functional by exactly reformulating a convex regularization problem as an unconstrained differentiable stabilization problem.

This paper evaluates the four leading techniques proposed in the literature for construction of prediction intervals (PIs) for neural network point forecasts. The delta, Bayesian, bootstrap, and mean-variance estimation (MVE) methods are reviewed and their performance for generating high-quality PIs is compared. PI-based measures are proposed and applied for the objective and quantitative assessment of each method's performance. A selection of 12 synthetic and real-world case studies is used to examine each method's performance for PI construction. The comparison is performed on the basis of the quality of generated PIs, the repeatability of the results, the computational requirements and the PIs variability with regard to the data uncertainty. The obtained results in this paper indicate that: 1) the delta and Bayesian methods are the best in terms of quality and repeatability, and 2) the MVE and bootstrap methods are the best in terms of low computational load and the width variability of PIs. This paper also introduces the concept of combinations of PIs, and proposes a new method for generating combined PIs using the traditional PIs. Genetic algorithm is applied for adjusting the combiner parameters through minimization of a PI-based cost function subject to two sets of restrictions. It is shown that the quality of PIs produced by the combiners is dramatically better than the quality of PIs obtained from each individual method.

The algorithm and the results of a blind multiuser detector using a machine learning technique called support vector machine (SVM) on a chaos-based code division multiple access system is presented in this paper. Simulation results showed that the performance achieved by using SVM is comparable to existing minimum mean square error (MMSE) detector under both additive white Gaussian noise (AWGN) and Rayleigh fading conditions. However, unlike the MMSE detector, the SVM detector does not require the knowledge of spreading codes of other users in the system or the estimate of the channel noise variance. The optimization of this algorithm is considered in this paper and its complexity is compared with the MMSE detector. This detector is much more suitable to work in the forward link than MMSE. In addition, original theoretical bit-error rate expressions for the SVM detector under both AWGN and Rayleigh fading are derived to verify the simulation results.

Negative correlation learning (NCL) is a neural network ensemble learning algorithm that introduces a correlation penalty term to the cost function of each individual network so that each neural network minimizes its mean square error (MSE) together with the correlation of the ensemble. This paper analyzes NCL and reveals that the training of NCL (when lambda = 1) corresponds to training the entire ensemble as a single learning machine that only minimizes the MSE without regularization. This analysis explains the reason why NCL is prone to overfitting the noise in the training set. This paper also demonstrates that tuning the correlation parameter lambda in NCL by cross validation cannot overcome the overfitting problem. The paper analyzes this problem and proposes the regularized negative correlation learning (RNCL) algorithm which incorporates an additional regularization term for the whole ensemble. RNCL decomposes the ensemble's training objectives, including MSE and regularization, into a set of sub-objectives, and each sub-objective is implemented by an individual neural network. In this paper, we also provide a Bayesian interpretation for RNCL and provide an automatic algorithm to optimize regularization parameters based on Bayesian inference. The RNCL formulation is applicable to any nonlinear estimator minimizing the MSE. The experiments on synthetic as well as real-world data sets demonstrate that RNCL achieves better performance than NCL, especially when the noise level is nontrivial in the data set.

This paper considers the lag synchronization (LS) issue of unknown coupled chaotic delayed Yang-Yang-type fuzzy neural networks (YYFCNN) with noise perturbation. Separate research work has been published on the stability of fuzzy neural network and LS issue of unknown coupled chaotic neural networks, as well as its application in secure communication. However, there have not been any studies that integrate the two. Motivated by the achievements from both fields, we explored the benefits of integrating fuzzy logic theories into the study of LS problems and applied the findings to secure communication. Based on adaptive feedback control techniques and suitable parameter identification, several sufficient conditions are developed to guarantee the LS of coupled chaotic delayed YYFCNN with or without noise perturbation. The problem studied in this paper is more general in many aspects. Various problems studied extensively in the literature can be treated as special cases of the findings of this paper, such as complete synchronization (CS), effect of fuzzy logic, and noise perturbation. This paper presents an illustrative example and uses simulated results of this example to show the feasibility and effectiveness of the proposed adaptive scheme. This research also demonstrates the effectiveness of application of the proposed adaptive feedback scheme in secure communication by comparing chaotic masking with fuzziness with some previous studies. Chaotic signal with fuzziness is more complex, which makes unmasking more difficult due to the added fuzzy logic.

This paper considers an output feedback learning control for a class of uncertain nonlinear systems with flexible components. The distinct time delay caused by system flexibility leads to the phase lag phenomenon and low system bandwidth. Therefore, the tracking problem of such systems is very difficult and challenging. To improve the tracking performance of such systems, an iterative learning control scheme using the Fourier neural network (FNN) is presented in this paper. This scheme uses only local output information for feedback. FNN employs orthogonal complex Fourier exponentials as its activation functions and the physical meaning of its hidden-layer neurons is clear. The FNN-based learning controller introduced here relies on the frequency-domain method, which converts the tracking problem in the time domain into a number of regulation problems in the frequency domain. A novel phase compensation method is introduced to deal with the phase lag phenomenon, so that the bandwidth of the closed-loop system is increased. Experiments on a belt-driven positioning table are conducted to show the effectiveness of the proposed controller.

This paper proposes an incremental multiple-object recognition and localization (IMORL) method. The objective of IMORL is to adaptively learn multiple interesting objects in an image. Unlike the conventional multiple-object learning algorithms, the proposed method can automatically and adaptively learn from continuous video streams over the entire learning life. This kind of incremental learning capability enables the proposed approach to accumulate experience and use such knowledge to benefit future learning and the decision making process. Furthermore, IMORL can effectively handle variations in the number of instances in each data chunk over the learning life. Another important aspect analyzed in this paper is the concept drifting issue. In multiple-object learning scenarios, it is a common phenomenon that new interesting objects may be introduced during the learning life. To handle this situation, IMORL uses an adaptive learning principle to autonomously adjust to such new information. The proposed approach is independent of the base learning models, such as decision tree, neural networks, support vector machines, and others, which provide the flexibility of using this method as a general learning methodology in multiple-object learning scenarios. In this paper, we use a neural network with a multilayer perceptron (MLP) structure as the base learning model and test the performance of this method in various video stream data sets. Simulation results show the effectiveness of this method.

In this paper, we present a novel neural network (NN) adaptive control architecture with guaranteed transient performance. With this new architecture, both input and output signals of an uncertain nonlinear system follow a desired linear system during the transient phase, in addition to stable tracking. This new architecture uses a low-pass filter in the feedback loop, which consequently enables to enforce the desired transient performance by increasing the adaptation gain. For the guaranteed transient performance of both input and output signals of the uncertain nonlinear system, the L1 gain of a cascaded system, comprised of the low-pass filter and the closed-loop desired reference model, is required to be less than the inverse of the Lipschitz constant of the unknown nonlinearities in the system. The tools from this paper can be used to develop a theoretically justified verification and validation framework for NN adaptive controllers. Simulation results illustrate the theoretical findings.

This paper presents a hardware implementation of multilayer feedforward neural networks (NN) using reconfigurable field-programmable gate arrays (FPGAs). Despite improvements in FPGA densities, the numerous multipliers in an NN limit the size of the network that can be implemented using a single FPGA, thus making NN applications not viable commercially. The proposed implementation is aimed at reducing resource requirement, without much compromise on the speed, so that a larger NN can be realized on a single chip at a lower cost. The sequential processing of the layers in an NN has been exploited in this paper to implement large NNs using a method of layer multiplexing. Instead of realizing a complete network, only the single largest layer is implemented. The same layer behaves as different layers with the help of a control block. The control block ensures proper functioning by assigning the appropriate inputs, weights, biases, and excitation function of the layer that is currently being computed. Multilayer networks have been implemented using Xilinx FPGA "XCV400hq240". The concept used is shown to be very effective in reducing resource requirements at the cost of a moderate overhead on speed. This implementation is proposed to make NN applications viable in terms of cost and speed for online applications. An NN-based flux estimator is implemented in FPGA and the results obtained are presented.

Temporal sequence learning is one of the most critical components for human intelligence. In this paper, a novel hierarchical structure for complex temporal sequence learning is proposed. Hierarchical organization, a prediction mechanism, and one-shot learning characterize the model. In the lowest level of the hierarchy, we use a modified Hebbian learning mechanism for pattern recognition. Our model employs both active 0 and active 1 sensory inputs. A winner-take-all (WTA) mechanism is used to select active neurons that become the input for sequence learning at higher hierarchical levels. Prediction is an essential element of our temporal sequence learning model. By correct prediction, the machine indicates it knows the current sequence and does not require additional learning. When the prediction is incorrect, one-shot learning is executed and the machine learns the new input sequence as soon as the sequence is completed. A four-level hierarchical structure that isolates letters, words, sentences, and strophes is used in this paper to illustrate the model.

In this paper, we introduce feedback-based associative learning in self-organized learning arrays (SOLAR). SOLAR structures are hierarchically organized networks of sparsely connected neurons that define their own functions and select their interconnections locally. This paper provides a description of neuron self-organization and signal processing. Feedforward processing is used to make necessary correlations and learn the input patterns. Discovered associations between neuron inputs are used to generate feedback signals. These feedback signals, when propagated to the primary inputs, can establish the expected input values. This can be used for heteroassociative (HA) and autoassociative (AA) learning and pattern recognition. Example applications in HA learning are given.

In this paper, a novel stochastic (or online) training algorithm for neural networks, named parameter incremental learning (PIL) algorithm, is proposed and developed. The main idea of the PIL strategy is that the learning algorithm should not only adapt to the newly presented input-output training pattern by adjusting parameters, but also preserve the prior results. A general PIL algorithm for feedforward neural networks is accordingly presented as the first-order approximate solution to an optimization problem, where the performance index is the combination of proper measures of preservation and adaptation. The PIL algorithms for the multilayer perceptron (MLP) are subsequently derived. Numerical studies show that for all the three benchmark problems used in this paper the PIL algorithm for MLP is measurably superior to the standard online backpropagation (BP) algorithm and the stochastic diagonal Levenberg-Marquardt (SDLM) algorithm in terms of the convergence speed and accuracy. Other appealing features of the PIL algorithm are that it is computationally as simple as the BP algorithm, and as easy to use as the BP algorithm. It, therefore, can be applied, with better performance, to any situations where the standard online BP algorithm is applicable.

This paper investigates the method of forecasting stock price difference on artificially generated price series data using neuro-fuzzy systems and neural networks. As trading profits is more important to an investor than statistical performance, this paper proposes a novel rough set-based neuro-fuzzy stock trading decision model called stock trading using rough set-based pseudo outer-product (RSPOP) which synergizes the price difference forecast method with a forecast bottleneck free trading decision model. The proposed stock trading with forecast model uses the pseudo outer-product based fuzzy neural network using the compositional rule of inference [POPFNN-CRI(S)] with fuzzy rules identified using the RSPOP algorithm as the underlying predictor model and simple moving average trading rules in the stock trading decision model. Experimental results using the proposed stock trading with RSPOP forecast model on real world stock market data are presented. Trading profits in terms of portfolio end values obtained are benchmarked against stock trading with dynamic evolving neural-fuzzy inference system (DENFIS) forecast model, the stock trading without forecast model and the stock trading with ideal forecast model. Experimental results showed that the proposed model identified rules with greater interpretability and yielded significantly higher profits than the stock trading with DENFIS forecast model and the stock trading without forecast model.

In this paper, we are concerned with the concept of fuzzy logic networks and logic-based data analysis realized within this framework. The networks under discussion are homogeneous architectures comprising of OR/AND neurons originally introduced by Hirota and Pedrycz. Being treated here as generic processing units, OR/AND neurons are neurofuzzy constructs that exhibit well-defined logic characteristics and are endowed with a high level of parametric flexibility and come with significant interpretation abilities. The composite logic nature of the logic neurons becomes instrumental in covering a broad spectrum of logic dependencies whose character spread in-between between those being captured by plain and and or logic descriptors (connectives). From the functional standpoint, the developed network realizes a logic approximation of multidimensional mappings between unit hypercubes, that is transformations from [0, 1]n to [0, 1]m. The way in which the structure of the network has been formed is highly modular and becomes reflective of a general concept of decomposition of logic expressions and Boolean functions (as being commonly encountered in two-valued logic). In essence, given a collection of input variables, selected is their subset and transformed into new composite variable, which in turn is used in the consecutive module of the network. These intermediate synthetic variables are the result of the successive problem (mapping) decomposition. The development of the network is realized through genetic optimization. This helps address important issues of structural optimization (where we are concerned with a selection of a subset of variables and their allocation within the network) and reaching a global minimum when carrying out an extensive parametric optimization (adjustments of the connections of the neurons). The paper offers a comprehensive and user-interactive design procedure including a simple pruning mechanism whose intention is to enhance the interpretability of the network while reducing its size. The experimental studies comprise of three parts. First, we demonstrate the performance of the network on Boolean data (that leads to some useful comparative observations considering a wealth of optimization tools available in two-valued logic and digital systems). Second, we discuss synthetic multivalued data that helps focus on the approximation abilities of the network. Finally, show the generation of logic expressions describing selected data sets coming from the machine learning repository.

This paper presents a neural network architecture using a support vector machine (SVM) as an inference engine (IE) for classification of light detection and ranging (Lidar) data. Lidar data gives a sequence of laser backscatter intensities obtained from laser shots generated from an airborne object at various altitudes above the earth surface. Lidar data is pre-filtered to remove high frequency noise. As the Lidar shots are taken from above the earth surface, it has some air backscatter information, which is of no importance for detecting underwater objects. Because of these, the air backscatter information is eliminated from the data and a segment of this data is subsequently selected to extract features for classification. This is then encoded using linear predictive coding (LPC) and polynomial approximation. The coefficients thus generated are used as inputs to the two branches of a parallel neural architecture. The decisions obtained from the two branches are vector multiplied and the result is fed to an SVM-based IE that presents the final inference. Two parallel neural architectures using multilayer perception (MLP) and hybrid radial basis function (HRBF) are considered in this paper. The proposed structure fits the Lidar data classification task well due to the inherent classification efficiency of neural networks and accurate decision-making capability of SVM. A Bayesian classifier and a quadratic classifier were considered for the Lidar data classification task but they failed to offer high prediction accuracy. Furthermore, a single-layered artificial neural network (ANN) classifier was also considered and it failed to offer good accuracy. The parallel ANN architecture proposed in this paper offers high prediction accuracy (98.9%) and is found to be the most suitable architecture for the proposed task of Lidar data classification.

The additive recurrent network structure of linear threshold neurons represents a class of biologically-motivated models, where nonsaturating transfer functions are necessary for representing neuronal activities, such as that of cortical neurons. This paper extends the existing results of dynamics analysis of such linear threshold networks by establishing new and milder conditions for boundedness and asymptotical stability, while allowing for multistability. As a condition for asymptotical stability, it is found that boundedness does not require a deterministic matrix to be symmetric or possess positive off-diagonal entries. The conditions put forward an explicit way to design and analyze such networks. Based on the established theory, an alternate approach to study such networks is through permitted and forbidden sets. An application of the linear threshold (LT) network is analog associative memory, for which a simple design method describing the associative memory is suggested in this paper. The proposed design method is similar to a generalized Hebbian approach, but with distinctions of additional network parameters for normalization, excitation and inhibition, both on a global and local scale. The computational abilities of the network are dependent on its nonlinear dynamics, which in turn is reliant upon the sparsity of the memory vectors.

In a previous paper, the self-trapping network (STN) was introduced as more biologically realistic than attractor neural networks (ANNs) based on the Ising model. This paper extends the previous analysis of a one-dimensional (1-D) STN storing a single memory to a model that stores multiple memories and that possesses generalized sparse connnectivity. The energy, Lyapunov function, and partition function derived for the 1-D model are generalized to the case of an attractor network with only near-neighbor synapses, coupled to a system that computes memory overlaps. Simulations reveal that 1) the STN dramatically reduces intra-ANN connectivity without severly affecting the size of basins of attraction, with fast self-trapping able to sustain attractors even in the absence of intra-ANN synapses; 2) the basins of attraction can be controlled by a single free parameter, providing natural attention-like effects; 3) the same parameter determines the memory capacity of the network, and the latter is much less dependent than a standard ANN on the noise level of the system; 4) the STN serves as a useful memory for some correlated memory patterns for which the standard ANN totally fails; 5) the STN can store a large number of sparse patterns; and 6) a Monte Carlo procedure, a competitive neural network, and binary neurons with thresholds can be used to induce self-trapping.

Original Hebbian encoding scheme of bidirectional associative memory (BAM) provides a poor pattern capacity and recall performance. Based on Rosenblatt's perceptron learning algorithm, the pattern capacity of BAM is enlarged, and perfect recall of all training pattern pairs is guaranteed. However, these methods put their emphases on pattern capacity, rather than error correction capability which is another critical point of BAM. This paper analyzes the sensitivity to noise in BAM and obtains an interesting idea to improve noise immunity of BAM. Some researchers have found that the noise sensitivity of BAM relates to the minimum absolute value of net inputs (MAV). However, in this paper, the analysis on failure association shows that it is related not only to MAV but also to the variance of weights associated with synapse connections. In fact, it is a positive monotone increasing function of the quotient of MAV divided by the variance of weights. This idea provides an useful principle of improving error correction capability of BAM. Some revised encoding schemes, such as small variance learning for BAM (SVBAM), evolutionary pseudorelaxation learning for BAM (EPRLAB) and evolutionary bidirectional learning (EBL), have been introduced to illustrate the performance of this principle. All these methods perform better than their original versions in noise immunity. Moreover, these methods have no negative effect on the pattern capacity of BAM. The convergence of these methods is also discussed in this paper. If there exist solutions, EPRLAB and EBL always converge to a global optimal solution in the senses of both pattern capacity and noise immunity. However, the convergence of SVBAM may be affected by a preset function.

This work presents a novel learning algorithm for efficient construction of the radial basis function (RBF) networks that can deliver the same level of accuracy as the support vector machines (SVMs) in data classification applications. The proposed learning algorithm works by constructing one RBF subnetwork to approximate the probability density function of each class of objects in the training data set. With respect to algorithm design, the main distinction of the proposed learning algorithm is the novel kernel density estimation algorithm that features an average time complexity of O(nlogn), where n is the number of samples in the training data set. One important advantage of the proposed learning algorithm, in comparison with the SVM, is that the proposed learning algorithm generally takes far less time to construct a data classifier with an optimized parameter setting. This feature is of significance for many contemporary applications, in particular, for those applications in which new objects are continuously added into an already large database. Another desirable feature of the proposed learning algorithm is that the RBF networks constructed are capable of carrying out data classification with more than two classes of objects in one single run. In other words, unlike with the SVM, there is no need to resort to mechanisms such as one-against-one or one-against-all for handling datasets with more than two classes of objects. The comparison with SVM is of particular interest, because it has been shown in a number of recent studies that SVM generally are able to deliver higher classification accuracy than the other existing data classification algorithms. As the proposed learning algorithm is instance-based, the data reduction issue is also addressed in this paper. One interesting observation in this regard is that, for all three data sets used in data reduction experiments, the number of training samples remaining after a na/spl inodot//spl uml/ve data reduction mechanism is applied is quite close to the number of support vectors identified by the SVM software. This paper also compares the performance of the RBF networks constructed with the proposed learning algorithm and those constructed with a conventional cluster-based learning algorithm. The most interesting observation learned is that, with respect to data classification, the distributions of training samples near the boundaries between different classes of objects carry more crucial information than the distributions of samples in the inner parts of the clusters.

This paper presents a new sequential learning algorithm for radial basis function (RBF) networks referred to as generalized growing and pruning algorithm for RBF (GGAP-RBF). The paper first introduces the concept of significance for the hidden neurons and then uses it in the learning algorithm to realize parsimonious networks. The growing and pruning strategy of GGAP-RBF is based on linking the required learning accuracy with the significance of the nearest or intentionally added new neuron. Significance of a neuron is a measure of the average information content of that neuron. The GGAP-RBF algorithm can be used for any arbitrary sampling density for training samples and is derived from a rigorous statistical point of view. Simulation results for bench mark problems in the function approximation area show that the GGAP-RBF outperforms several other sequential learning algorithms in terms of learning speed, network size and generalization performance regardless of the sampling density function of the training data.

This paper reports a study on the problem of the blind simultaneous extraction of specific groups of independent components from a linear mixture. This paper first presents a general overview and unification of several information theoretic criteria for the extraction of a single independent component. Then, our contribution fills the theoretical gap that exists between extraction and separation by presenting tools that extend these criteria to allow the simultaneous blind extraction of subsets with an arbitrary number of independent components. In addition, we analyze a family of learning algorithms based on Stiefel manifolds and the natural gradient ascent, present the nonlinear optimal activations (score) functions, and provide new or extended local stability conditions. Finally, we illustrate the performance and features of the proposed approach by computer-simulation experiments.

It has been proven that there are synchrony (or phase-locking) phenomena present in multiple oscillating systems such as electrical circuits, lasers, chemical reactions, and human neurons. If the measurements of these systems cannot detect the individual oscillators but rather a superposition of them, as in brain electrophysiological signals (electro- and magneoencephalogram), spurious phase locking will be detected. Current source-extraction techniques attempt to undo this superposition by assuming properties on the data, which are not valid when underlying sources are phase-locked. Statistical independence of the sources is one such invalid assumption, as phase-locked sources are dependent. In this paper, we introduce methods for source separation and clustering which make adequate assumptions for data where synchrony is present, and show with simulated data that they perform well even in cases where independent component analysis and other well-known source-separation methods fail. The results in this paper provide a proof of concept that synchrony-based techniques are useful for low-noise applications.

Many applications of machine learning involve analysis of sparse high-dimensional data, in which the number of input features is larger than the number of data samples. Standard inductive learning methods may not be sufficient for such data, and this provides motivation for nonstandard learning settings. This paper investigates a new learning methodology called learning through contradictions or Universum support vector machine (U-SVM). U-SVM incorporates a priori knowledge about application data, in the form of additional Universum samples, into the learning process. This paper investigates possible advantages of U-SVM versus standard SVM, and describes the practical conditions necessary for the effectiveness of the U-SVM. These conditions are based on the analysis of the univariate histograms of projections of training samples onto the normal direction vector of (standard) SVM decision boundary. Several empirical comparisons are presented to illustrate the practical utility of the proposed approach.

This paper describes a practical framework for using multilayer feedforward neural networks to simultaneously fit both a function and its first derivatives. This framework involves two steps. The first step is to train the network to optimize a performance index, which includes both the error in fitting the function and the error in fitting the derivatives. The second step is to prune the network by removing neurons that cause overfitting and then to retrain it. This paper describes two novel types of overfitting that are only observed when simultaneously fitting both a function and its first derivatives. A new pruning algorithm is proposed to eliminate these types of overfitting. Experimental results show that the pruning algorithm successfully eliminates the overfitting and produces the smoothest responses and the best generalization among all the training algorithms that we have tested.

This paper presents a novel offline modeling for product quality prediction of mineral processing which consists of a number of unit processes in series. The prediction of the product quality of the whole mineral process (i.e., the mixed concentrate grade) plays an important role and the establishment of its predictive model is a key issue for the plantwide optimization. For this purpose, a hybrid modeling approach of the mixed concentrate grade prediction is proposed, which consists of a linear model and a nonlinear model. The least-squares support vector machine is adopted to establish the nonlinear model. The inputs of the predictive model are the performance indices of each unit process, while the output is the mixed concentrate grade. In this paper, the model parameter selection is transformed into the shape control of the probability density function (PDF) of the modeling error. In this context, both the PDF-control-based and minimum-entropy-based model parameter selection approaches are proposed. Indeed, this is the first time that the PDF shape control idea is used to deal with system modeling, where the key idea is to turn model parameters so that either the modeling error PDF is controlled to follow a target PDF or the modeling error entropy is minimized. The experimental results using the real plant data and the comparison of the two approaches are discussed. The results show the effectiveness of the proposed approaches.

Neural networks (NNs) have numerous applications to online processes, but the problem of stability is rarely discussed. This is an extremely important issue because, if the stability of a solution is not guaranteed, the equipment that is being used can be damaged, which can also cause serious accidents. It is true that in some research papers this problem has been considered, but this concerns continuous-time NN only. At the same time, there are many systems that are better described in the discrete time domain such as population of animals, the annual expenses in an industry, the interest earned by a bank, or the prediction of the distribution of loads stored every hour in a warehouse. Therefore, it is of paramount importance to consider the stability of the discrete-time NN. This paper makes several important contributions. 1) A theorem is stated and proven which guarantees uniform stability of a general discrete-time system. 2) It is proven that the backpropagation (BP) algorithm with a new time-varying rate is uniformly stable for online identification and the identification error converges to a small zone bounded by the uncertainty. 3) It is proven that the weights' error is bounded by the initial weights' error, i.e., overfitting is eliminated in the proposed algorithm. 4) The BP algorithm is applied to predict the distribution of loads that a transelevator receives from a trailer and places in the deposits in a warehouse every hour, so that the deposits in the warehouse are reserved in advance using the prediction results. 5) The BP algorithm is compared with the recursive least square (RLS) algorithm and with the Takagi-Sugeno type fuzzy inference system in the problem of predicting the distribution of loads in a warehouse, giving that the first and the second are stable and the third is unstable. 6) The BP algorithm is compared with the RLS algorithm and with the Kalman filter algorithm in a synthetic example.

This paper presents a new approach for identifying and validating the F/A-18 aeroservoelastic model, based on flight flutter tests. The neural network (NN), trained with five different flight flutter cases, is validated using 11 other flight flutter test (FFT) data. A total of 16 FFT cases were obtained for all three flight regimes (subsonic, transonic, and supersonic) at Mach numbers ranging between 0.85 and 1.30 and at altitudes of between 5000 and 25 000 ft. The results obtained highlight the efficiency of the multilayer perceptron NN in model identification. Optimization of the NN requires mixing of two proprieties: the hidden layer size reduction and four-layered NN performances. This paper shows that a four-layer NN with only 16 neurons is enough to create an accurate model. The fit coefficients were higher than 92% for both the identification and the validation test data, thus demonstrating accuracy of the NN.

This paper proposes three novel training methods, two of them based on the backpropagation approach and a third one based on information theory for multilayer perceptron (MLP) binary classifiers. Both backpropagation methods are based on the maximal-margin (MM) principle. The first one, based on the gradient descent with adaptive learning rate algorithm (GDX) and named maximum-margin GDX (MMGDX), directly increases the margin of the MLP output-layer hyperplane. The proposed method jointly optimizes both MLP layers in a single process, backpropagating the gradient of an MM-based objective function, through the output and hidden layers, in order to create a hidden-layer space that enables a higher margin for the output-layer hyperplane, avoiding the testing of many arbitrary kernels, as occurs in case of support vector machine (SVM) training. The proposed MM-based objective function aims to stretch out the margin to its limit. An objective function based on Lp-norm is also proposed in order to take into account the idea of support vectors, however, overcoming the complexity involved in solving a constrained optimization problem, usually in SVM training. In fact, all the training methods proposed in this paper have time and space complexities O(N) while usual SVM training methods have time complexity O(N (3)) and space complexity O(N (2)) , where N is the training-data-set size. The second approach, named minimization of interclass interference (MICI), has an objective function inspired on the Fisher discriminant analysis. Such algorithm aims to create an MLP hidden output where the patterns have a desirable statistical distribution. In both training methods, the maximum area under ROC curve (AUC) is applied as stop criterion. The third approach offers a robust training framework able to take the best of each proposed training method. The main idea is to compose a neural model by using neurons extracted from three other neural networks, each one previously trained by MICI, MMGDX, and Levenberg-Marquard (LM), respectively. The resulting neural network was named assembled neural network (ASNN). Benchmark data sets of real-world problems have been used in experiments that enable a comparison with other state-of-the-art classifiers. The results provide evidence of the effectiveness of our methods regarding accuracy, AUC, and balanced error rate.

This paper studies a general class of dynamical neural networks with lateral inhibition, exhibiting winner-take-all (WTA) behavior. These networks are motivated by a metal-oxide-semiconductor field effect transistor (MOSFET) implementation of neural networks, in which mutual competition plays a very important role. We show that for a fairly general class of competitive neural networks, WTA behavior exists. Sufficient conditions for the network to have a WTA equilibrium are obtained, and rigorous convergence analysis is carried out. The conditions for the network to have the WTA behavior obtained in this paper provide design guidelines for the network implementation and fabrication. We also demonstrate that whenever the network gets into the WTA region, it will stay in that region and settle down exponentially fast to the WTA point. This provides a speeding procedure for the decision making: as soon as it gets into the region, the winner can be declared. Finally, we show that this WTA neural network has a self-resetting property, and a resetting principle is proposed.

Motivated by the reportedly strong performance of particle filters (PFs) for noise reduction on essentially linear speech production models, and the mounting evidence that the introduction of nonlinearities can lead to a refined speech model, this paper presents a study of PF solutions to the problem of speech enhancement in the context of nonlinear, neural-type speech models. Several variations of a global model are presented (single/multiple neurons; bias/no bias), and corresponding PF solutions are derived. Different importance functions are given when beneficial, Rao-Blackwellization is proposed when possible, and dual/nondual versions of each algorithms are presented. The method shown can handle both white and colored noise. Using a variety of speech and noise signals and different objective quality measures, the performance of these algorithms are evaluated against other PF solutions running on linear models, as well as some traditional enhancement algorithms. A certain hierarchy in performance is established between each algorithm in the paper. Depending on the experimental conditions, the best-performing algorithms are a classical Rao-Blackwellized particle filter (RBPF) running on a linear model, and a proposed PF employing a nondual, nonlinear model with multiple neurons and no biases. With consistence, the neural-network-based PF outperforms RBPF at low signal-to-noise ratio (SNR).

The objective of this paper is to introduce a neural-network-based algorithm for word clustering as an extension of the neural-network-based Boolean factor analysis algorithm (Frolov , 2007). It is shown that this extended algorithm supports even the more complex model of signals that are supposed to be related to textual documents. It is hypothesized that every topic in textual data is characterized by a set of words which coherently appear in documents dedicated to a given topic. The appearance of each word in a document is coded by the activity of a particular neuron. In accordance with the Hebbian learning rule implemented in the network, sets of coherently appearing words (treated as factors) create tightly connected groups of neurons, hence, revealing them as attractors of the network dynamics. The found factors are eliminated from the network memory by the Hebbian unlearning rule facilitating the search of other factors. Topics related to the found sets of words can be identified based on the words' semantics. To make the method complete, a special technique based on a Bayesian procedure has been developed for the following purposes: first, to provide a complete description of factors in terms of component probability, and second, to enhance the accuracy of classification of signals to determine whether it contains the factor. Since it is assumed that every word may possibly contribute to several topics, the proposed method might be related to the method of fuzzy clustering. In this paper, we show that the results of Boolean factor analysis and fuzzy clustering are not contradictory, but complementary. To demonstrate the capabilities of this attempt, the method is applied to two types of textual data on neural networks in two different languages. The obtained topics and corresponding words are at a good level of agreement despite the fact that identical topics in Russian and English conferences contain different sets of keywords.

This paper gives a detailed analysis of the error surfaces of certain recurrent networks and explains some difficulties encountered in training recurrent networks. We show that these error surfaces contain many spurious valleys, and we analyze the mechanisms that cause the valleys to appear. We demonstrate that the principle mechanism can be understood through the analysis of the roots of random polynomials. This paper also provides suggestions for improvements in batch training procedures that can help avoid the difficulties caused by spurious valleys, thereby improving training speed and reliability.

"For original paper see O. De Jesus, .ibid., vol. 18, no. 1, p.14-27,(2007)",. In a recent paper, De Jesus proposed a general framework for describing dynamic neural networks. Gradient and Jacobian calculations were discussed based on backpropagation-through-time (BPTT) algorithm and real-time recurrent learning (RTRL). Some errors in the paper of De Jesus bring difficulties for other researchers who want to implement the algorithms. This comments paper shows the critical parts of the publication and gives errata to facilitate understanding and implementation.

Aging effects, environmental changes, thermal drifts, and soft and hard faults affect physical systems by changing their nature and behavior over time. To cope with a process evolution adaptive solutions must be envisaged to track its dynamics; in this direction, adaptive classifiers are generally designed by assuming the stationary hypothesis for the process generating the data with very few results addressing nonstationary environments. This paper proposes a methodology based on $k$ -nearest neighbor (NN) classifiers for designing adaptive classification systems able to react to changing conditions just-in-time (JIT), i.e., exactly when it is needed. $k$ -NN classifiers have been selected for their computational-free training phase, the possibility to easily estimate the model complexity $k$ and keep under control the computational complexity of the classifier through suitable data reduction mechanisms. A JIT classifier requires a temporal detection of a (possible) process deviation (aspect tackled in a companion paper) followed by an adaptive management of the knowledge base (KB) of the classifier to cope with the process change. The novelty of the proposed approach resides in the general framework supporting the real-time update of the KB of the classification system in response to novel information coming from the process both in stationary conditions (accuracy improvement) and in nonstationary ones (process tracking) and in providing a suitable estimate of $k$ . It is shown that the classification system grants consistency once the change targets the process generating the data in a new stationary state, as it is the case in many real applications.

In this paper, dynamic output feedback control problem is investigated for a class of nonlinear interconnected systems with time delays. Decentralized observer independent of the time delays is first designed. Then, we employ the bounds information of uncertain interconnections to construct the decentralized output feedback controller via backstepping design method. Based on Lyapunov stability theory, we show that the designed controller can render the closed-loop system asymptotically stable with the help of the changing supplying function idea. Furthermore, the corresponding decentralized control problem is considered under the case that the bounds of uncertain interconnections are not precisely known. By employing the neural network approximation theory, we construct the neural network output feedback controller with corresponding adaptive law. The resulting closed-loop system is stable in the sense of semiglobal boundedness. The observers and controllers constructed in this paper are independent of the time delays. Finally, simulations are done to verify the effectiveness of the theoretic results obtained.

The phenomenon of self-organization has been of special interest to the neural network community throughout the last couple of decades. In this paper, we study a variant of the self-organizing map (SOM) that models the phenomenon of self-organization of the particles forming a string when the string is tightened from one or both of its ends. The proposed variant, called the string tightening self-organizing neural network (STON), can be used to solve certain practical problems, such as computation of shortest homotopic paths, smoothing paths to avoid sharp turns, computation of convex hull, etc. These problems are of considerable interest in computational geometry, robotics path-planning, artificial intelligence (AI) (diagrammatic reasoning), very large scale integration (VLSI) routing, and geographical information systems. Given a set of obstacles and a string with two fixed terminal points in a 2-D space, the STON model continuously tightens the given string until the unique shortest configuration in terms of the Euclidean metric is reached. The STON minimizes the total length of a string on convergence by dynamically creating and selecting feature vectors in a competitive manner. Proof of correctness of this anytime algorithm and experimental results obtained by its deployment have been presented in the paper.

A common problem encountered in disciplines such as statistics, data analysis, signal processing, textual data representation, and neural network research, is finding a suitable representation of the data in the lower dimension space. One of the principles used for this reason is a factor analysis. In this paper, we show that Hebbian learning and a Hopfield-like neural network could be used for a natural procedure for Boolean factor analysis. To ensure efficient Boolean factor analysis, we propose our original modification not only of Hopfield network architecture but also its dynamics as well. In this paper, we describe neural network implementation of the Boolean factor analysis method. We show the advantages of our Hopfield-like network modification step by step on artificially generated data. At the end, we show the efficiency of the method on artificial data containing a known list of factors. Our approach has the advantage of being able to analyze very large data sets while preserving the nature of the data.

This paper deals with problems of stability and the stabilization of discrete-time neural networks. Neural structures under consideration belong to the class of the so-called locally recurrent globally feedforward networks. The single processing unit possesses dynamic behavior. It is realized by introducing into the neuron structure a linear dynamic system in the form of an infinite impulse response filter. In this way, a dynamic neural network is obtained. It is well known that the crucial problem with neural networks of the dynamic type is stability as well as stabilization in learning problems. The paper formulates stability conditions for the analyzed class of neural networks. Moreover, a stabilization problem is defined and solved as a constrained optimization task. In order to tackle this problem two methods are proposed. The first one is based on a gradient projection (GP) and the second one on a minimum distance projection (MDP). It is worth noting that these methods can be easily introduced into the existing learning algorithm as an additional step, and suitable convergence conditions can be developed for them. The efficiency and usefulness of the proposed approaches are justified by using a number of experiments including numerical complexity analysis, stabilization effectiveness, and the identification of an industrial process.

Recognition of temporal/dynamical patterns is among the most difficult pattern recognition tasks. In this paper, based on a recent result on deterministic learning theory, a deterministic framework is proposed for rapid recognition of dynamical patterns. First, it is shown that a time-varying dynamical pattern can be effectively represented in a time-invariant and spatially distributed manner through deterministic learning. Second, a definition for characterizing similarity of dynamical patterns is given based on system dynamics inherently within dynamical patterns. Third, a mechanism for rapid recognition of dynamical patterns is presented, by which a test dynamical pattern is recognized as similar to a training dynamical pattern if state synchronization is achieved according to a kind of internal and dynamical matching on system dynamics. The synchronization errors can be taken as the measure of similarity between the test and training patterns. The significance of the paper is that a completely dynamical approach is proposed, in which the problem of dynamical pattern recognition is turned into the stability and convergence of a recognition error system. Simulation studies are included to demonstrate the effectiveness of the proposed approach.

The face is a complex multidimensional visual model and developing a computational model for face recognition is difficult. In this paper, we present a method for face recognition based on parallel neural networks. Neural networks (NNs) have been widely used in various fields. However, the computing efficiency decreases rapidly if the scale of the NN increases. In this paper, a new method of face recognition based on fuzzy clustering and parallel NNs is proposed. The face patterns are divided into several small-scale neural networks based on fuzzy clustering and they are combined to obtain the recognition result. In particular, the proposed method achieved a 98.75 % recognition accuracy for 240 patterns of 20 registrants and a 99.58% rejection rate for 240 patterns of 20 nonregistrants. Experimental results show that the performance of our new face-recognition method is better than those of the backpropagation NN (BPNN) system, the hard c-means (HCM) and parallel NNs system, and the pattern-matching system.