By Topic

Neural Information Processing, 2002. ICONIP '02. Proceedings of the 9th International Conference on

Date 18-22 Nov. 2002

Go

Filter Results

Displaying Results 1 - 25 of 115
  • Segmentation of pathological features in MRI brain datasets

    Publication Year: 2002 , Page(s): 2673 - 2677 vol.5
    Cited by:  Patents (1)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (670 KB) |  | HTML iconHTML  

    One of the major clinical applications of magnetic resonance imaging (MRI) is to detect pathological features in human body parts. While results are available in a digital format, their evaluation is performed by a trained human observer, which is still considered as the "gold standard". However, providing additional quantitative figures (e.g., lesion size or count) is tedious for a human and may better be obtained from automatical image processing methods. Three example brain lesion types (as revealed by MRI) and methods for their detection are described. Special emphasis is led on the way prior knowledge about the specific lesion type is incorporated in the algorithm. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A reinforcement learning algorithm for neural networks with incremental learning ability

    Publication Year: 2002 , Page(s): 2566 - 2570 vol.5
    Cited by:  Papers (2)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (536 KB) |  | HTML iconHTML  

    When neural networks are used for approximating action-values of Reinforcement Learning (RL) agents, the "interference" caused by incremental learning can be serious. To solve this problem, in this paper, a neural network model with incremental learning ability was applied to RL problems. In this model, correctly acquired input-output relations are stored into long-term memory, and the memorized data are effectively recalled in order to suppress the interference. In order to evaluate the incremental learning ability, the proposed model was applied to two problems: Extended Random-Walk Task and Extended Mountain-Car Task. In these tasks, the working space of agents is extended as the learning proceeds. In the simulations, we certified that the proposed model could acquire proper action-values as compared with the following three approaches to the approximation of action-value functions: tile coding, a conventional neural network model and the previously proposed neural network model. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Learning to control a joint driven double inverted pendulum using nested actor/critic algorithm

    Publication Year: 2002 , Page(s): 2610 - 2614 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (507 KB) |  | HTML iconHTML  

    In recent years, 'Reinforcement Learning' which can acquire reflective and adaptive actions, is becoming the center of attention as a learning method for robotics control. However, there are many unsolved problems that have to be cleared in order to put the method into practical use. One of the problems is the handling of the state space and the action space. Many algorithms of existing reinforcement learning deal with discrete state space and action space. When the unit of search space is rough, a subtle control cannot be achieved (imperfect perception). On the contrary, when the unit of search space is too fine, searching space is enlarged accordingly and the stable convergence of learning cannot be obtained (curse of dimensionality). In this paper, we propose a nested actor/critic algorithm that can deal with the continuous state and action space. The method proposed in this paper inserts a child actor/critic into the actor part of parent actor/critic algorithm. We examined the proposed algorithm for a stable control problem in both simulation and prototype model of a joint-driven double inverted pendulum. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Growing hierarchical self organising map (GHSOM) toolbox: visualisations and enhancements

    Publication Year: 2002 , Page(s): 2537 - 2541 vol.5
    Cited by:  Papers (7)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (640 KB) |  | HTML iconHTML  

    The Growing Hierarchical Self Organising Map (GHSOM) presents a method of dynamically modeling the data set that is presented. To a certain extent the GHSOM provides a solution to determine the size of the SOM needed, which is done through a growing fashion of neurons. In our development of the GHSOM Toolbox for Matlab presented in this paper, we have discovered that the GHSOM algorithm also provides a visualisation advantage of having the ability of presenting classes and sub-classes of similar data. We also propose two enhancements to the algorithm: (1) Usage of cumulative quantisation errors for better resolution in the growth process and (2) Tidier algorithm for initialisation of sub layer neurons for orientation. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Nonlinear Phillips curves in the Euro area and USA? Evidence from linear and neural networks models

    Publication Year: 2002 , Page(s): 2142 - 2146 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (511 KB) |  | HTML iconHTML  

    This paper applies neural network methodology to inflation forecasting in the Euro-area and the USA. Neural network methodology outperforms linear forecasting methods for the Euro Area at forecast horizons of one, three, and six month horizons, while the linear model is preferable for US data. The nonlinear estimation shows that unemployment is a significant predictor of inflation for the Euro Area. Neither model detects a significant effect of unemployment on inflation for the US data. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Artificial neural network approach in determining voltage stability in power system networks

    Publication Year: 2002 , Page(s): 2304 - 2307 vol.5
    Cited by:  Papers (1)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (405 KB) |  | HTML iconHTML  

    Voltage stability problems have been one of the major concerns for electric utilities as a result of system heavy loading. As electric power systems are operated under increasingly stressed conditions, the ability to maintain voltage stability becomes a growing concern. This paper reports on an investigation on the application of artificial neural networks (ANNs) in voltage stability assessment. A multilayer feedforward ANN with error back propagation learning is proposed for calculation of voltage stability index (L). Extensive testing of the proposed ANN based approach indicates its viability for power system voltage assessment. Test results are presented on two sample power systems. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Research on correct convergence of the EM algorithm for Gaussian mixtures

    Publication Year: 2002 , Page(s): 2660 - 2664 vol.5
    Cited by:  Papers (2)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (484 KB) |  | HTML iconHTML  

    In this paper, we present a theoretical analysis on the correct convergence of expectation-maximization algorithm for Gaussian mixtures. We first introduce the expectation-maximization algorithm and its general convergence properties. We also give a variation of the expectation-maximization algorithm for Gaussian mixtures. We then prove that the expectation-maximization algorithm becomes a compact mapping in certain neighborhood of a consistent solution when a measure of the average overlap of Gaussian in the mixtures is small enough while the sample size is large enough. We further obtain and prove the condition of the correct convergence of it. And finally, we demonstrate the theoretical results by the simulation. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Hybrid HMM-NN for speech recognition and prior class probabilities

    Publication Year: 2002 , Page(s): 2391 - 2395 vol.5
    Cited by:  Papers (2)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (496 KB) |  | HTML iconHTML  

    During the last years, speech recognition technologies have started their migration from research laboratories to real word applications gaining market shares. Although this shows that paradigms like Neural Networks have reached a high level of accuracy in modeling speech, it must be realized that there is still room for improving recognition performances exploiting the feedbacks coming from the applicative fields. In these cases, in fact, precious application dependent speech material can be recorded, and used to train the acoustic models in order to improve the behaviour of the recognizer on target dictionaries. The best results can be achieved when an iterative, refining process is set up. Unfortunately, speech corpora coming from the field are seldom phonetically balanced and this can cause the performances of the Neural Network to get worse, wasting the benefits of the refining process. In this paper, the problem of Prior Probability normalization has been faced and a method for Prior Probability normalization has been investigated, with the important characteristic of being applicable simply through a modification of the biases at the end of the training phase (therefore on trained nets). An experimentation on several languages is reported, showing the Prior Probability normalization seems quite useful to improve recognition accuracy and to get rid of some undesired effects of training data-bases not perfectly phonetically balanced. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Constructing error correcting output coding classifiers

    Publication Year: 2002 , Page(s): 2635 - 2639 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (527 KB) |  | HTML iconHTML  

    The topic of this paper is a special family of classifiers known as error correcting output coding (ECOC) classifiers. These are one of ensemble methods which prepare redundant discriminant functions and then construct a classifier by combining these discriminant functions. We focus on a method for combining the discriminant functions and develop a extension of the ECOC classifiers. Experiments on artificial data demonstrate that a much better performance can be obtained by the new classifier. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Accelerated reinforcement learning control using modified CMAC neural networks

    Publication Year: 2002 , Page(s): 2575 - 2578 vol.5
    Cited by:  Papers (3)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (417 KB) |  | HTML iconHTML  

    Reinforcement learning is a class of model-free learning control methods that can solve Markov decision problems. One difficulty for the application of reinforcement learning control is its slow convergence, especially in MDPs with continuous state space. In this paper, a modified structure of CMAC neural networks is proposed to accelerate reinforcement learning control. The modified structure is designed by incorporating a priori knowledge of learning control problems so that the efficiency and generalization ability of reinforcement learning can be improved. Simulation results on the cart-pole balancing problem illustrate the effectiveness of the proposed method. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Neural network algorithm for solving ray-tracing problem

    Publication Year: 2002 , Page(s): 2665 - 2668 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (419 KB) |  | HTML iconHTML  

    This work is dedicated to the study of neural network method for solving of ray-tracing task, which appears in 3D visualization algorithms. Physical representation of the task is the problem of finding the nearest point of the "vision" ray crossing with the surfaces of the scene. Application: Real time 3D visualization, rendering of the complex scenes, containing semitransparent, reflecting, diffusive objects, soft shadows and volume light sources. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • On data based learning using support vector clustering

    Publication Year: 2002 , Page(s): 2516 - 2521 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (636 KB) |  | HTML iconHTML  

    This paper addresses the effect of applying clustering algorithms, based on a distance metric rule, prior to support kernel learning in classification and regression problems. Self-Organising Maps (SOMs), which place emphasis in data domain description, and K-means clustering algorithms have been selected before applying a support vector algorithm which is based on a margin rule. Moreover, the recently developed support vector clustering algorithm, based on a cluster boundary rule, is applied in benchmark problems for comparison. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Effectiveness of directional microphones and utilization of source arriving directions in source separation

    Publication Year: 2002 , Page(s): 2190 - 2194 vol.5
    Cited by:  Papers (1)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (439 KB) |  | HTML iconHTML  

    In this paper, we present the effectiveness of directional microphones in source separation and present a three-source separation method in which microphones are directed toward source directions obtained by image processing. The separation performance with directional microphones is much higher than that with omni-directional microphones. In the proposed method for more than two sources, we direct microphones toward sources estimated by flesh color extraction for source separation. The method estimates a convergence point of separating parameters based on source directions. We show that using the estimated convergence point as an initial value considerably improves convergence performance and also demonstrate the effectiveness of the proposed method in three source-separation in a reverberant environment. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Temporal differences learning with the scaled conjugate gradient algorithm

    Publication Year: 2002 , Page(s): 2625 - 2629 vol.5
    Cited by:  Papers (1)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (513 KB) |  | HTML iconHTML  

    This paper investigates the use of the scaled conjugate gradient algorithm in temporal differences learning for time series prediction more than one time interval ahead. Although neural networks trained with the traditional backpropagation (BP) algorithm are successfully applied in this area, the temporal differences (TD) methodology is potentially more applicable for multi-step predictions. A combination of TD with advanced algorithms like the scaled conjugate gradient (SCG) algorithm may prove more promising, resulting to robust learning systems. Whether, though, TD is better than supervised learning when examined with a solid training algorithm like SCG is an open issue. The results of this study indicate that the SCG algorithm, which was developed for supervised learning, cannot be directly applied in TD(λ) learning. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Associative memory neural networks for information retrieval of text word pairs

    Publication Year: 2002 , Page(s): 2200 - 2203 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (414 KB) |  | HTML iconHTML  

    Natural language information processing remains a challenge in linguistics. Existing methods for retrieval of text often use stemming to retain common roots and base recall on these root words. This requires the removal of stop words, e.g., numbers, symbols, high frequency bid low semantic weight words, thereby precluding phrases using these words. In addition stemming is a morphologic approach that cannot readily process homonyms, synonyms and certain inflectional and derived forms. Machine learning approaches assign words to categories, but application to a large corpus remains in debate. For this study, we describe the application of the Cortronic theory and methods developed by Robert Hecht-Nielsen (2002) to information retrieval of text word pairs. Hecht-Nielsen's theories build on associative memory artificial neural networks (AMNNs) introduced by Steinbuch (1961) and extended by Willshaw et al. (1969) especially in regards sparseness. The AMNNs are used to process a large corpus without excluding stop words, and retain the joint probability of mutual occurrences that allows rapid retrieval of word pairs. This AMNN approach includes three key components: representation of arbitrary objects (words), learning and knowledge accumulation based on measurement of co-occurrence and use of all the learned knowledge to produce (predict) the missing word in a phrase. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • DevLex: a self-organizing neural network model of the development of lexicon

    Publication Year: 2002 , Page(s): 2546 - 2551 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (586 KB) |  | HTML iconHTML  

    In this paper we present the DevLex model of language acquisition. DevLex consists of two self-organizing maps (a growing semantic map and a phonological map) that are connected via associative links. It simulates the early stages of lexical development in children, in particular, word confusion as evidenced in naming errors. The simulation results indicate that the rate of word confusion is modulated by developmental profile of vocabulary increase, word density of competing neighbors, and rate of lexical growth. These results match up with hypotheses from empirical research on lexical development. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Time-series data prediction based on reconstruction of missing samples and selective ensembling of FIR neural networks

    Publication Year: 2002 , Page(s): 2152 - 2156 vol.5
    Cited by:  Papers (2)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (546 KB) |  | HTML iconHTML  

    This paper considers the problem of time-series forecasting by a selective ensemble neural network when the input data are incomplete. Five fill-in methods, viz. cubic smoothing spline interpolation, EM (Expectation maximization), regularized EM, average EM, and average regularized EM, are simultaneously employed in a first step for reconstructing the missing values of time-series data. A set of complete data from each individual fill-in method is used to train a finite impulse response (FIR) neural network to predict the time series. The outputs from individual network are combined by a selective ensemble method in the second step. Experimental results show that the prediction made by the proposed method is more accurate than those predicted by neural networks without a fill-in process or by a single fill-in process. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Analysis of natural images by independent quadratic forms and temporally coherent quadratic forms

    Publication Year: 2002 , Page(s): 2424 - 2429 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (854 KB) |  | HTML iconHTML  

    Several studies have succeeded to correlate natural image statistics with receptive field properties of neurons in the primary visual cortex such as simple cells. However, complex cell properties have not fully explained by previous studies of natural image statistics. In this study, we deal with quadratic forms. Because they form a class of functions that includes complex cell responses and many other functions. We employ two criteria for learning parameters of quadratic forms. They are independence of output and temporal coherence of output. By independence criterion, squared responses of simple cells were obtained and complex cell properties were not reproduced. On the other hand, by maximizing the sparseness of difference of output, we obtained complex cell properties among other kind of quadratic forms. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Classifying brain shapes

    Publication Year: 2002 , Page(s): 2678 - 2682 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (708 KB) |  | HTML iconHTML  

    Brain shapes do not necessarily form a continuum in some descriptor space, but may form clusters related to pre-determined genetical factors or acquired diseases. As a feasibility study for introducing a suitable descriptor space, the use of modal analysis was tested on a large brain database acquired in healthy young subjects. Significant shape differences due to gender were found, and intra-gender clusters were determined. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A pruning algorithm of neural networks using impact factor regularization

    Publication Year: 2002 , Page(s): 2605 - 2609 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (470 KB) |  | HTML iconHTML  

    In general, small-sized networks, even though they show good generalization performance, tend to fail to learn the training data within a given error bound, whereas large-sized networks learn easily the training data but yield poor generalization. In this paper, a pruning algorithm of neural networks using impact factor regularization is described to train network without overfitting and to achieve a small-sized network. In order to achieve this goal, an automatic determination method of the regularization parameter and an extended Levenberg-Marquardt algorithm are developed as learning algorithms of neural networks. We tested the proposed method on four regression problems and the simulation results showed our algorithm is effective in regression. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Sparse coding and rough set theory-based hybrid approach to the classificatory decomposition of cortical evoked potentials

    Publication Year: 2002 , Page(s): 2264 - 2268 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (515 KB) |  | HTML iconHTML  

    This paper presents a novel approach to classification of decomposed cortical evoked potentials (EPs). The decomposition is based on learning of a sparse set of basis functions using an artificial neural network (ANN). The basis functions are generated according to a probabilistic model of the data. In contrast to the traditional signal decomposition techniques (i.e. principle component analysis or independent component analysis), this allows for an overcomplete representation of the data (i.e. number of basis functions that is greater than the dimensionality of the input signals). Obviously, this can be of a great advantage. However, there arises an issue of selecting the most significant components from the whole collection. This is especially important in classification problems based upon the decomposed representation of the data, where only those components that provide a substantial discernibility between EPs of different groups are relevant. To deal with this problem, we propose an approach based on the rough set theory's (RS) feature selection mechanisms. We design a sparse coding- and RS-based hybrid system capable of signal decomposition and, based on a reduced component set, signal classification. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Hebbian learning in an automatic gender identification by speech system

    Publication Year: 2002 , Page(s): 2409 - 2413 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (404 KB) |  | HTML iconHTML  

    This work presents an Automatic Gender Identification (AGI) algorithm based on Eigenfiltering. A Maximum Eigenfilter is implemented by means of an Artificial Neural Network (ANN) trained via Generalized Hebbian Learning (GHL). The Eigenfilter uses Principal Component Analysis (PCA) to perform maximum information extraction from the speech signal, which enhances correlated information and improves the pattern analysis. Also, a well known speech processing technique is applied, the Mel-Frequency Cepstral Coefficients (MFCC). This technique is a classical approach for speech feature extraction, and it is a very efficient way to represent physiological voice parameters. The pattern classification uses a Radial Basis Function (RBF) ANN. Experimental results have shown that the identification algorithm overall performance was widely increased by the Eigenfiltering process. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A services-oriented architecture applied to artificial neural network

    Publication Year: 2002 , Page(s): 2650 - 2654 vol.5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (546 KB) |  | HTML iconHTML  

    In this work, a services-oriented architecture that allows the utilization of artificial neural network models, through the Internet, is proposed. The Web Services, the Message Queuing and the Neural Network Markup Language constitute the technology used for the development of this approach. The module called by Artificial Neural Network-Web Service is the main module of the proposed architecture. It is an independent module, that accepts as entrance a description of the ANN characteristics and the necessary data to its training in a neural network markup language format returning as result the trained ANN. A main advantage of this approach is that any user can invoke this application programmatically over the Internet, without that it needs to have a computer with great capacity of processing. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A system for fingerprint minutiae classification and recognition

    Publication Year: 2002 , Page(s): 2474 - 2478 vol.5
    Cited by:  Papers (5)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (491 KB) |  | HTML iconHTML  

    This paper shows a study about biometrics characteristics for recognition/classification and presents how it is used for individual recognition. The approach uses the automated fingerprint recognition based on minutia, which are extracted directly from the finger prints and the methodology used to its recognition is the artificial neural networks (ANN) based system. The neocognitron model was the ANN chosen. Inasmuch as neocognitron was originally implemented for handwritten characters recognition, it is possible to verify its usefulness for another kind of pattern recognition. Finally it is presented the results for this system and the conclusions according to the number of samples and recognition rate. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A new kernel clustering algorithm

    Publication Year: 2002 , Page(s): 2527 - 2531 vol.5
    Cited by:  Papers (1)
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (492 KB) |  | HTML iconHTML  

    We propose a new kernel clustering algorithm. It estimates an in advance fixed number of vectors and margins in a feature space. Each pair of vector and margin defines a hyperplane in feature space and thus separates the data in two clusters. All the clusters together carry important information about the data set. The estimation in feature space is done implicitly by the use of a kernel. Therefore nonlinear clusters in the space of the data can be obtained. The clusters are estimated by optimizing a homogeneous quadratic program. We show how our algorithm can be efficiently implemented and we demonstrate the usefulness with a real world example. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.