Catálogo de publicaciones - libros

We deal with a problem of cleaning up a contaminated network by mutual copying. This problem involves not only an aspect of “the double-edged sword” where copying could further spread contamination but an aspect of mutual cooperation where resource consuming copying could be left for others. The framework of “prisoner’s dilemma” has been applied, aiming at emergence of appropriate copying strategies in an adaptive manner to the network environment.

This paper reports a critical phenomenon in a self-repair network by mutual copying. Extensive studies have been done on critical phenomena in many fields such as in epidemic theory and in percolation theory with an effort of identification of critical points. However, from the viewpoints of cleaning up a network by mutual copying, critical phenomena have not much studied. A critical phenomenon has been observed in a self-repair network. Self-repairing by mutual copying is “the double-edged sword” that could cause outbreaks with inappropriate parameters, and careful investigations are needed.

We propose a new filter for preventing computer worms from spreading. The new worm filter limits the number of unacknowledged requests, rather than the rate of connections to new computers. Normal network traffic is analyzed to determine appropriate parameters for the worm filter. Performance evaluation showed that the worm filter stops not only high-speed worms in the wild, but also simulated slow-speed worms. Finally, the weaknesses of the worm filter is discussed.

This paper presents nonlinear static and dynamic system modeling using wavenet and neuralnet. Wavenet combines wavelet theory and feed-forward neuralnet, so learning approach is similar to neuralnet. The selection of transfer function is crucial for the approximation property and the convergence of the network. The purelin and the tansig functions are used as the transfer functions for neuralnet and the first derivative of a gaussian function is used as the transfer function for wavenet. Wavenet and neuralnet parameters are optimized during learning phase. Selecting all initial values random, but for wavenet, it may be unsuitable for process modeling because wavelets have localization feature. For this reason heuristic procedure has been used for wavenet. In this study gradient methods have been applied for parameters updating with momentum. Error minimization is computed by quadratic cost function for wavenet and neuralnet. Nonlinear static and dynamic functions have been used for the simulations. Recently wavenet has been used as an alternative of the neuralnet because interpretation of the model with neuralnet is so hard. For wavenet learning approach, training algorithms require smaller number of iterations when compared with neuralnet. Consequently, according to the number of training iteration and TMSE, dynamic and static system modeling with wavenet is better as shown in results.

This paper describes Neurone Editor as an editing tool that allows different types of neuronal cells to be modelled by using Java and Java3D. It provides a means of generating realistic neuronal geometries and their expected growth patterns. Each neurone type, as described by the parameters and the variance attributed to each parameter, allows the generation of multiple instances of similar neurons. In this way, we can generate large neurone networks in a biological pattern with the aim of generating computational structures of that type and also as an aid to visualize these structures. With these aims in view, detailed and anatomically correct renditions are not provided but the aim is to replicate natural growth patterns and the resulting interconnections on the assumption that shape defines function.

The learning strategy employed in neural networks offers a good performance even in the situations where a model is presented with incomplete and noisy data. However, neural networks are known as ‘black boxes’ as how the outputs are produced is not clear. In this study, a hybrid learning strategy, namely RDC-ANNE (Rules Driven by Consistency in Artificial Neural Networks Ensemble) is proposed. This paper looks at the use of RDC-ANNE in the graft outcome prediction domain as a prototypical medical application. At first, for a better generalization, a committee of binary neural networks is trained. Then, a partial C4.5 decision tree is built from a specifically selected dataset, generated based on the graft data used to test the trained neural networks ensemble. Finally the most appropriate leaf in every path is converted into an understandable rule. In this approach, for the rule generation process, we enforced the model to mainly consider the patterns that their class labels were consistently causing agreement across the neural network classifiers. Experimental results show that the RDC-ANNE method is able to extract partial rules from an ensemble model and reveal the important embedded information of a trained neural network ensemble.

This paper represents the performance comparison of the Multilayered Perceptron (MLP) networks using various back propagation (BP) algorithms for breast cancer diagnosis. The training algorithms used are gradient descent with momentum and adaptive learning, resilient back propagation, Quasi-Newton and Levenberg-Marquardt. The performances of these four algorithms are compared with the standard steepest descent back propagation algorithm. The current study investigates and compares the accuracy, sensitivity, specificity, false negative and false positive results of the selected four algorithms to train MLP networks. The Papinicolou image of breast cancer cells were captured via an image analyzer and thirteen morphological features were extracted to numerical scores. The feature scores are used as data sets to train the MLP network. The MLP network using the Levenberg-Marquardt algorithm displays the best performance for all the five measurement criteria’s (accuracy, specificity, sensitivity, true positive and true negative) at a lower number of hidden nodes.

A knowledge based system for detection of honeycombing patterns in HRCT lung images is described. In the system, rules generated by machine learning on low level image pixel-based features and heuristic rules from the domain expert on high level region-based features are combined using a generalized ripple down rules (GRDR) framework. Results demonstrate that the systems’ performance can be incrementally improved.

Past statistics have revealed that breast cancer is the world’s leading cause of death among women. One popular method of screening breast cancer is ultrasound. However, reading an ultrasound image is not an easy task because it lacks spatial resolution, subject to image distortion, susceptible to noise and is highly operator dependant. Several image processing techniques have been introduced to enhance the detection of diagnostic features. This study proposes modified seed based region growing algorithm to detect the edges and segment the area of solid masses in an ultrasound image without having to specify the location of the seed and the grey level threshold value manually. Automatic seed selection is done by using moving k -means clustering. A performance analysis has been carried out towards 3 different ultrasound images. The results reveal that this algorithm can detect the edges of solid masses and segment it from the rest of the image effectively.

This paper presents an efficient method to determine a body region-of-interest (ROI) enclosing a fetus in a two-dimensional (2D) key frame that removes some irrelevant matters such as the abdominal area in front of a fetus to visualize a fetal ultrasound volume along with the key frame. In the body ROI determination, a clear frontal view of a fetus lying down floating in amniotic fluid mainly depends on the successful determination of the top bound among the four bounds of an ROI. The key idea of our top-bound setting is to locate it in amniotic fluid areas between a fetus and its mother’s abdomen, which are dark so as to typically induce local minima of the vertical projection of a key frame. The support vector machines (SVM) classifier, known as an effective tool for classification, tests whether the candidate top bound, located at each of the local minima which are sorted in an increasing order, is valid or not. The test, using textural characterization of neighbor regions around each candidate bound, determines the first valid one as the top bound. The body ROI determination rate as well as resulting 3D images demonstrate that our system could replace a user in allocation of a fetus for its 3D visualization.