NCTA 2012 Abstracts


Full Papers
Paper Nr: 5
Title:

Diffusion Ensemble Classifiers

Authors:

Alon Schclar, Lior Rokach and Amir Amit

Abstract: We present a novel approach for the construction of ensemble classifiers based on the Diffusion Maps (DM) dimensionality reduction algorithm. The DM algorithm embeds data into a low-dimensional space according to the connectivity between every pair of points in the ambient space. The ensemble members are trained based on dimension-reduced versions of the training set. These versions are obtained by applying the DM algorithm to the original training set using different values of the input parameter. In order to classify a test sample, it is first embedded into the dimension reduced space of each individual classifier by using the Nyström out-of-sample extension algorithm. Each ensemble member is then applied to the embedded sample and the classification is obtained according to a voting scheme. A comparison is made with the base classifier which does not incorporate dimensionality reduction. The results obtained by the proposed algorithms improve on average the results obtained by the non-ensemble classifier.

Paper Nr: 7
Title:

Extension of Backpropagation through Time for Segmented-memory Recurrent Neural Networks

Authors:

Stefan Glüge, Ronald Böck and Andreas Wendemuth

Abstract: We introduce an extended Backpropagation Through Time (eBPTT) learning algorithm for Segmented-Memory Recurrent Neural Networks. The algorithm was compared to an extension of the Real-Time Recurrent Learning algorithm (eRTRL) for these kind of networks. Using the information latching problem as benchmark task, the algorithms’ ability to cope with the learning of long-term dependencies was tested. eRTRL was generally better able to cope with the latching of information over longer periods of time. On the other hand, eBPTT guaranteed a better generalisation when training was successful. Further, due to its computational complexity, eRTRL becomes impractical with increasing network size, making eBPTT the only viable choice in these cases.

Paper Nr: 13
Title:

Geometric Image of Neurodynamics

Authors:

Germano Resconi and Robert Kozma

Abstract: We know that the brain is composed of simple neural units given by dendrites, soma, and axons. Every neural unit can be modelled by electrical circuits with capacitors and adaptive resistors. To study the neural dynamic we use special Ordinary Differential Equations (ODE) whose solutions give us the behaviour or trajectory of the neural states in time. The problem with ODE is in the definition of the parameters and in the complexity of the solutions that in many cases cannot be found. The key elements that we use are the multidimensional vector spaces of the electrical charges, currents and voltages. So currents and voltages are geometric references for states in the central neural system (CNS). Any neuro –biological architecture can be modelled by an adaptive electrical circuit or neuromorphic network that relates voltage with current by conductance matrix or on the contrary by impedance matrix. Given a straight line with a change of reference we reshape the straight line in a geodetic and in a new form for the distance. The change of the reference transforms a set of variables into another so this transformation is similar to a statement in the digital computer that we associate to the software. Every change of variables can be reproduced by a similar change of voltages (currents) into currents (voltages) by conductance (impedance) matrix. We use the CNS as a material support or hardware in the digital computer to realise the wanted transformation. In conclusion geometry fuses the digital computer structure with neuromorphic computing to give efficient computation where conceptual intention is the change of the reference space , while material intention is given by the neurodynamical processes modelled by the change of the electrical charge space where we define the metric geometry and distance.

Paper Nr: 16
Title:

Artificial Intelligence Methods in Reactive Navigation of Mobile Robots Formation

Authors:

Zenon Hendzel, Marcin Szuster and Andrzej Burghardt

Abstract: The article presents a hierarchical control system build using artificial intelligence methods, that generates a trajectory of the wheeled mobile robots formation, and realises the tracking control task of all agents. The hierarchical control system consists of a navigator, based on a conception of behavioural control signals coordination, and individual tracking control systems for all mobile robots in the formation. The navigator realises a sensor-based approach to the path planning process in the unknown 2-D environment with static obstacles. The navigator presents a new approach to the behavioural control, where one Neural dynamic programming algorithm generates the control signal for the complex behaviour, which is a composition of two individual behaviours: “goal-seeking”and “obstacle avoiding”. Influence of individual behaviours on the navigator control signal depends on the environment conditions and changes fluently. On the basis of control signal generated by the navigator are computed the desired collision-free trajectories for all robots in formation, realised by the tracking control systems. Realisation of generated trajectories guarantees reaching the goal by selected point of the robots formation with obstacles avoiding by all agents. Computer simulations have been conducted to illustrate the process of path planning in the different environment conditions.

Paper Nr: 24
Title:

Adaptive Sequential Feature Selection for Pattern Classification

Authors:

Liliya Avdiyenko, Nils Bertschinger and Juergen Jost

Abstract: Feature selection helps to focus resources on relevant dimensions of input data. Usually, reducing the input dimensionality to the most informative features also simplifies subsequent tasks, such as classification. This is, for instance, important for systems operating in online mode under time constraints. However, when the training data is of limited size, it becomes difficult to define a single small subset of features sufficient for classification of all data samples. In such situations, one should select features in an adaptive manner, i.e. use different feature subsets for every testing sample. Here, we propose a sequential adaptive algorithm that for a given testing sample selects features maximizing the expected information about its class. We provide experimental evidence that especially for small data sets our algorithm outperforms two the most similar information-based static and adaptive feature selectors.

Paper Nr: 32
Title:

Combined Input Training and Radial Basis Function Neural Networks based Nonlinear Principal Components Analysis Model Applied for Process Monitoring

Authors:

Messaoud Bouakkaz and Mohamed-Faouzi Harkat

Abstract: In this paper a novel Nonlinear Principal Component Analysis (NLPCA) is proposed. Generally, a NLPCA model is performed by using two sub-models, mapping and demapping. The proposed NLPCA model consists of two cascade three-layer neural networks for mapping and demapping, respectively. The mapping model is identified by using a Radial Basis Function (RBF) neural networks and the demapping is performed by using an Input Training neural networks (IT-Net). The nonlinear principal components, which represents the desired output of the first network, are obtained by the IT-NET. The proposed approach is illustrated by a simulation example and then applied for fault detection and isolation of the TECP process.

Paper Nr: 54
Title:

MR Damper Identification using ANN based on 1-Sensor - A Tool for Semiactive Suspension Control Compliance

Authors:

Juan C. Tudón-Martínez, Ruben Morales-Menendez, Ricardo A. Ramirez-Mendoza and Luis E. Garza-Castañón

Abstract: A model for a Magneto-Rheological (MR) damper based on Artifical Neural Networks (ANN) is proposed. The ANN model does not require regressors in the input and output vector, i.e. is considered static. Only one sensor is used to achieve a reliable MR damper model which is compared with experimental data provided from two MR dampers with different properties. The RMS of the error is used to measure the model accuracy; from both MR dampers, an average value of 7.1% of total error in the force signal is obtained by taking into account 5 different experiments. The ANN model, which represents the nonlinear behavior of an MR damper, is used in a suspension control system of a Quarter of Vehicle (QoV) in order to evaluate the comfort of passengers maintaining the road holding. A control technique with the MR damper model is compared with a passive suspension system. Simulation results show the effectiveness of a semiactive suspension versus the passive one. The RMS of the comfort signal improves 7.4% with the MR damper while the road holding gain in the frequency response shows that the safety in the vehicle can be increased until 40.4% with the semiactive suspension system. The accurate MR damper model validates a realistic QoV response compliance.

Short Papers
Paper Nr: 17
Title:

Adaptive Neural Network Control of Underactuated System

Authors:

Andrzej Burghardt and Zenon Hendzel

Abstract: The article presents a synthesis of the control system of an underactuated object of ball-beam type. Based on a mathematical description of the object, we proposed an adaptational control algorithm, ensuring stabilization of the ball position on the beam. The synthesis of the control system was conducted on the basis of Lyapunov’s stability theory, using artificial neural networks in the adaptation process. The proposed solution was simulated with Matlab/Simulink software and verified on the real object.

Paper Nr: 27
Title:

A Binary Neural Network Framework for Attribute Selection and Prediction

Authors:

Victoria J. Hodge, Tom Jackson and Jim Austin

Abstract: In this paper, we introduce an implementation of the attribute selection algorithm, Correlation-based Feature Selection (CFS) integrated with our k-nearest neighbour (k-NN) framework. Binary neural networks underpin our k-NN and allow us to create a unified framework for attribute selection, prediction and classification. We apply the framework to a real world application of predicting bus journey times from traffic sensor data and show how attribute selection can both speed our k-NN and increase the prediction accuracy by removing noise and redundant attributes from the data.

Paper Nr: 28
Title:

Focused Image Color Quantization using Magnitude Sensitive Competitive Learning Algorithm

Authors:

Enrique Pelayo, David Buldain and Carlos Orrite

Abstract: This paper introduces the Magnitude Sensitive Competitive Learning (MSCL) algorithm for Color Quantization. MSCL is a neural competitive learning algorithm, including a magnitude function as a factor of the measure used for the neuron competition. This algorithm has the property of distributing color vector prototypes in certain data-distribution zones according to an arbitrary magnitude locally calculated for every unit. Therefore, it opens the possibility not only to distribute the codewords (colors of the palette) according to their frequency, but also to do it in function of any other data-dependent magnitude focused on a given task. This work shows some examples of focused Color Quantization where the objective is to represent with high detail certain regions of interest in the image (salient area, center of the image, etc.). The oriented behavior of MSCL permits to surpass other standard Color Quantization algorithms in these tasks.

Paper Nr: 33
Title:

FPGA Implementation of Hodgkin-Huxley Neuron Model

Authors:

Safa Yaghini Bonabi, Hassan Asgharian, Reyhaneh Bakhtiari, Saeed Safari and Majid Nili Ahmadabadi

Abstract: In this paper an implementation of Hodgkin-Huxley single neuron is provided. Unlike almost all of the existing implementations, the arithmetic logics are implemented with computation techniques (i.e. CORDIC) and look-up-tables (LUTs) are used only in few modules. This makes our design more robust and flexible to simulate the functionality of a large network of neurons. Most of the previous works are based on the software implementations which overshadow the parallel nature of the neural system or using LUTs for hardware implementation which needs more space and also limited flexibility. In this paper, an FPGA is selected as our hardware implementation platform to provide an appropriate reconfigurable platform for simulating the functionality of a network of neurons. We validated our design based on our high level implementation of Hodgkin-Huxley neuron in MATLAB and report our implementation results based on Xilinx SPARTAN 3 FPGA in Xilinx ISE Design Suite.

Paper Nr: 34
Title:

Contradiction Resolution for Foreign Exchange Rates Estimation

Authors:

Ryotaro Kamimura

Abstract: In this paper, we propose a new type of information-theoretic method called ”contradiction resolution.” In this method, we suppose that a neuron should be evaluated for itself (self-evaluation) and by all the other neurons (outer-evaluation). If some difference or contradiction between two types of evaluation can be found, the contradiction should be decreased as much as possible. We applied the method to the self-organizing maps with an output layer, which is a kind of combination of the self-organizing maps with the RBF networks. When the method was applied to the dollar-yen exchange rates, prediction and visualization performance could be improved simultaneously.

Paper Nr: 36
Title:

Enhancing the Accuracy of Mapping to Multidimensional Optimal Regions using PCA

Authors:

Elham Bavafaye Haghighi and Mohammad Rahmati

Abstract: Mapping to Multidimensional Optimal Regions (M2OR) is a special purposed method for multiclass classification task. It reduces computational complexity in comparison to the other concepts of classifiers. In order to increase the accuracy of M2OR, its code assignment process is enriched using PCA. In addition to the increment in accuracy, corresponding enhancement eliminates the unwanted variance of the results from the previous version of M2OR. Another advantage is more controllability on the upper bound of V.C. dimension of M2OR which results in a better control on its generalization ability. Additionally, the computational complexity of the enhanced-optimal code assignment algorithm is reduced in training phase. By the other side, partitioning the feature space in M2OR is an NP hard problem. PCA plays a key role in the greedy feature selection presented in this paper. Similar to the new code assignment process, corresponding greedy strategy increases the accuracy of the enhanced M2OR.

Paper Nr: 39
Title:

Recovery of Sequential and Non Sequential Memories with a Neural Mass Model

Authors:

Filippo Cona and Mauro Ursino

Abstract: A neural model for the recovery of learnt patterns is presented. The model simulates the theta-gamma activity associated to memory recall. Two versions of the model are described: the first can learn generic patterns without a given order, while the second learns patterns in a specific sequence. The latter has been implemented to overcome the limited recovery capacity of the former. The network is trained using Hebbian and anti-Hebbian paradigms, and exploits excitatory and inhibitory mutual synapses. The results show that autoassociative memories for storage and recovery of multiple patterns can be built using biologically inspired models which simulate brain rhythms, and that the model which learns sequences can recover much more patterns.

Paper Nr: 48
Title:

Approximation of Geometric Structures with Growing Cell Structures and Growing Neural Gas - A Performance Comparison

Authors:

Hendrik Annuth and Christian-A. Bohn

Abstract: We compare Growing Cell Structures and Growing Neural Gas, which were introduced by Bernd Fritzke and which are famous for their facilities in classification, clustering, dimensionality reduction, data visualization, and approximation tasks. We practically test and analyze their capabilities in geometric approximation and focusing on the application of surface reconstruction from 3D point-data. Our focus is to work out the differences of the algorithms that are especially relevant concerning approximation purposes. We address the issue of suitable input data, their applied graphs, their topological properties, their run time complexities and we present a summary of suggested alternations to both approaches and evaluate our results.

Paper Nr: 50
Title:

Modeling of River Water Temperatures using Feed-forward Artificial Neural Networks

Authors:

Cindie Hébert, Daniel Caissie, Mysore G. Satish and Nassir El-Jabi

Abstract: Water temperature influences most physical, chemical and biological processes of the river environment. It plays an important role in the distribution of fishes and on the growth rates of many aquatic organisms. It is therefore important to develop water temperature models in order to effectively manage aquatic habitats, to study the thermal regime of rivers and to have effective tools for environmental impact studies. The objective of the present study was to develop a water temperature model based on artificial neural networks (ANN) for two thermally different watercourses. The ANN model performed best in summer and autumn and showed a poorer (but still good) performance in spring. The many advantages of ANN models are their simplicity, low data requirements, their capability of modelling long-term series as well as have an overall good performance.

Paper Nr: 53
Title:

On Modelling Cognitive Styles of Users in Adaptive Interactive Systems using Artificial Neural Networks

Authors:

Efi Papatheocharous, Marios Belk, Panagiotis Germanakos and George Samaras

Abstract: User modelling in interactive Web systems is an essential quality to optimally filter, personalise and adapt their content and functionality to serve the intrinsic needs of individual users. The mechanism for obtaining the user model needs to be intelligent, adaptive and transparent to the user, in the sense that user experience should not be disrupted or compromised. Human factors are extensively employed lately for enriching user models by capturing more intrinsic perceptual characteristics of the users. Accordingly, this paper proposes the use of Artificial Neural Networks (ANNs) for attaining cognitive styles of users in adaptive interactive systems. One of the main benefits is the automatic prediction of cognitive typologies of users by avoiding psychometric tests, which are among the typical ways of constructing user profiles and are particularly time-consuming. Furthermore, ANNs can efficiently model the relationship between cognitive styles and user interaction. The experimental setup and the results obtained show that ANNs are suitable for predicting the cognitive styles ratio of users in respect to their actual cognitive style ratio value.

Paper Nr: 62
Title:

Sliding Global Attractors of Neural Learning and Memory

Authors:

Yoram Baram

Abstract: The highly variable nature of neural firing has been recognized by diverse empirical and analytic findings. Here, the underlying morphology of neural firing is shown to be governed by a bilinear map, prescribing eight types of neuronal global attractors and their points of local bifurcation. While synaptic learning gives rise to irregular firing, membrane memory is shown to guarantee that, under the same external activation, learning and retrieval end at the same global attractor. Forced and spontaneous changes in membrane conductance are shown to cause sliding of the global attractors, switching them from passive to active state and vice versa, and creating secondary firing modes. Selective activation of interacting neurons is shown to create a shunting effect, yielding combinatorial retrieval, concealment and revelation of stored global attractors. The utility of the global attractors is explained not only by their individual dynamic characteristics, but also by their high power of combinatorial expression.

Paper Nr: 63
Title:

Hopfield Neural Network for Microscopic Evacuation of Buildings

Authors:

Boutheina Amina Aoun, Zouhour Neji Ben Salem and Hend Bouziri

Abstract: The problem of evacuation raised a lot of interest as its objective of saving lives is of an extreme importance. In this context, many researches supplied solutions allowing to plan the process of evacuation in case of disaster. Certain solutions took into account the behavior of the crowd, while others treated the evacuees in an independent way. For that purpose, we dedicate our study to this last type of evacuation, namely the microscopic evacuation. Our approach is based on the artificial neural networks which we considered capable of generating a human behavior thanks to their neuronal aspect. We proposed a solution capable of planning a microscopic evacuation of building by having recourse to Hopfield neural networks. We supplied an experimental study on the real cases of two hospitals. This study also brought a comparison of our model with another neuronal model for evacuation which is the self organizing map.

Paper Nr: 65
Title:

Attractor Neural Networks for Simulating Dyslexic Patients’ Behavior

Authors:

Shin-ichi Asakawa

Abstract: It was investigated that the ability of an attractor neural network. The attractor neural network can be applicable to various symptoms of brain damaged patients. It can account for delays in reaction times in word reading and word identification tasks. Because the iteration numbers of mutual connections between an output and a cleanup layers might increase, when they are partially damaged. This prolongation looks or behaves the delays of reaction times of brain damaged patients. When we applied the attractor neural network to the data of Tyler et al. (2000) for categorization task, it showed a kind of category specific phenomenon. In this sense, the attractor neural network could explain an aspect of the category specific disorders. In this sense the attractor network might simulate the human semantic memory organization. In spite of variations in data, and in spite of the simplicity of the architecture, the attractor network showed good performances. We could say that the attractor network succeeded in mimicking human normal subjects and brain damaged patients. The possibility of explaining the triangle model (Plaut & McClelland,1989; Plaut, McClelland, Seidenberg, and Patterson, 1996) also discussed.

Paper Nr: 66
Title:

Single Input Single Output Time Series Artificial Neural Network Models for Free Residual Chlorine Forecasting in Water Distribution Networks

Authors:

Selcuk Soyupak, Hurevren Kilic, Ibrahim Ethem Karadirek and Habib Muhammetoglu

Abstract: The aim of this study is to investigate the utilization of Single Input Single Output Time Series Artificial Neural Networks models as a forecasting tool for estimating Free Residual Chlorine levels at critical locations of fairly complex Water Distribution Systems. The response surface methodology was adopted in identifying performance and precision trends as a function of number of steps used as inputs and number of steps ahead to predict (Horizons). The utilized response surfaces were for coefficient of determination and mean absolute error. The creation of response surfaces was achieved by developing Artificial Neural Network models for several combinations of number of steps used as inputs and number of steps ahead to predict that enable the calculations of coefficient of determination and mean absolute error for the selected combinations. Then these results have been assembled to obtain contour maps by distance weighted least square technique. The maximum attained coefficient of determination levels were within the range 0.656 to 0.974, while minimum achievable mean absolute error levels were within the range 0.0080 to 0.0284 ppm. The achieved mean absolute error is very low when compared with the followings: a) the applied Free Residual Chlorine levels from the source which is about 0.5 ppm and b) the minimum detection limit of the chlorine analyzers given as 0.01 ppm.

Paper Nr: 71
Title:

Recurrent Neural Networks - A Natural Model of Computation beyond the Turing Limits

Authors:

Jérémie Cabessa and Alessandro E. P. Villa

Abstract: According to the Church-Turing Thesis, the classical Turing machine model is capable of capturing all possible aspects of algorithmic computation. However, in neural computation, several basic neural models were proven to be capable of computational capabilities located beyond the Turing limits. In this context, we present an overview of recent results concerning the super-Turing computational power of recurrent neural networks, and show that recurrent neural networks provide a suitable and natural model of computation beyond the Turing limits. We nevertheless don’t draw any hasty conclusion about the controversial issue of a possible predominance of biological intelligence over the potentialities of artificial intelligence.

Posters
Paper Nr: 2
Title:

Robust Stability Analysis of a Class of Delayed Neural Networks

Authors:

Neyir Ozcan and Sabri Arik

Abstract: This paper studies the global robust stability of delayed neural networks. A new sufficient condition that ensures the existence, uniqueness and global robust asymptotic stability of the equilibrium point is presented. The obtained condition is derived by using the Lyapunov stability and Homomorphic mapping theorems and by employing the Lipschitz activation functions. The result presented establishes a relationship between the network parameters of the neural system independently of time delays. We show that our results is new and improves some of the previous global robust stability results expressed for delayed neural networks.

Paper Nr: 14
Title:

A Non-lineal Mathematical Model for Annealing Stainless Steel Coils

Authors:

Raquel González Corral, J. Bonelo Sánchez, Carlos G. Spinola, C. Galvez-Fernández and M. Martín-Vázquez

Abstract: Stainless steel manufacturing has experienced a high growth. Nowadays the stainless steel manufacturing is an industry with many applications. Annealing process is an important process in the production of stainless steel coils. The aim of this research is to obtain the classification of defective annealed coils. So a nonlinear mathematical model has been developed for the annealing process. In this research the following techniques have been used: SOM neural networks and classifications methods. For testing, temperature signals were collected along the annealing furnace, also speed signal of the production line were collected. These signals are correlated with each one of the manufactured coils.

Paper Nr: 18
Title:

Prediction of the Behaviours by the Prismatic Beams with Polypropylene Fibers under High Temperature Effects through Artificial Neural Networks

Authors:

Fatih Altun and Tamer Dirikgil

Abstract: In order to improve the mechanical qualities of a concrete, various kinds of fibers are added to the concrete. In the studies, polypropylene (PP) fibers are employed as a fiber type. It has a significant place in the researches that PP fibers not only improve the mechanical qualities of the concrete under normal temperatures, but also prevents the bursting of the concrete with the internal vapour compression under high temperatures. The distributions and locations of the fibers in the concrete and the variables employed for experimental proceedings affect the mechanical results. This makes it difficult to link the obtained results to each other. In order to establish a complicated link, it is inevitable to create a learning mechanism. In this study, multilayered perceptrons (MLP) and radial basis function artificial neural network (RBFNN) models were used and their flexure strengths were sought to be predicted. Both of the neural network models put in a successful performance and enabled the prediction of the experimental results with a satisfying approximation.

Paper Nr: 26
Title:

Influence of Data Orthogonality - On the Accuracy and Stability of Financial Market Predictions

Authors:

A. A. Maknickas and N. Maknickiene

Abstract: Input selection is always important for adapting artificial intelligence systems for forecasting. Recurrent neural networks could predict using the historical data of financial markets but the predictions are very unstable. The goal of our paper is to study the influence of two historical data inputs on accuracy and stability of recurrent neural network forecasting. It is proposed to use orthogonal recurrent neural network inputs for the prediction of financial market exchange rates. Statistical comparison of the predicted results for different degrees of orthogonality of the data inputs shows much tighter distribution of the predicted results, when the more orthogonal input data are used. This proposed data input concept was tested using evolution of recurrent systems with linear Outputs recurrent neural network with historical input data of currency exchange rates.

Paper Nr: 35
Title:

Induction Motor Speed Control using Fuzzy Neural Network Speed Estimation

Authors:

Tien-Chi Chen and Wei-Chung Wang

Abstract: The field-oriented control (FOC) of induction motor has high static and dynamic performance. In order to achieve the speed loop feedback control, precise rotor speed information is important for induction motor control. In the past, encoder was widely used to obtain the speed information of induction motor. However, speed sensor would increase the cost of entire system and reduce the system reliability. In addition, for some special applications such as very high speed motor drives, some difficulties are encountered in mounting these speed sensors. The speed sensorless control would overcome these problems. This paper proposes a fuzzy neural network speed estimation for induction motor speed sensorless control. The speed estimation is based on the deduction of rotor flux and estimated rotor flux, which is calculated by fuzzy neural network. The fuzzy neural network includes a four-layer network. The steepest descent algorithm and back-propagation algorithm are used to adjust the parameters of fuzzy neural network in order to minimize the error between the rotor flux and the estimated rotor flux, which is implied to enable precise estimation of the rotor speed.

Paper Nr: 42
Title:

Component-based Gender Classification based on Hair and Facial Geometry Features

Authors:

Wen-Shiung Chen, Wen-Jui Chang, Lili Hsieh and Zong-Yi Lin

Abstract: In this paper, a component-based gender classification based on hair and facial geometrical features are presented. By way of these preprocessing, hair and facial geometry features can then be extracted automatically from the face images. We compare hair detection methods by examining their color and texture features, and also analyze some geometrical features from references. The best performance of 87.15% in gender classification rate is achieved by combining the most significant hair and geometrical features which is better than some of the literature before.

Paper Nr: 43
Title:

Sign Recognition with HMM/SVM Hybrid for the Visually-handicapped in Subway Stations

Authors:

Dong-jin Lee and Ho-sub Yoon

Abstract: In this paper, we propose a sign classification system to recognize exit number and arrow signs in natural scene images. The purpose of the sign classification system is to provide assistance to a visually-handicapped person in subway stations. For automatically extracting sign candidate regions, we use Adaboost algorithm, however, our detector not only extracts sign regions, but also non-sign (noise) regions in natural scene images. Thus, we suggest a verification technique to discriminate sign regions from non-sign regions. In addition, we suggest a novel feature extraction algorithm cooperated with Hidden Markov Model. To evaluate the system, we tested a total of 20,177 sign candidate regions including the number of 8,414 non-sign regions on the captured images under several real environments in Daejeon in South Korea. We achieved an exit number and arrow sign recognition rate of each 99.5% and 99.8% and a false positive rate (FPR) of 0.3% to discriminate between sign regions and non-sign regions.

Paper Nr: 44
Title:

Deductive Reasoning - Using Artificial Neural Networks to Simulate Preferential Reasoning

Authors:

Marco Ragni and Andreas Klein

Abstract: Composition tables are used in AI for knowledge representation and to compute transitive inferences. Most of these tables are computed by hand, i.e., there is the need to generate them automatically. Furthermore, human preferred solutions and errors in reasoning can be explained as well based on these tables. First, we will report briefly psychological results about the preferences in calculi. Then we show that we can train ANNs on a simple calculus like the point algebra and the trained ANN is able to correctly solve larger calculi such as the Cardinal Direction Calculus. As human prefer specific conclusions, we are able to show that based on the ANN, which is trained on the preferred conclusions of the point algebra alone, is able to reproduce the results on the larger calculi as well. Finally, we show that humans preferred solutions can be adequately described by the networks. A brief discussion of the structure of successful ANNs conclude the paper.

Paper Nr: 45
Title:

A Neural Network Model of Cortical Auditory-visual Interactions - A Neurocomputational Analysis of the Shams-Illusion

Authors:

Cristiano Cuppini, Elisa Magosso and Mauro Ursino

Abstract: The perception of the external world is based on the integration of data from different sensory modalities. Recent theories and experimental findings have suggested that this phenomenon is present since the early low-level cortical areas. The mechanisms underlying these early processes and the organization of the underlying circuitries is still a matter of debate. Here, by using a simple neural network to reproduce and analyse a well-known cross-modal illusion occurring in the visual cortex, we suggest that a fundamental role is played by direct excitatory synapses between visual and auditory regions.

Paper Nr: 46
Title:

Improved Iris Recognition using Parabolic Normalization and Multi-layer Perceptron Neural Network

Authors:

A. Hilal, B. Daya and P. Beauseroy

Abstract: Iris signature is considered as one of the richest, unique, and stable biometrics. This permits to an iris identification system to identify a person even after many years from his first iris signature extraction. In this paper we investigate a new method of iris normalization where iris features are normalized in a parabolic function. Thus iris information close to the pupil is privileged to that close to the sclera. A multilayer perceptron artificial neural network is then used to test the normalization effect and compare it with classical linear normalization method. The study is tested on CASIA V3 database iris images. Accuracy at the equal error rate operating point and receiver operating characteristics curves show better results with the parabolic normalization method and thus propose its use for better iris recognition system performance.

Paper Nr: 68
Title:

Rate of Penetration Prediction and Optimization using Advances in Artificial Neural Networks, a Comparative Study

Authors:

Khoukhi Amar and Alarfaj Ibrahim

Abstract: An important aspect of oil industry is rate of penetration (ROP) prediction. Many studies have been implemented to predict it. Mainly, multiple regression and artificial neural network models were used. In this paper, the objective is to compare the traditional multiple regression with two artificial intelligence techniques; extreme learning machines (ELM) and radial basis function networks (RBF). ELM and RBF are artificial neural network (ANNs) techniques. ANNs are cellular systems which can acquire, store, and utilize experiential knowledge. The techniques are implemented using MATLAB function codes. For ELM, the activation functions, number of hidden neurons, and number of data points in the training data set are varied to find the best combination. Different input parameters of ELM give different results. The comparison is made based on field data with no correction, then with weight on bit (WOB) correction, and finally with interpolated WOB and rotary speed (RPM) correction. Seven input parameters are used for ROP prediction. These are depth, bit weight, rotary speed, tooth wear, Reynolds number function, ECD and pore gradient. The techniques are compared in terms of training time and accuracy, and testing time and accuracy. Simulation experiments show that ELM gave the best results in terms of accuracy and processing time.

Paper Nr: 70
Title:

ARTMAP with Modified Internal Category Geometry to Reduce the Category Proliferation

Authors:

Robinson Alves, Carlos Padilha, Jorge Melo and Adrião Dória Neto

Abstract: In this paper the concept of polytopes was explored in order to change the internal category geometry of the ARTMAP network. It was possible to achieve reduction of recruited categories and improving the quality of generalization of the neural network. The use of polytopes in the network achieves excellent results for both the generalization and the number of recruited categories. These significant improvements occurred in relation to other networks from ART family and also for other types of classifiers.