NCTA 2014 Abstracts


Full Papers
Paper Nr: 1
Title:

A Walsh Analysis of Multilayer Perceptron Function

Authors:

Kevin Swingler

Abstract: The multilayer perceptron (MLP) is a widely used neural network architecture, but it suffers from the fact that its knowledge representation is not readily interpreted. Hidden neurons take the role of feature detectors, but the popular learning algorithms (back propagation of error, for example) coupled with random starting weights mean that the function implemented by a trained MLP can be difficult to analyse. This paper proposes a method for understanding the structure of the function learned by MLPs that model functions of the class f : f􀀀1;1gn ! Rm. The approach characterises a given MLP using Walsh functions, which make the interactions among subsets of variables explicit. Demonstrations of this analysis used to monitor complexity during learning, understand function structure and measure the generalisation ability of trained networks are presented.

Paper Nr: 2
Title:

A Cortico-Collicular Model for Multisensory Integration

Authors:

Federico Giovannini and Elisa Magosso

Abstract: Remarkable visual-auditory cross-modal phenomena occur at perceptual level: a visual stimulus enhances or biases auditory localization in case of spatially coincident or spatially disparate stimuli. Hemianopic patients (with one blind hemifield resulting from damage to primary visual cortex) retain visual enhancement but not visual bias of auditory localization in the blind hemifield. Here, we propose a neural network model to investigate which cortical and subcortical regions may be involved in these phenomena in intact and damaged conditions. The model includes an auditory cortical area, the primary and extrastriate visual cortices and the Superior Colliculus (a subcortical structure). Model simulations suggest that: i) Visual enhancement of auditory localization engages two circuits (one involving the primary visual cortex and one involving the Superior Colliculus) that act in a redundant manner. In absence of primary visual cortex (hemianopia), enhancement still occurs thanks to the Superior Colliculus strongly activated by the spatially coincident stimuli. ii) Visual bias of auditory localization is due to an additive contribution of the two circuits. In hemianopia, the effect disappears as the Superior Colliculus is not sufficiently activated by the spatially disparate stimuli. The model helps interpreting perceptual visual-auditory phenomena and their retention or absence in brain damage conditions.

Paper Nr: 5
Title:

Artificial Neural Network Models of Intersegmental Reflexes

Authors:

Alicia Costalago Meruelo, David M. Simpson, S. Veres and Philip L. Newland

Abstract: In many animals intersegmental reflexes are important for postural control and movement making them ideal candidates for the bio-inspired design of medical treatment for neuromuscular injuries in cases such as drop foot and possibly in robot design. In this paper we study an intersegmental reflex of the foot (tarsus) of the locust hind leg, which is a reflex that raises the tarsus when the tibia is flexed and depresses it when the tibia is extended. A novel method is described to quantify the intersegmental responses in which an Artificial Neural Network, the Time Delay Neural Network, is applied. The architecture of the network is optimised through a metaheuristic algorithm to produce accurate predictions with short computational time and complexity and high generalisation to different individual responses. The results show that ANNs provide accurate predictions when trained with an average reflex response to Gaussian White Noise stimulation compared to autoregressive models. Furthermore, the network model can calculate the individual responses from each of the animals and responses to another input such as a sinusoid. A detailed understanding of such a reflex response could be included in the design of orthoses or functional electrical stimulation treatments to improve walking in patients with neuromuscular disorders.

Paper Nr: 8
Title:

Are Non-Standard Neural Behaviors Computationally Relevant?

Authors:

Stylianos Kampakis

Abstract: An idea that has recently appeared in the neural network community is that networks with heterogeneous neurons and non-standard neural behaviors can provide computational advantages. A theoretical investigation of this idea was given by Kampakis (2013) for spiking neurons. In artificial neural networks this idea has been recently researched through Neural Diversity Machines (Maul, 2013). However, this idea has not been tested experimentally for spiking neural networks. This paper provides a first experimental investigation of whether neurons with non-standard behaviors can provide computational advantages. This is done by using a spiking neural network with a biologically realistic neuron model that is tested on a supervised learning task. In the first experiment the network is optimized for the supervised learning task by adjusting the parameters of the neurons in order to adapt the neural behaviors. In the second experiment, the parameter optimization is used in order to improve the network’s performance after the weights have been trained. The results confirm that neurons with non-standard behaviors can provide computational advantages for a network. Further implications of this study and suggestions for future research are discussed.

Paper Nr: 12
Title:

A Novel Particle Swarm Optimization Algorithm Based on Iterative Chaotic Map with Infinite Collapses for Global Optimization

Authors:

He Yuyao and Song Cheng

Abstract: As a novel optimization technique, particle swarm optimization (PSO) has gained much attention during the past decades. To enhance the performance of PSO, a novel chaotic PSO algorithm based on iterative chaotic map with Infinite collapses for global optimization is proposed. Firstly we propose a novel iterative chaotic map with Infinite collapses model (called ICMIC), which shows a good chaotic performance. Secondly, by incorporating ICMIC perturbation into the velocity updating equation of the standard PSO, a hybrid particle swarm optimization (called ICMICCPSO) is presented, which has a chaotic dynamics first and then a steepest descent dynamics. That is, the chaotic search is firstly implemented for every particle, and then, with the decaying of chaotic perturbation, the chaotic dynamics and searching space are undergoing condensation over generations for finding an area that may contain the global minima. Finally, the optimization process without ICMIC perturbations will be dominated by the stochastic steepest gradient descent property for finding global minimum point. The proposed ICMICPSO method is tested on several widely used benchmark multimodal continuous functions. Numerical results and comparisons with other Chaotic PSO variants demonstrate that the proposed ICMICPSO has good searching efficiency and can effectively balance the exploration and exploitation abilities.

Paper Nr: 20
Title:

Exploiting Local Class Information in Extreme Learning Machine

Authors:

Alexandros Iosifidis, Anastasios Tefas and Ioannis Pitas

Abstract: In this paper we propose an algorithm for Single-hidden Layer Feedforward Neural networks training. Based on the observation that the learning process of such networks can be considered to be a non-linear mapping of the training data to a high-dimensional feature space, followed by a data projection process to a lowdimensional space where classification is performed by a linear classifier, we extend the Extreme Learning Machine (ELM) algorithm in order to exploit the local class information in its optimization process. The proposed Local Class Variance Extreme Learning Machine classifier is evaluated in facial image classification problems, where we compare its performance with that of other ELM-based classifiers. Experimental results show that the incorporation of local class information in the ELMoptimization process enhances classification performance.

Paper Nr: 26
Title:

Volatile Organic Compound Detection with FET Sensors and Neural Network Data Processing as a Preliminary Step to Early Lung Cancer Diagnosis

Authors:

John C. Cancilla, Bin Wang, Pablo Diaz-Rodriguez, Gemma Matute, Hossam Haick and Jose S. Torrecilla

Abstract: Cancer is currently one of deadliest and most feared diseases in the developed world, and, particularly, lung cancer (LC) is one of the most common types and has one of the highest death/incidence ratios. An early diagnosis for LC is probably the most accessible possibility to try and save patients and lower this ratio. Recently, research concerning LC-related breath biomarkers has provided optimistic results and has become a real option to try and obtain a fast, reliable, and early LC diagnosis. In this paper, a combination of field-effect transistor (FET) sensors and artificial neural networks (ANNs) has been employed to classify and estimate the partial pressures of a series of polar and nonpolar volatile organic compounds (VOCs) present in prepared gaseous mixtures. The objective of these preliminary tests is to give an idea of how well this technology can be used to analyze artificial or real breath samples by quantifying the LC-related VOCs or biomarkers. The results of this step are very promising and indicate that this methodology deserves further research using more complex samples to find the existing limitations of the FET-ANN combination.

Paper Nr: 37
Title:

Pattern Recognition by Probabilistic Neural Networks - Mixtures of Product Components versus Mixtures of Dependence Trees

Authors:

Jiri Grim and Pavel Pudil

Abstract: We compare two probabilistic approaches to neural networks - the first one based on the mixtures of product components and the second one using the mixtures of dependence-tree distributions. The product mixture models can be efficiently estimated from data by means of EM algorithm and have some practically important properties. However, in some cases the simplicity of product components could appear too restrictive and a natural idea is to use a more complex mixture of dependence-tree distributions. By considering the concept of dependence tree we can explicitly describe the statistical relationships between pairs of variables at the level of individual components and therefore the approximation power of the resulting mixture may essentially increase. Nonetheless, in application to classification of numerals we have found that both models perform comparably and the contribution of the dependence-tree structures decreases in the course of EM iterations. Thus the optimal estimate of the dependence-tree mixture tends to converge to a simple product mixture model. Regardless of computational aspects, the dependence-tree mixtures could help to clarify the role of dendritic branching in the highly selective excitability of neurons.

Paper Nr: 42
Title:

Derivative Free Training of Recurrent Neural Networks - A Comparison of Algorithms and Architectures

Authors:

Branimir Todorović, Miomir Stanković and Claudio Moraga

Abstract: The problem of recurrent neural network training is considered here as an approximate joint Bayesian estimation of the neuron outputs and unknown synaptic weights. We have implemented recursive estimators using nonlinear derivative free approximation of neural network dynamics. The computational efficiency and performances of proposed algorithms as training algorithms for different recurrent neural network architectures are compared on the problem of long term, chaotic time series prediction.

Paper Nr: 44
Title:

Neuron Dynamics of Two-compartment Traub Model for Hardware-based Emulation

Authors:

Juan Carlos Moctezuma, Jose Luis Nunez-Yanez and Joseph P. McGeehan

Abstract: The two-compartment Pinsky and Rinzel version of the Traub model offers a suitable solution for hardware-based emulation, since it has a good trade-off between biophysical accuracy and computational resources. Many applications based on conductance-based models require a proper characterization of the neuron behaviour in terms of its parameters, such as tuning firing parameters, changing parameters during learning processes, replication and analysis of neuron recordings, etc. This work presents a study of the dynamics of such model especially suitable for hardware-based development. The morphology of the neuron is taken into account while the analysis focuses primarily on the relation between the firing/bursting properties and the relevant parameters of the model, such as current applied and morphology of the cell. Two different applied currents were considered: short duration and long steady. Seven different types of burst patterns were detected and analysed. The transformation process of the membrane voltage when a long steady current varies was classified into five stages. Finally, examples of neuron recording replication using the present methodology are developed.

Short Papers
Paper Nr: 6
Title:

Using a Hopfield Iterative Neural Network to Explain Diffusion in the Brain’s Extracellular Space Structure

Authors:

Abir Alharbi

Abstract: Many therapies for drug delivery to the brain are based on diffusion, and diffusion in this extracellular space is based on micro-techniques that can be modelled with classical differential equations such as the point source diffusion equation. In this paper an energy function is constructed using a finite-difference approximation to the governing diffusion equation and then minimized by a Hopfield neural network. The synergy of Hopfield neural networks with finite difference approximation is promising. The neural network approach is capable of giving insight to the complex brain activity better than any other classical numerical method and the parallelism nature of the Hopfield neural networks approach is easier to implement on fast parallel computers and this will make them faster than the traditional methods for modelling this complex problem. Moreover, the effect of the involved parameters on the diffusion distribution and drug delivery in the ECS is investigated.

Paper Nr: 7
Title:

Surgical Skill Evaluation by Means of a Sensory Glove and a Neural Network

Authors:

Giovanni Costantini, Giovanni Saggio, Laura Sbernini, Nicola Di Lorenzo, Franco Di Paolo and Daniele Casali

Abstract: In this work we used the HiTEg data glove to measure the skill of a physician or physician student in the execution of a typical surgical task: the suture. The aim of this project is to develop a system that, analyzing the movements of the hand, could tell if they are correct. To collect a set of measurements, we asked 18 subjects to performing the same task wearing the sensory glove. Nine subjects were skilled surgeons and nine subjects were non-surgeons, every subject performed ten repetitions of the same task, for two sessions, yielding to a dataset of 36 instances. Acquired data has been processed and classified with a neural network. A feature selection has been done considering only the features that have less variance among the expert subjects. The cross-validation of the classifier shows an error of 5.6%.

Paper Nr: 11
Title:

A Neural Model of Moral Decisions

Authors:

Alessio Plebe

Abstract: In this paper a neural model of moral decisions is proposed. It is based on the fact, supported by neuroimaging studies as well as theoretical analysis, that moral behavior is supported by brain circuits engaged more generally in emotional responses and in decision making. The model has two components, the first is composed by artificial counterpart of the orbitofrontal cortex, connected with sensorial cortical sheets and with the ventral striatum, the second by the ventromedial prefrontal cortex, that evaluate representations of values from the orbitofrontal cortex, comparing with negative values, encoded in the amygdala. The model is embedded in a simple environmental context, in which it learns that certain actions, although potentially rewarding, are morally forbidden.

Paper Nr: 19
Title:

Brain Modeling with Brytes - Making Big Brains from a Lot of Little Brains

Authors:

David Zipser

Abstract: Brytes are small brains used as subunits to model the cognitive processes of larger, smarter brains. A previously developed model of scratching behaviour that uses brytes to generate the coordinated movements of two arms, one with the itch site the other with the scratching hand is described. Then new strategies are described for using large sets of brytes with virtual locations all over the body to make decisions about whether scratching is safe in the current context and, if so, which appendage to use. Finally, the biological plausibility of brytes is examined in the contest of brain evolution and brain functional architecture.

Paper Nr: 23
Title:

Automated Segmentation of Folk Songs Using Artificial Neural Networks

Authors:

Andreas Neocleous, Nicolai Petkov and Christos N. Schizas

Abstract: Two different systems are introduced, that perform automated audio annotation and segmentation of Cypriot folk songs into meaningful musical information. The first system consists of three artificial neural networks (ANNs) using timbre low-level features. The output of the three networks is classifying an unknown song as “monophonic” or “polyphonic”. The second system employs one ANN using the same feature set. This system takes as input a polyphonic song and it identifies the boundaries of the instrumental and vocal parts. For the classification of the “monophonic – polyphonic”, a precision of 0.88 and a recall of 0.78 has been achieved. For the classification of the “vocal – instrumental” a precision of 0.85 and recall of 0.83 has been achieved. From the obtained results we concluded that the timbre low-level features were able to capture the characteristics of the audio signals. Also, that the specific ANN structures were suitable for the specific classification problem and outperformed classical statistical methods.

Paper Nr: 27
Title:

Modelling and Analysis of Retinal Ganglion Cells Through System Identification

Authors:

Dermot Kerr, Martin McGinnity and Sonya Coleman

Abstract: Modelling biological systems is difficult due to insufficient knowledge about the internal components and organisation, and the complexity of the interactions within the system. At cellular level existing computational models of visual neurons can be derived by quantitatively fitting particular sets of physiological data using an input-output analysis where a known input is given to the system and its output is recorded. These models need to capture the full spatio-temporal description of neuron behaviour under natural viewing conditions. At a computational level we aspire to take advantage of state-of-the-art techniques to accurately model non-standard types of retinal ganglion cells. Using system identification techniques to express the biological input-output coupling mathematically, and computational modelling techniques to model highly complex neuronal structures, we will "identify" ganglion cell behaviour with visual scenes, and represent the mapping between perception and response automatically.

Paper Nr: 28
Title:

STDP Learning Under Variable Noise Levels

Authors:

Dalius Krunglevicius

Abstract: Spike-timing-dependent plasticity (STDP) is a set of Hebbian learning rules which are firmly based on biological evidence. It has been demonstrated that one of the STDP learning rules is suited for learning spatiotemporal patterns in a very noisy environment. Parameters of the neuron are only optimal, however, for a certain range of quantity of injected noise. This means the level of noise must be known beforehand so that the parameters can be set accordingly. That could be a real problem when noise levels vary over time. We found that the model of a leaky-integrate-and-fire inhibitory neuron with an inverted STDP learning rule is capable of adjusting its response rate to a particular level of noise. In this paper we suggest a method that uses an inverted SDTP learning rule to modulate spiking rate of the trained neuron. This method is adaptive to noise levels; subsequently spiking neuron can be trained to learn the same spatiotemporal pattern with a wide range of background noise injected during the learning process.

Paper Nr: 32
Title:

Singularity Stairs Following with Limited Numbers of Hidden Units

Authors:

Seiya Satoh and Ryohei Nakano

Abstract: In a search space of a multilayer perceptron having J hidden units, MLP(J), there exist flat areas called singular regions that cause serious stagnation of learning. Recently a method called SSF1.3 utilizing singular regions has been proposed to systematically and stably find excellent solutions. SSF1.3 starts search from a search space of MLP(1), increasing J one by one. This paper proposes SSF2 that performs MLP search by utilizing singular regions with J changed bidirectionally within a certain range. The proposed method was evaluated using artificial and real data sets.

Paper Nr: 34
Title:

Graph-based Kernel Representation of Videos for Traditional Dance Recognition

Authors:

Christina Chrysouli, Vasileios Gavriilidis and Anastasios Tefas

Abstract: In this paper, we propose a novel graph-based kernel method in order to construct histograms for a bag of words approach, by using similarity measures, applied in activity recognition problems. Bag of words is the most popular framework for performing classification on video data. This framework, however, is an orderless collection of features. We propose a better way to encode action in videos, via altering the histograms. The creation of such histograms is performed based on kernel methods, inspired from graph theory, computed with no great additional computational cost. Moreover, when using the proposed algorithm to construct the histograms, a richer representation of videos is attained. Experiments on folk dances recognition were conducted based on our proposed method, by comparing histograms extracted from a typical bag-of-words framework against histograms of the proposed method, which provided promising results on this challenging task.

Paper Nr: 36
Title:

Stability Evaluation of Combined Neural Networks

Authors:

Ibtissem Ben Othman and Faouzi Ghorbel

Abstract: In the industrial field, the artificial neural network classifiers are currently used and they are generally integrated of technologic systems which need efficient classifier. However, the lack of control over its mathematical formulation explains the instability of its classification results. In order to improve the prediction accuracy, most of researchers refer to the classifiers combination approach. This paper tries to illustrate the capability of an example of combined neural networks to improve the stability criterion of the single neural classifier. The stability comparison is performed by the error rate probability densities function estimated by a new variant of the kernel-diffeomorphism semi-bounded Plug-in algorithm.

Paper Nr: 38
Title:

A Walk in the Statistical Mechanical Formulation of Neural Networks - Alternative Routes to Hebb Prescription

Authors:

Elena Agliari, Adriano Barra, Andrea Galluzzi, Daniele Tantari and Flavia Tavani

Abstract: Neural networks are nowadays both powerful operational tools (e.g., for pattern recognition, data mining, error correction codes) and complex theoretical models on the focus of scientific investigation. As for the research branch, neural networks are handled and studied by psychologists, neurobiologists, engineers, mathematicians and theoretical physicists. In particular, in theoretical physics, the key instrument for the quantitative analysis of neural networks is statistical mechanics. From this perspective, here, we review attractor networks: starting from ferromagnets and spin-glass models, we discuss the underlying philosophy and we recover the strand paved by Hopfield, Amit-Gutfreund-Sompolinky. As a sideline, in this walk we derive an alternative (with respect to the original Hebb proposal) way to recover the Hebbian paradigm, stemming from mixing ferromagnets with spin-glasses. Further, as these notes are thought of for an Engineering audience, we highlight also the mappings between ferromagnets and operational amplifiers, hoping that such a bridge plays as a concrete prescription to capture the beauty of robotics from the statistical mechanical perspective.

Paper Nr: 39
Title:

Learning Kernel Label Decompositions for Ordinal Classification Problems

Authors:

M. Pérez-Ortiz, P. A. Gutiérrez and C. Hervás-Martínez

Abstract: This paper deals with the idea of decomposing ordinal multiclass classification problems when working with kernel methods. The kernel parameters are optimised for each classification subtask in order to better adjust the kernel to the data. More flexible multi-scale Gaussian kernels are considered to increase the goodness of fit of the kernel matrices. Instead of learning independent models for all the subtasks, the optimum convex combination of the kernel matrices is then obtained, leading to a single model able to better discriminate the classes in the feature space. The results of the proposed algorithm shows promising potential for the acquisition of better suited kernels.

Paper Nr: 41
Title:

The Ideomotor Principle Simulated - An Artificial Neural Network Model for Intentional Movement and Motor Learning

Authors:

Neri Accornero and Marco Capozza

Abstract: Although the ideomotor principle (IMP), the notion positing that the nervous system initiates voluntary actions by anticipating their sensory effects, has long been around it still struggles to gain widespread acknowledgement. Supporting this theory, we present an artificial neural network model driving a simulated arm, designed as simply as possible to focus on the essential IMP features, that demonstrates by simulation how the IMP could work in biological intentional movement and motor learning. The simulation model shows that IMP motor learning is fast and effective and shares features with human motor learning. An IMP extension offers new insights into the so-called mirror neuron and canonical neuron systems.

Paper Nr: 52
Title:

Combining Empirical Mode Decomposition with Neural Networks for the Prediction of Exchange Rates

Authors:

J. Mouton and A. J. Hoffman

Abstract: This paper proposes a neural network based model applied to empirical mode decomposition (EMD) filtered data for multi-step-ahead prediction of exchange rates. EMD is used to decompose the returns of exchange rates into intrinsic mode functions (IMFs) which are partially recomposed to produce a low-pass filtered time series. This series is used to train a neural network for multi-step-ahead prediction. Out-of-sample tests on EUR/USD and USD/JPY rates show superior performance compared to random walk and neural network models that do not employ EMD filtering. The novel approach of using EMD as a filtering technique in combination with neural networks consistently delivers higher returns on investment and demonstrates its utility in multi-step-ahead prediction.

Paper Nr: 54
Title:

Document Image Classification Via AdaBoost and ECOC Strategies Based on SVM Learners

Authors:

Mehmet Ahat, Cagdas Ulas and Onur Agin

Abstract: In this paper, we describe easily extractable features and an approach for document image retrieval and classification at spatial level. The approach is based on the content of the image and utilizing visual similarity, it provides high speed classification of noisy text document images without optical character recognition (OCR). Our method involves a bag-of-visual words (BoVW) model on the designed descriptors and a Random- Window (RW) technique to capture the structural relationships of the spatial layout. Using the features based on these information, we analyze different multiclass classification methods as well as ensemble classifiers method with Support Vector Machine (SVM) as a base learner. The results demonstrate that the proposed method for obtaining structural relations is competitive for noisy document image categorization.

Paper Nr: 57
Title:

Empirical Models as a Basis for Synthesis of Large Spiking Neural Networks with Pre-Specified Properties

Authors:

Mikhail Kiselev

Abstract: Analysis of behaviour of large neuronal ensembles using mean-field equations and similar approaches was an important instrument in theory of spiking neural networks during almost all its history. However, it often implies dealing with complex systems of integro-differential equations which are very hard not only for obtaining explicit analytical solution but also for simpler tasks like stability analysis. Building empirical models on the basis of experimental data gathered in process of simulation of small size networks is considered in the paper as a practical alternative to these traditional methods. A methodology for creation and verification of such models using decision trees, multiple adaptive regression splines and other data mining algorithms is discussed. This idea is illustrated by the two examples – prediction of probability of avalanche-like excitation growth in the network and analysis of conditions necessary for development of strong firing frequency oscillations.

Paper Nr: 58
Title:

Neural Control using EEG as a BCI Technique for Low Cost Prosthetic Arms

Authors:

Shyam Diwakar, Sandeep Bodda, Chaitanya Nutakki, Asha Vijayan, Krishnashree Achuthan and Bipin Nair

Abstract: There have been significant advancements in brain computer interface (BCI) techniques using EEG-like methods. EEG can serve as non-invasive BMI technique, to control devices like wheelchairs, cursors and robotic arm. In this paper, we discuss the use of EEG recordings to control low-cost robotic arms by extracting motor task patterns and indicate where such control algorithms may show promise towards the humanitarian challenge. Studies have shown robotic arm movement solutions using kinematics and machine learning methods. With iterative processes for trajectory making, EEG signals have been known to be used to control robotic arms. The paper also showcases a case-study developed towards this challenge in order to test such algorithmic approaches. Non-traditional approaches using neuro-inspired processing techniques without implicit kinematics have also shown potential applications. Use of EEG to resolve temporal information may, indeed, help understand movement coordination in robotic arm.

Paper Nr: 61
Title:

Combining Different Computational Techniques in the Development of Financial Prediction Models

Authors:

A. J. Hoffman

Abstract: The prediction of financial time series to enable improved portfolio management is a complex topic that has been widely researched. Modelling challenges include the high level of noise present in the signals, the need to accurately model extreme rather than average behaviour, the inherent non-linearity of relationships between explanatory and predicted variables and the need to predict the future behaviour of a large number of independent investment instruments that must be considered for inclusion into a well-diversified portfolio. This paper demonstrates that linear time series prediction does not offer the ability to develop reliable prediction models, due to the inherently non-linear nature of the relationship between explanatory and predicted variables. It is shown that the results of histogram based sorting techniques can be used to guide the selection of suitable variables to be included in the development of a neural network model. We find that multivariate neural network models can outperform the best models using only a single explanatory variable. We furthermore demonstrate that the stochastic nature of the signals can be addressed by training common models for a number of similar instruments which forces the neural network to model the underlying relationships rather than the noise in the signals.

Paper Nr: 62
Title:

A Computational Model for Simulation of Moral Behavior

Authors:

Fernanda M. Eliott and Carlos H. C. Ribeiro

Abstract: The extension of our integration to technologies brings about the possibility of inserting moral prototypes into artificial agents, no matter if they are going to interact with other artificial agents or biological creatures. We describe here MultiA, a computational model for simulating moral behavior derived from changes over a biologically inspired architecture. MultiA uses reinforcement learning techniques and is intended to produce selective cooperative behavior as a consequence of a biologically plausible model of morality inspired from a perusal of empathy. MultiA has its sensorial information translated into emotions and homeostatic variable values, which feed cognitive and learning systems. The moral behavior is expected to emerge from the artificial social emotion of sympathy and its associated feeling of empathy, based on an ability to internally emulate other agents internal states.

Paper Nr: 65
Title:

Intrinsic Fault Tolerance of Hopfield Artificial Neural Network Model for Task Scheduling Technique in SoC

Authors:

Rajhans Singh and Daniel Chillet

Abstract: Due to the technology evolution, one of the main problems for future System-on-Chips (SoC) concerns the difficulties to produce circuits without defaults. While designers propose new structures able to correct the faults occurring during computation, this article addresses the control part of SoCs, and focuses on task scheduling for processors embedded in SoC. Indeed, to ensure the execution of application in presence of faults on such systems, operating system services will need to be fault tolerant. This is the case for the task scheduling service, which is an optimization problem that can be solved by Artificial Neural Network. In this context, this paper explores the intrinsic fault tolerance capability of Hopfield Artificial Neural Network (HANN). Our work shows that even if some neurons are in fault, a HANN can provide valid solutions for task scheduling problem. We define the intrinsic limit of fault tolerance capability of Hopfield model and illustrate the impact of fault on the network convergence.

Paper Nr: 69
Title:

On Error Probability of Search in High-Dimensional Binary Space with Scalar Neural Network Tree

Authors:

Vladimir Kryzhanovsky, Magomed Malsagov, Juan Antonio Clares Tomas and Irina Zhelavskaya

Abstract: The paper investigates SNN-tree algorithm that extends the binary search tree algorithm so that it can deal with distorted input vectors. Unlike the SNN-tree algorithm, popular methods (LSH, k-d tree, BBF-tree, spill-tree) stop working as the dimensionality of the space grows (N > 1000). The proposed algorithm works much faster than exhaustive search (26 times faster at N=10000). However, some errors may occur during the search. In this paper we managed to obtain an estimate of the upper bound on the error probability for SNN-tree algorithm. In case when the dimensionality of input vectors is N≥500 bits, the probability of error is so small (P<10-15) that it can be neglected according to this estimate and experimental results. In fact, we can consider the proposed SNN-tree algorithm to be exact for high dimensionality (N ≥ 500).

Paper Nr: 73
Title:

Extraction of Dynamics-correlated Factors from Image Features in Pushing Motion of a Mobile Robot

Authors:

Takahiro Inaba and Yuichi Kobayashi

Abstract: It is important for autonomous robots to improve capability of extracting information that is relevant to their motions. This paper presents an extraction and estimation of factors that affect behavior of object from image features in object pushing manipulation by a two-wheeled robot. Motions of image features (SIFT keypoints) are approximated with variance. By detecting correlation between the variance and positions of the keypoints, the robot can detect keypoints whose positions affect behaviour of some keypoints. Position information of the keypoints is expected to be useful for the robot to decide its pushing motion. The proposed scheme was verified in experiment with a camera-mounted mobile robot which has no pre-defined knowledge about its environment.

Paper Nr: 75
Title:

Planning of Pushing Manipulation by a Mobile Robot Considering Cost of Recovery Motion

Authors:

Takahiro Saito, Yuichi Kobayashi and Tatsuya Naruse

Abstract: This paper presents a planning method of pushing manipulation by a mobile robot. It is sometimes very useful if the robot can take recovery action, namely, re-approaching and re-pushing, when it turns out to be ineffective to keep current pushing motion. The proposed planning framework is based on the idea of mode switching, where three modes; approaching, pushing and re-pushing, are considered. The pushing motion is first built with dynamic programming, which provides value function of the state. Based on the value, planning of re-approaching to the object and re-pushing is conducted using a value iteration algorithm extended to state space with uncertainty. The proposed planning framework was evaluated in simulation, and it was shown that it provides more effective behaviour of the robot by recovery motion at an appropriate timing.

Paper Nr: 76
Title:

Oscillatory Model of Neuromorphic Processors by Embedding Orthogonal Filters

Authors:

Wieslaw Citko and Wieslaw Sienko

Abstract: The purpose of this article is to present a model of the computational intelligence system based on a network of coupled phase oscillators. The structure of such a model consists of a net of phase-locked loops (PLL) and orthogonal filters based on a Hamiltonian neural network embedded in this net.

Paper Nr: 79
Title:

Study of an EEG based Brain Machine Interface System for Controlling a Robotic Arm

Authors:

Yicong Gong, Carly Gross, David Fan, Ahmed Nasrallah, Nathaniel Maas, Kelly Cashion and Vijayan K. Asari

Abstract: We present a methodology to explore the capabilities of an existing interface for controlling a robotic arm with information extracted from brainwaves. Brainwaves are collected through the use of an Emotiv EPOC headset. The headset utilizes electroencephalography (EEG) technology to collect active brain signals. We employ the Emotiv software suites to classify the thoughts of a subject representing specific actions. The system then sends an appropriate signal to a robotic interface to control the robotic arm. We identified several actions for mapping, implemented these chosen actions, and evaluated the system’s performance. We also present the limitations of the proposed system and provide groundwork for future research.

Paper Nr: 80
Title:

A Vision Architecture

Authors:

Christoph von der Malsburg

Abstract: We are offering a particular interpretation (well within the range of experimentally and theoretically accepted notions) of neural connectivity and dynamics and discuss it as the data-and-process architecture of the visual system. In this interpretation the permanent connectivity of cortex is an overlay of well-structured networks, “nets”, which are formed on the slow time-scale of learning by self-interaction of the network under the influence of sensory input, and which are selectively activated on the fast perceptual time-scale. Nets serve as an explicit, hierarchical representation of visual structure in the various sub-modalities, as constraint networks favouring mutually consistent sets of latent variables and as projection mappings to deal with invariance.

Posters
Paper Nr: 17
Title:

Intelligent Recognition of Ancient Persian Cuneiform Characters

Authors:

Fahimeh Mostofi and Adnan Khashman

Abstract: This paper presents an intelligent character recognition system based on utilising a back propagation neural network model. The characters in question are unique and rare to be addressed in such applications. These are the ancient Persian Cuneiform alphanumerical characters. The recognition system comprises firstly, image processing phase where clear and noisy or degraded images of the ancient script are prepared for processing by the neural model in the second phase. The importance of such application lies in its potential to make translating ancient scripts and language easier, faster and more efficient. Experimental results indicate that our proposed method can be further applied successfully to other ancient languages and may be utilised in museums and similar environments.

Paper Nr: 21
Title:

Model Identification for Photovoltaic Panels Using Neural Networks

Authors:

Antonino Laudani, Gabriele Maria Lozito, Martina Radicioni, Francesco Riganti Fulginei and Alessandro Salvini

Abstract: The present work documents the study on the usage of Neural Networks to compute the parameters used in solar panel modelling. The approach followed starts from a dataset obtained by a process of model identification via numerical solution of nonlinear equations. After a preliminary analysis pointing out the intrinsic difficulty in the classic identification of the parameters via NN, by taking advantage of closed form relations, a hybrid neural system, composed by neural network based identifiers and explicit equations, was implemented. The generalization capabilities of the neural identifier were investigated, showing the effectiveness of this approach.

Paper Nr: 22
Title:

Sensor Reduction on EMG-based Hand Gesture Classification

Authors:

Giovanni Costantini, Gianni Saggio, Lucia Quitadamo, Daniele Casali, Alberto Leggieri and Emanuele Gruppioni

Abstract: This work concerns a system based on EMG sensors, signal conditioning circuitry, classification algorithm based on Artificial Neural Network, and virtual avatar representation, useful to identify hand movements within a set of five. This is to potentially make any trans-radial upper-limb amputee able to drive a virtual or real limb prosthetic hand. When using six EMG sensors, the system is able to recognize with an accuracy of 88.8% the gestures performed by a subject, and replicated by an avatar. Here we focused on differences resulting with the adoption of a different number of sensors and therefore, by means of a very simple heuristic method, we compared different subsets of features, excluding the less significant sensors. We found optimal subsets of one, two, three, four and five sensors, demonstrating a decrease of the performance of only 0.8% when using five sensors, while with three sensors the accuracy can be as high as 81.7%.

Paper Nr: 25
Title:

Automatic Segmentation of MR Brain Tumor Images using Support Vector Machine in Combination with Graph Cut

Authors:

Elisabetta Binaghi, Massimo Omodei, Valentina Pedoia, Sergio Balbi, Desiree Lattanzi and Emanuele Monti

Abstract: This work focuses the attention on the automatic segmentation of meningioma from multispectral brain Magnetic Resonance imagery. The Authors address the segmentation task by proposing a fully automatic method hierarchically structured in two phases. The preliminary unsupervised phase is based on Graph Cut framework. In the second phase, preliminary segmentation results are refined using a supervised classification based on Support Vector Machine. The overall segmentation procedure is conceived fully automatic and tailored to non-volumetric data characterized by poor inter-slice spacing, in an attempt to facilitate the insertion in clinical practice. The results obtained in this preliminary study are encouraging and prove that the segmentation benefits from the allied use of Graph Cut and Support Vector Machine frameworks.

Paper Nr: 31
Title:

Prediction Model Adaptation Thanks to Control Chart Monitoring - Application to Pollutants Prediction

Authors:

Philippe Thomas, William Derigent and Marie-Christine Suhner

Abstract: Indoor air quality is a major determinant of personal exposure to pollutants in today’s world since people spend much of their time in numerous different indoor environments. The Anaximen company develops a smart and connected object named Alima, which can measure every minute several physical parameters: temperature, humidity, concentrations of COV, CO2, formaldehyde and particulate matter (pm). Beyond the measurement aspect, Alima presents some data analysis feature named ‘predictive analytics’, whose primary aim is to predict the evolution of indoor pollutants in time. In this article, the neural network (NN) model,embedded in this object and designed for pollutant prediction, is presented. In addition with this NN model, this article also details an approach where batch learning is performed periodically when a too important drift between the model and the system is detected. This approach is based on control charts.

Paper Nr: 33
Title:

Incorporating Privileged Information to Improve Manifold Ordinal Regression

Authors:

M. Pérez-Ortiz, P. A. Gutiérrez and C. Hervás-Martínez

Abstract: Manifold learning covers those learning algorithms where high-dimensional data is assumed to lie on a lowdimensional manifold (usually nonlinear). Specific classification algorithms are able to preserve this manifold structure. On the other hand, ordinal regression covers those learning problems where the objective is to classify patterns into labels from a set of ordered categories. There have been very few works combining both ordinal regression and manifold learning. Additionally, privileged information refers to some special features which are available during classifier training, but not in the test phase. This paper contributes a new algorithm for combining ordinal regression and manifold learning, based on the idea of constructing a neighbourhood graph and obtaining the shortest path between all pairs of patterns. Moreover, we propose to exploit privileged information during graph construction, in order to obtain a better representation of the underlying manifold. The approach is tested with one synthetic experiment and 5 real ordinal datasets, showing a promising potential.

Paper Nr: 43
Title:

Q-Routing in Cognitive Packet Network Routing Protocol for MANETs

Authors:

Amal Alharbi, Abdullah Al-Dhalaan and Miznah Al-Rodhaan

Abstract: Mobile Ad hoc Networks (MANET) are self-organized networks which are characterized by dynamic topologies in time and space. This creates an instable environment, where classical routing approaches cannot achieve high performance. Thus, adaptive routing is necessary to handle the random changing network topology. This research uses Reinforcement Learning approach with Q-Routing to introduce our MANET routing algorithm: Stability-Aware Cognitive Packet Network (CPN). This new algorithm extends the work on CPN to adapt it to the MANET environment with focus on path stability metric. CPN is a distributed adaptive routing protocol that uses three types of packets: Smart Packets for route discovery, Data Packets for carrying data payload, and Acknowledgments to bring back feedback information for the Reinforcement Learning reward function. The research defines a reward function as a combination of high stability and low delay path criteria to discover long-lived routes without disrupting the overall delay. The algorithm uses Acknowledgment-based Q-routing to make routing decisions which adapt on line to network changes allowing nodes to learn efficient routing policies.

Paper Nr: 55
Title:

Boosting of Neural Networks over MNIST Data

Authors:

Eva Volna, Vaclav Kocian and Martin Kotyrba

Abstract: The methods proposed in the article come out from a technique called boosting, which is based on the principle of combining a large number of so-called weak classifiers into a strong classifier. The article is focused on the possibility of increasing the efficiency of the algorithms via their appropriate combination, and particularly increasing their reliability and reducing their time exigency. Time exigency does not mean time exigency of the algorithm itself, nor its development, but time exigency of applying the algorithm to a particular problem domain. Simulations and experiments of the proposed processes were performed in the designed and created application environment. Experiments have been conducted over the MNIST database of handwritten digits that is commonly used for training and testing in the field of machine learning. Finally, a comparative experimental study with other approaches is presented. All achieved results are summarized in a conclusion.

Paper Nr: 67
Title:

The Parameter Selection and Average Run Length Computation for EWMA Control Charts

Authors:

Sheng Shu Cheng, Fong-Jung Yu, Shih-Ting Yang and Jiang-Liang Hou

Abstract: In the Statistical Process Control (SPC) field, an Exponentially Weighted Moving Average for Stationary processes (EWMAST) chart with proper control limits has been proposed to monitor the process mean of a stationary autocorrelated process. There are two issues of note when using the EWMAST charts. These are the smoothing parameter selections for the process mean shifts, and the determination of the control limits to meet the required average run length (ARL). In this paper, a guideline for selecting the smoothing parameter is discussed. These results can be used to select the optimal smoothing parameter in the EWMAST chart. Also, a numerical procedure using an integration approach is presented for the ARL computation with the specified control limits. The proposed approach is easy to implement and provides a good approximation to the average run length of EWMAST charts.

Paper Nr: 70
Title:

On the Claim for the Existence of “Adversarial Examples” in Deep Learning Neural Networks

Authors:

Costas Neocleous and Christos N. Schizas

Abstract: A recent article in which it is claimed that adversarial examples exist in deep artificial neural networks (ANN) is critically examined. The newly discovered properties of ANNs are critically evaluated. Specifically, we point that adversarial examples can be serious problems in critical applications of pattern recognition. Also, they may stall the further development of artificial neural networks. We challenge the absolute existence of these examples, as this has not been universally proven yet. We also suggest that ANN structures, that correctly recognize adversarial examples, can be developed.

Paper Nr: 74
Title:

Feature, Configuration, History - A Bio-inspired Framework for Information Representation in Neural Networks

Authors:

Frédéric Alexandre, Maxime Carrere and Randa Kassab

Abstract: Artificial Neural Networks are very efficient adaptive models but one of their recognized weaknesses is about information representation, often carried out in an input vector without a structure. Beyond the classical elaboration of a hierarchical representation in a series of layers, we report here inspiration from neuroscience and argue for the design of heterogenous neural networks, processing information at feature, configuration and history levels of granularity, and interacting very efficiently for high-level and complex decision making. This framework is built from known characteristics of the sensory cortex, the hippocampus and the prefrontal cortex and is exemplified here in the case of pavlovian conditioning, but we propose that it can be advantageously applied in a wider extent, to design flexible and versatile information processing with neuronal computation.

Paper Nr: 78
Title:

A New Artificial Neural Network Approach for Fluid Flow Simulations

Authors:

Osama Sabir and T. M. Y. S. Tuan Ya

Abstract: In this research we describe our attempt to get instantaneous numerical simulation for fluid flow by using Artificial Neural networks (ANN). Such simulation should provide a reliable perception about the fluid behaviour with respect to both momentum and energy equations. In addition to the preceding recorded data, the proposed method consider the geometrical boundaries profile as a major contributions for ANN training phase. Our study is driven by the need of rapid response especially in medical cases, surgeon diagnosis, engineering emergency situations, and when novel circumstances occurs. Furthermore, the existing computational fluid dynamics tools require long time to response and the present of professional expert to set the parameters for the different cases. In fact, ANN can deal with the lack of proper physical models or the present of uncertainty about some conditions that usually affect the outcomes form the other approaches. We manage to get acceptable result for 1D-flow equations with respect to both energy and momentum equations. Our ANN approach is able to handle fluid flow prediction with known boundaries velocity. This approach can be the first step for neural network computational program that can tackle variance type of problems.

Paper Nr: 84
Title:

Improving the Accuracy of Face Detection for Damaged Video and Distant Targets

Authors:

Jun-Horng Chen

Abstract: This work aims at improving the accuracy of face detection in two scenarios, when the video quality is deteriorated by the transmission link and when the target is far away from the camera. In block based coding, the packet loss inevitably makes the corrupted face image lacks some blocks. This work proposes the sparse modeling error concealment can coarsely recover the lost blocks, the fine texture can be obtained by diminishing the edge discontinuity, and a satisfied result for face detection can thus be recovered. Furthermore, this work utilizes the relationship learning super-resolution method to enhance the resolution in the case of face image taken from a long distance. Experimental results demonstrate that the proposed approach can effectively increase the accuracy of face detection for severely degraded and low resolution face images.

Paper Nr: 85
Title:

RNN-based Model for Self-adaptive Systems - The Emergence of Epilepsy in the Human Brain

Authors:

Emanuela Merelli and Marco Piangerelli

Abstract: The human brain is the self-adaptive system par excellence. We claim that a hierarchical model for self-adaptive system can be built on two levels, the upper structural level S and the lower behavioral level B. The higher order structure naturally emerges from interactions of the system with its environment and it acts as coordinator of local interactions among simple reactive elements. The lower level regards the topology of the network whose elements self-organize to perform the behavior of the system. The adaptivity feature follows the self-organizing principle that supports the entanglement of lower level elements and the higher order structure. The challenging idea in this position paper is to represent the two-level model as a second order Long Short-Term Memory Recurrent Neural Network, a bio-inspired class of artificial neural networks, very powerful for dealing with the dynamics of complex systems and for studying the emergence of brain activities. It is our aim to experiment the model over real Electrocorticographical data (EcoG) for detecting the emergence of long-term neurological disorders such as epileptic seizures.