Neural Nets and Surroundings

We focus on learning the probability matrix for discrete random variables in factor graphs. We review the problem and its variational approximation and, via entropic priors, we show that soft quantization can be included in a probabilistically-consistent fashion in a factor graph that learns the mutual relationship among the variables involved. The framework is explained with reference the ”Tipper” example and the results of a Matlab simulation are included.

A major recurring problem in exploratory phases of data mining is the task of finding the number of clusters in a dataset. In this paper we illustrate a variant of the competitive clustering method which introduces a rival penalization mechanism, and show how it can be used to solve such problem. Additionally, we present some tests aimed at comparing the performance of our rival-penalized technique with other classical procedures.

This paper describes how the classification of imbalanced datasets through support vector machines using the boundary movement method can be easily explained in terms of a cost-sensitive learning algorithm characterized by giving each example a cost in function of its class. Moreover, it is shown that under this interpretation the boundary movement is measured in terms of the squared norm of the separator’s slopes in feature space, thus providing practical insights in order to properly choose the boundary surface shift.

We present a multi-purpose genetic algorithm, designed and implemented with GPGPU / CUDA parallel computing technology. The model was derived from a multi-core CPU serial implementation, named GAME, already scientifically successfully tested and validated on astrophysical massive data classification problems, through a web application resource (DAMEWARE), specialized in data mining based on Machine Learning paradigms. Since genetic algorithms are inherently parallel, the GPGPU computing paradigm has provided an exploit of the internal training features of the model, permitting a strong optimization in terms of processing performances and scalability.

In this paper we investigate the introduction of Reservoir Computing (RC) neural network models in the context of AAL (Ambient Assisted Living) and self-learning robot ecologies, with a focus on the computational constraints related to the implementation over a network of sensors. Specifically, we experimentally study the relationship between architectural parameters influencing the computational cost of the models and the performance on a task of user movements prediction from sensors signal streams. The RC shows favorable scaling properties results for the analyzed AAL task.

Computational finance is one of the fields where machine learning and data mining have found in recent years a large application. Neverthless, there are still many open issues regarding the predictability of the stock market, and the possibility to build an automatic intelligent trader able to make forecasts on stock prices, and to develop a profitable trading strategy. In this paper, we propose an automatic trading strategy based on support vector machines, which employs recall-precision curves in order to allow a buying action for the trader only when the confidence of the prediction is high. We present an extensive experimental evaluation which compares our trader with several classic competitors.

Most of the scientist assume that epileptic seizures are triggered by an abnormal electrical activity of groups of neural populations that yields to dynamic changes in the properties of Electroencephalography (EEG) signals. To understand the pathogenesis of the epileptic seizures, it is useful detect them by using a tool able to identify the dynamic changes in EEG recordings. In the last years, many measures in the complex network theory have been developed. The aim of this paper is the use of Permutation Entropy (PE) with the addition of a threshold method to create links between the different electrodes placed over the scalp, in order to simulate the network phenomena that occur in the brain. This technique was tested over two EEG recordings: a healthy subject and an epileptic subject affected by absence seizures.

This work proposes a preliminary study of an automatic recognition system for the Italian Sign Language (Lingua Italiana dei Segni - LIS). Several other attempts have been made in the literature, but they are typically oriented to international languages. The system is composed of a feature extraction stage, and a sign recognition stage. Each sign is represeted by a single Hidden Markov Model, with parameters estimated through the resubstitution method. Then, starting from a set of features related to the position and the shape of head and hands, the Sequential Forward Selection technique has been applied to obtain feature vectors with the minimum dimension and the best recognition performance. Experiments have been performed using the cross-validation method on the Italian Sign Language Database A3LIS-147, maintaining the orthogonality between training and test sets. The obtained recognition accuracy averaged across all signers is 47.24%, which represents an encouraging result and demonstrates the effectiveness of the idea.

A random ensemble of random perceptrons is studied and applied in fall detection and categorization, an important and growing problem in Ambient Assisted Living and other fields related to the care of elder and in general of “fragile” people. The classifier ensemble is designed around an ECOC aggregator and compensates for the lack of an accurate training with the number of base learners, which increases accuracy and strengthens the error-correcting capabilities of class codewords. The approach is suitable when some memory is available, but computational power is limited: this is the standard situation in mobile computing, and to an even larger extent in wearable computing. Performances on the two applicative tasks of fall recognition (dichotomic) and categorization (multi-class) are compared with those of support vector machines.

Short term prediction of air pollution is gaining increasing attention in the research community, due to its social and economical impact. In this paper we study the application of a Kernel Adaptive Filtering (KAF) algorithm to the problem of predicting PM

10

data in the Italian province of Ancona, and we show how this predictor is able to achieve a significant low error with the inclusion of chemical data correlated with the PM

10

such as NO

2

.

In this paper we propose a filter combination for the adaptive noise cancellation (ANC) problem in nonlinear environment. The architecture consists in a convex combination of two adaptive filters: a classical filter and a nonlinear filter based on Functional Links. While the convergence of the linear filter is very fast, the convergence of the nonlinear one might be slower, even if it provides a more accurate solution. The convex combination of both filters allows to reach good performances in terms of convergence and speed. In addition a variable step size is used in order to obtain better performance. Several experimental results, in different reverberant conditions, demonstrate the effectiveness of the proposed approach.

In the period from June to September 2011, the Stromboli volcano was affected by an activity characterized by an increase of the volcanic tremor amplitude, in the magnitude of explosions and with some lava overflows. In order to examine and understand in more detail this particular phase of the volcano, we present here an unsupervised investigation of the waveform variation of the explosion-quakes recorded during this period. The aim is to identify a possible relationship between the temporal changes of these events and the volcano seismic activity. The analysis is performed on a dataset of about 8400 explosion-quakes by using a SOM neural network. This technique works well with large datasets allowing to find out unpredicted characteristics among them. The SOM clustering highlights sudden changes occurring at the end of July and of August and a permanent variation between June and September reflecting a modification in the volcano activity. These results could be interesting for focusing the analysis of the seismological dataset in these intervals in order to evidence minor, but important variations, which were previously undetected and to improve the knowledge on the explosive dynamics of the volcano.

Multiple Sclerosis (MS) is an autoimmune condition in which the immune system attacks the Central Nervous System. Magnetic Resonance Imaging (MRI) is today a crucial tool for diagnosis of MS by allowing in-vivo detection of lesions. New lesions may represent new inflammation; they may increase in size during acute phase to contract later while the disease severity is reduced. To monitor evolution in time of lesions and to correlate this to MS phases, we focused on the application of Artificial Neural Network (ANN) based classification of MS lesions. An euclidean distance histogram, representing the distribution of edge inter-pixel distances, is used as input. In this work, we have extended the study already published, increasing to 21 the number of images. We can observe that the percentage of correct results on 21 images (93.81%) increased if compared to the study performed on 13 images (92.31%). This methodology could be used to monitor evolution in time of lesions of each patient and to correlate this to MS phases (i.e. to know if the lesions change their form).

Automated video surveillance using video analysis and understanding technology has become an important research topic in the area of computer vision. Most cameras used in surveillance are fixed, allowing to only look at one specific view of the surveilled area. Recently, the progress in sensor technologies is leading to a growing dissemination of Pan-Tilt-Zoom (PTZ) cameras, that can dynamically modify their field of view. Since PTZ cameras are mainly used for object detection and tracking, it is important to extract moving object regions from images taken with this type of camera. However, this is a challenging task because of the dynamic background caused by camera motion.

After reviewing background subtraction-based approaches to moving object detection in image sequences taken from PTZ cameras, we present a neural-based background subtraction approach where the background model automatically adapts in a self-organizing way to changes in the scene background. Experiments conducted on real image sequences demonstrate the effectiveness of the presented approach.

Standardization and the assessment of the quality of the final product is fundamental in food industry. Coffee particle properties are monitored continuously during coffee beans grinding. Operators control the grinders in order to keep coffee particle granulometry within specific thresholds. In this work, a general regression neural network approach is used to learn to control two grinders used for coffee production at LAVAZZA factory, obtaining average control error of the order of a few

μ

m. The results appear promising for the future development of an automatic decision support system.

In-line automated inspection of raw materials is one of the major concerns in food industry. The aim of this paper is to devise a method to sort pistachio nuts, in order to reject the bad ones. X-ray images are used to compute a set of fuzzy features, with membership functions automatically inferred from the positive samples. A functional-link neural network is then used for the proper classification task. By means of a repeated cross-validation, the proposed solution showed a correct recognition rate of 99.6%, with a false positive rate of 0.3% with a single classifier and 0.1% with a combined one.

This paper presents a real-time hand gesture recognizer based on a

Learning Vector Quantization

(

LVQ

) classifier. The recognizer is formed by two modules. The first module, mainly composed of a data glove, performs the feature extraction. The second module, the classifier, is performed by means of LVQ. The recognizer, tested on a dataset of 3900 hand gestures, performed by people of different gender and physique, has shown very high recognition rate.

Single nucleotide polymorphisms (SNPs) are the foremost part of many genome association studies. Selecting a subset of SNPs that is sufficiently informative but still small enough to reduce the genotyping overhead is an important step towards disease-gene association. In this work, a Random Forest (RF) approach to informative SNPs selection in Familial Combined Hyperlipidemia (FCH) is proposed. FCH is the most common form of familial hyperlipidemia. Affected patients have elevated levels of plasma triglycerides and/or total cholesterol and show increased risk of premature coronary heart disease. In order to identify susceptibility markers for FCH we perform the analysis of 21 SNPs in ten genes associated with high cardiovascular risk. RF appears to be a useful technique in identifying gene polymorphisms involved in FCH: the identified SNPs confirmed some variants in a couple of genes as genetic markers of FCH as proved in various studies in scientific literature and lead us to report for the first time a further gene association with FCH. This result could be promising and encourages to further investigate on the role of the identified gene in the development of FCH phenotype.

The graph classification problem consists, given a weighted graph and a partial node labeling, in extending the labels to all nodes. In many real-world context, such as Gene Function Prediction, the partial labeling is unbalanced: positive labels are much less than negatives. In this paper we present a new neural algorithm for predicting labels in presence of label imbalance. This algorithm is based on a family of Hopfield networks, described by 2 continuous parameters and 1 discrete parameter, and it consists of two main steps: 1) the network parameters are learnt through a cost-sensitive optimization procedure based on local search; 2) a suitable Hopfield network restricted to unlabeled nodes is considered and simulated. The reached equilibrium point induces the classification of unlabeled nodes. An experimental analysis on real-world unbalanced data in the context of genome-wide prediction of gene functions show the effectiveness of the proposed approach.

The human brain processes information showing learning and prediction abilities but the underlying neuronal mechanisms still remain unknown. Recently, many studies prove that neuronal networks are able of both generalizations and associations of sensory inputs.

In this paper, following a set of neurophysiological evidences, we propose a learning framework with a strong biological plausibility that mimics prominent functions of cortical circuitries. We developed the Inductive Conceptual Network (ICN), that is a hierarchical bio-inspired network, able to learn invariant patterns by Variable-order Markov Models implemented in its nodes. The outputs of the top-most node of ICN hierarchy, representing the highest input generalization, allow for automatic classification of inputs. We found that the ICN clusterized MNIST images with an error of 5.73% and USPS images with an error of 12.56%.

Net migration is the net total of migrants during the period, that is, the total number of immigrants less the annual number of emigrants, including both citizens and noncitizens. To derive estimates of net migration, the United Nations Population Division takes into account the past migration history of a country or area, the migration policy of a country, and the influx of refugees in recent periods. The data to calculate these official estimates come from a variety of sources, including border statistics, administrative records, surveys, and censuses. When no official estimates can be made because of insufficient data, net migration is derived through the balance equation, which is the difference between overall population growth and the natural increase during the intercensal period. In this contribution, we apply the functional data model to Italian data, for forecasting net migration numbers.

By 1951, average fertility had fallen to just over two children per woman, and only five percent of children would die in their first ten years of life. A similar pattern of declining fertility and mortality rates, collectively known as the demographic transition, has been observed in every industrializing country. Financial projections for Social Security systems depend on many demographic, economic and social factors as well as the reduction of fertility rates and the ageing of a population. In order to address the need to develop reliable projections, it is unavoidable to detect appropriate measures to represent the future trends of the quantities of interest. The aim of the paper is apply to Italian data a mathematical scheme suitable for projecting the fertility rates and for measuring the uncertainty around these estimates. Finally a numerical application is provided.

The aim of this note is to provide a global performance index that allows to evaluate the performance of each faculty member and which is able to consider the multidimensional nature of the academic activity in terms of research, teaching and other activities that academics should/might exercise. In order to model also the case in which there could be synergic and redundant connections among the different areas of the academic activity, we propose to use fuzzy measures and the Choquet integral as an aggregator of the different components.

This paper investigates the use of feed forward neural networks for testing the weak form market efficiency. In contrast to approaches that compare out-of-sample predictions of non-linear models to those generated by the random walk model, we directly focus on testing for unpredictability by considering the null hypothesis that a given set of past lags has no effect on current returns. To avoid the data-snooping problem the testing procedure is based on the StepM approach in order to control the familiwise error rate. The procedure is used to test for predictive power in FTSE-MIB index of the italian stock market.

The inherent low conversion efficiency, from solar to electrical energy, of the photovoltaic cells makes the use of techniques and architectures aimed at maximizing the electrical power a photovoltaic array is able to produce at any weather condition mandatory. In order to understand what are the challenging problems cropping up in some modern applications, an overview of the main techniques for photovoltaic arrays modeling is given first. Afterwards, the control strategies for the maximum power point tracking used in commercial products dedicated to photovoltaic strings and modules are compared and their advantages and drawbacks are put into evidence, with a special emphasis on their efficiency. Some methods presented in literature and based on the use of artificial neural networks are compared with more classical ones. Finally, a brief overview of other applications of artificial neural networks to photovoltaic-related problems is also given.

The paper deals with the proposal of a new architecture, called Sustainable Energy Microsystem (SEM), for a smart grid project in urban context. SEM includes energy sub-systems (SS) currently independent, such as high efficiency buildings, sustainable mobility systems (Electrical Vehicles and metro transit-systems), dispersed generation from renewables and Combined Heat and Power units. The present paper includes the description of the main SEM elements and some results of an energy analysis on each subsystem, showing the effective possibilities of integration, aimed to energy saving and environmental sustainability.

The current electrical grid is undergoing a deep renovation that poses new problems in terms of technologies, communication and control. The increasing level of penetration of renewable energy is leading towards the concept of distributed energy production, and it is expected that Virtual Power Plants (VPPs) will play an important role in the future

smart grid

. The stochastic nature of the power flows in the VPP, caused by the fluctuating availability of renewables, by the users’ demand and by the energy market price, complicates the task of power balancing for the VPP. This paper proposes the use of simple learning mechanisms to support power scheduling decisions and to improve a correct supply of the connected loads.

In this paper the problem of the minimization of active power losses in a real Smart Grid located in the area of Rome is faced by defining and solving a suited multi-objective optimization problem. It is considered a portion of the

ACEA Distribuzione S.p.A.

network which presents backflow of active power for 20% of the annual operative time. The network taken into consideration includes about 100 nodes, 25 km of MV lines, three feeders and three distributed energy sources (two biogas generators and one photovoltaic plant). The grid has been accurately modeled and simulated in the phasor domain by Matlab/Simulink, relying on the SimPowerSystems ToolBox, following a Multi-Level Hierarchical and Modular approach. It is faced the problem of finding the optimal network parameters that minimize the total active power losses in the network, without violating operative constraints on voltages and currents. To this aim it is adopted a genetic algorithm, defining a suited fitness function. Tests have been performed by feeding the simulation environment with real data concerning dissipated and generated active and reactive power values. First results are encouraging and show that the proposed optimization technique can be adopted as the core of a hierarchical Smart Grid control system.

An on-line prediction algorithm able to estimate, over a determined time horizon, the solar irradiation of a specific site is considered. The learning algorithm is based on Radial Basis Function (RBF) networks and combines the growing criterion and the pruning strategy of the minimal resource allocating network technique. An adaptive extended Kalman filter is used to update all the parameters of the Neural Network (NN). The on-line learning mechanism avoids the initial training of the NN with a large data set. The proposed solution has been experimentally tested on a 14 kWp PhotoVoltaic (PV) plant and results are compared to a classical RBF neural network.

Energy saving is nowadays a mandatory requirement for buildings. In this paper, a novel holistic system for intelligent Smart Home environments is introduced, which is able to embrace energy production and consumption in a unique perspective and to handle them in conjunction with device and services management. Semantic Web technologies are employed to foster data interoperability and to supply inference power for intelligent tasks management, decisions making and energy saving. The proposed system uses an IP-based network as main communication channel and a semantic extension of the UPnP technology, to support zero-configuration, automatic discovering and intelligent control. The ontology framework, used to formally describe all relevant information of the Smart Home environment like devices, services and context, is also presented focusing on the device and the energy ontology. In addition, the issue of the lack of Semantic Web compliant device descriptions that currently are not provided by device vendors, posing a serious barrier toward the practical application of semantic technologies in Smart Home scenario, is also tackled.

Nowadays a correct use of energy is a crucial aspect, in fact cost and energy waste reduction are the main goals that must be achieved. To reach this objective an optimal energy management must be obtained through some techniques and optimization algorithms, in order to provide the best solution in terms of cost. In this work a comparison between different methods for energy scheduling is proposed and some analytical results are reported, in order to offer a clear overview for each technique, in terms of advantages and disadvantages. A residential scenario is considered for computer simulations, in which a system storage and renewable resources are available and exploitable to match the user load demand.

The recognition of human emotions by technical systems is regarded as a problem of pattern recognition. Here methods of machine learning are employed which require substantial amounts of ’emotionally labeled’ data, because model based approaches are not available. Problems of emotion recognition are discussed from this point of view, focusing on problems of data gathering and also touching upon modeling of emotions and machine learning aspects.

This article proposes experiments on decision making based on the “Winter Survival Task”, one of the scenarios most commonly applied in behavioral and psychological studies. The goal of the Task is to identify, out of a predefined list of 12 items, those that are most likely to increase the chances of survival after the crash of a plane in a polar area. In our experiments, 60 pairs of unacquainted individuals (120 subjects in total) negotiate a common choice of the items to be retained after that each subject has performed the task individually. The results of the negotiations are analyzed in causal terms and show that the choices made by the subjects individually act as a causal factor with respect to the outcome of the negotiation.

In the field of biology, the study of bonding has been renewed by the discovery of non genetic transmission of behavioural traits through early mother-infant interaction and the role of stress hormones and ocytocin. However, the study of early interaction is complex and Social Signal Processing (SSP) can help in addressing some issues. Based on works from our group, we will show data from diverse sources (e.g. experiments, home movies) showing how SSP was used to address synchrony between partners (e.g. infant, child, care giver, agent) and characteristics that participates to interpersonal exchanges (e.g. motherese, emotional prosody or faces).

Writing processes depend on the development and the capacity of working memory. Their execution is highly costly in cognitive resources. During writing, emotions are potentially present. According to Ellis and Ashbrook’s (1988) model, emotions are expected to cause interferences in working memory by creating extra cognitive load. Our main hypothesis was that emotions should be compared to a secondary task, overloading working memory capacities. Two experiments using emotional induction procedures were carried out on two different writing tasks (text production and dictation) with young graders. Results showed that emotional content interfered as cognitive overload within the limited working memory resources and had an impact on orthographic abilities. In terms of computational intelligence, as emotions seem to have an impact on the availability of cognitive resources, this could lead to important theoretical and practical implications for the elaboration of interactive scenarios or modeling learning and processing procedures.

In this paper we present preliminary results about on-line drawings acquired by means of a digitizing tablet, and performed by control population (left and right hand) as well as pathological subjects using their dominant hand. Experimental results reveal a clear difference between both groups, specially on the on-air movements. Although the acquired samples are not enough to extract significant conclusions we think that this preliminary results encourage the experimentation in this research line. Thus, the main purpose of this paper is to attract the attention of the scientific community.

This paper presents a novel method for gender recognition through anthropometric hand information. From a visual hand database of a hundred users and distributed in an unbalanced way, contains more men than women. It is designed a simple method to get some length and width measurements from the hand. This information has been passed through a quadratic discriminant classifier called Biometric Dispersion Matcher (BDM) that provides relevant information. In a first step, a discriminative threshold is applied in order to discard those measures which do not have enough information for gender recognition. In a second step, it provides a vector of the main measures. And, finally, it achieves performance rates from 95%, with a train data set of only 18 men and 9 women, to 98%, with a higher training data set.

Combining information from multiple sources is a vivid field of research. The problem of emotion recognition is inherently multi-modal. As automatic recognition of the emotional states is performed imperfectly by the single mode classifiers, its combination is crucial. In this work, the AVEC 2011 corpus is used to evaluate several machine learning techniques in the context of information fusion. In particular temporal integration of intermediate results combined with a reject option based on classifier confidences. The results for the modes are combined using a Markov random field that is designed to be able to tackle failures of individual channels.

This paper intends to contribute to the research on the perception of emotion with a case study on the recognition of realistic vs. stylized facial emotional expressions in typical developing three and six-year-old children. In particular, it reports on two perceptual experiments aimed at evaluating children’s ability in identifying human and stylized male and female facial emotional expressions of happiness, anger and surprise. Results show that six-years-old children are able to recognize facial expressions of happiness and anger exploiting stylized as well as realistic human figures, preferring stylized faces for the identification of surprise. Three-year-old children are not able to recognize surprise and are significantly better in recognizing happiness rather than anger, suggesting that the ability to recognize certain emotional faces emerges through experience. In addition, it also suggests that this ability is not affected by face gender.

Authors present parts of the work on the development of multimodal HCI control interface targeted at paralyzed people. In this paper the proposed emotional status determination solution is presented based on the data gathering from eye tracking sensors. The participants were provided with an audiovisual stimulus (slides/videos with sounds) and the emotional feedback was determined by the combination of gaze tracking and artificial neural network processing. Some initial experimental analysis and the data of the recognition accuracy of the emotional state based on the gaze tracking are provided along with the description of implemented algorithms.

Past research has demonstrated intercultural differences in emoticon use with effects of the topic of discourse (e.g. science vs. politics) interacting with the culture of online postings (e.g. UK, Italy, Sweden, Germany). The current research focuses within a discourse, and within a lingua franca for communication and attempts to assess whether emoticon use varies as a function of user-type within the online context. The online context is a web user-forum associated with a software technology company. The user categories are determined by a few orthogonal classifications: employees, novice users, and experts; recipients of kudos vs. non-recipients of kudos; etc. As part of a developing theory of presentation of “professional” selves, and perceptions thereof, we test the hypotheses that kudo recipients deploy markedly fewer negative emoticons than comparison categories and that non-employee experts use markedly more emoticons in general than other categories of forum users. Also interactivity across the different group of users and their correlation with emoticon use was explored.

Social interaction is one of the main channels to access reality and information about people. In this last years there is a growing interest in community websites that combine social interaction with music and entertainment exploration. Music is a language of emotions and music emotional recognition has been addressed by different disciplines (psychology, cognitive science and musicology). Aim of this work is to introduce a framework for music emotion recognition based on machine learning and soft computing techniques. First, musical emotional features are extract from audio songs and successively they are elaborated for classification or clustering. One user can submit a target song, representing his conceptual emotion, and to obtain a playlist of audio songs with similar emotional content. In the case of classification, a playlist is obtained from the songs of the same class. In the other case, the playlist is suggested by system exploiting the content of the audio songs and it could also contain songs of different classes. Several experiments are proposed to show the performance of the developed framework.

Since the ancient past, philosophers and scientists have developed a particular kind of representations aiming at describing the human nature. Many of these images could be considered as visual metaphors, so it is possible to study them using two different approaches. The first one aims at analysing the connection existing between these images, the scientific paradigms, and other elements of the cultural context within which they emerged. The second approach strives to underline some specific elements that characterise the inherent power of these visual metaphors, ensuring their success in scientific and cultural context.