Netpro2vec is a neural embedding framework, based on probability distribution representations of graphs. The goal is to look at node descriptions, such as those induced by the Transition Matrix and Node Distance Distribution. Netpro2vec provides embeddings completely independent from the task and nature of the data. The framework is evaluated on synthetic and various real biomedical network datasets through a comprehensive experimental classification phase and is compared to well-known competitors.

Nowadays, a lot of data is in the form of Knowledge Graphs aiming at representing information as a set of nodes and relationships between them. This paper proposes an efficient framework to create informative embeddings for node classification on large knowledge graphs. Such embeddings capture how a particular node of the graph interacts with his neighborhood and indicate if it is either isolated or part of a bigger clique. Since a homogeneous graph is necessary to perform this kind of analysis, the framework exploits the metapath approach to split the heterogeneous graph into multiple homogeneous graphs. The proposed pipeline includes an unsupervised attentive neural network to merge different metapaths and produce node embeddings suitable for classification. Preliminary experiments on the IMDb dataset demonstrate the validity of the proposed approach, which can defeat current state-of-the-art unsupervised methods.

Pupil dynamics alterations have been found in patients affected by a variety of neuropsychiatric conditions, including autism. Studies in mouse models have used pupillometry for phenotypic assessment and as a proxy for arousal. Both in mice and humans, pupillometry is non-invasive and allows for longitudinal experiments supporting temporal specificity, however, its measure requires dedicated setups. Here, we introduce a Convolutional Neural Network that performs online pupillometry in both mice and humans in a web app format. This solution dramatically simplifies the usage of the tool for the non-specialist and non-technical operators. Because a modern web browser is the only software requirement, this choice is of great interest given its easy deployment and set-up time reduction. The tested model performances indicate that the tool is sensitive enough to detect both locomotor-induced and stimulus-evoked pupillary changes, and its output is comparable with state-of-the-art commercial devices.

In future burning plasma devices measuring the plasma position and its shape with great spatial and temporal resolution is a crucial task. Nowadays these information are obtained by means of magnetic coils installed inside the vacuum vessel that in the future devices (like ITER and DEMO), due to the harsh environment caused by the high plasma temperature, will experience degradation over the time. Furthermore, the long plasma discharges will result in large integration drift causing a lack of accuracy in the measured signal. In order to assist the magnetic diagnostics and at the same time provide a novel tool to benchmark them, here the measurement of the plasma magnetic axis position by means of a collimated neutron flux monitor is proposed. Three different methods are here described and applied on JET by means of the neutron camera: a weighted average, the asymmetry method and a neural network. The methods are calibrated on a large database of plasma discharges including NBI and ICRH heated ones, and then compared with the magnetic axis position reconstructed by EFIT. The neural network outperforms the two other methods. In particular, the asymmetry method results to be sensitive to MHD activity, NBI power variation and to neutron emissivity profiles presenting a strong asymmetry, such as in case of peripheral NBI deposition due to high density plasmas or ICRH resonance layer. A possible application to vertical displacement events and disruptions is discussed and envisaged for future applications on DEMO. Finally, the performances of the neural network and of the asymmetry methods are studied in the case of one or two missing channels in the neutron flux monitor, showing how in general the reconstruction of the radial magnetic axis in both methods is more sensitive to the lack of channels than the vertical one. The methods here proposed can be applied and benchmarked on DTT and ITER neutron cameras as part of a future real-time control system.

Nowadays more and more elderly people need support in daily activities. This is due to the increase of cognitive diseases and other conditions which lead the elderly to not being self-sufficient. Considering this, providing an Ambient Assisted Living system could improve significantly people life quality and could support caregivers' tasks. The combination of Ambient Assisted Living systems and information and communication technologies achieve this purpose perfectly. They exploit internet of things and artificial intelligence paradigms to make daily challenges easier for people with neurodegenerative diseases. This work melds technologies mentioned above providing a smart system for elderly to manage goods and fill in shopping lists. It was possible using software, hardware, and cloud systems combined with a neural network aimed to recognise products. The proposed system has been validated both from a functional point of view through a proof-of-concept and quantitatively by a performance analysis of its components.

This work proposes a model for suggesting optimal process configuration in plunge centreless grinding operations. Seven different approaches were implemented and compared: first principles model, neural network model with one hidden layer, support vector regression model with polynomial kernel function, Gaussian process regression model and hybrid versions of those three models. The first approach is based on an enhancement of the well-known numerical process simulation of geometrical instability. The model takes into account raw workpiece profile and possible wheel-workpiece loss of contact, which introduces an inherent limitation on the resulting profile waviness. Physical models, because of epistemic errors due to neglected or oversimplified functional relationships, can be too approximated for being considered in industrial applications. Moreover, in deterministic models, uncertainties affecting the various parameters are not explicitly considered. Complexity in centreless grinding models arises from phenomena like contact length dependency on local compliance, contact force and grinding wheel roughness, unpredicted material properties of the grinding wheel and workpiece, precision of the manual setup done by the operator, wheel wear and nature of wheel wear. In order to improve the overall model prediction accuracy and allow automated continuous learning, several machine learning techniques have been investigated: a Bayesian regularized neural network, an SVR model and a GPR model. To exploit the a priori knowledge embedded in physical models, hybrid models are proposed, where neural network, SVR and GPR models are fed by the nominal process parameters enriched with the roundness predicted by the first principle model. Those hybrid models result in an improved prediction capability.

Reflectance transformation imaging (RTI) is a computational photography technique widely used in the cultural heritage and material science domains to characterize relieved surfaces. It basically consists of capturing multiple images from a fixed viewpoint with varying lights. Handling the potentially huge amount of information stored in an RTI acquisition that consists typically of 50-100 RGB values per pixel, allowing data exchange, interactive visualization, and material analysis, is not easy. The solution used in practical applications consists of creating "relightable images" by approximating the pixel information with a function of the light direction, encoded with a small number of parameters. This encoding allows the estimation of images relighted from novel, arbitrary lights, with a quality that, however, is not always satisfactory. In this paper, we present NeuralRTI, a framework for pixel-based encoding and relighting of RTI data. Using a simple autoencoder architecture, we show that it is possible to obtain a highly compressed representation that better preserves the original information and provides increased quality of virtual images relighted from novel directions, especially in the case of challenging glossy materials. We also address the problem of validating the relight quality on different surfaces, proposing a specific benchmark, SynthRTI, including image collections synthetically created with physical-based rendering and featuring objects with different materials and geometric complexity. On this dataset and as well on a collection of real acquisitions performed on heterogeneous surfaces, we demonstrate the advantages of the proposed relightable image encoding.

We summarize the main scientific contributions of our friend and colleague Alfredo Petrosino, full professor in computer science at the University of Naples Parthenope, Italy. They mainly cover topics in high-performance computing, neural network models, soft and granular computing, computer vision, and machine learning. We also highlight how most of his research activity lays the foundation for biometry and its applications.

Background: Logic Learning Machine (LLM) is an innovative method of supervised analysis capable of constructing models based on simple and intelligible rules. In this investigation the performance of LLM in classifying patients with cancer was evaluated using a set of eight publicly available gene expression databases for cancer diagnosis. LLM accuracy was assessed by summary ROC curve (sROC) analysis and estimated by the area under an sROC curve (sAUC). Its performance was compared in cross validation with that of standard supervised methods, namely: decision tree, artificial neural network, support vector machine (SVM) and k-nearest neighbor classifier. Results: LLM showed an excellent accuracy (sAUC = 0.99, 95%CI: 0.98-1.0) and outperformed any other method except SVM. Conclusions: LLM is a new powerful tool for the analysis of gene expression data for cancer diagnosis. Simple rules generated by LLM could contribute to a better understanding of cancer biology, potentially addressing therapeutic approaches.

This paper presents an experimental study where a vision camera integrates coaxially into a laser beam welding tool to monitor beam deviations (beam offset) in laser stake welding of T-joints. The aim is to obtain an early detection of deviations from the joint centreline in this type of welding where the joint is not visible from the top side. A polynomial surface fitting method is applied to extract features that can describe the behaviour of the melt pool. A nonlinear autoregressive with exogenous inputs neural network model is trained to relate eight image features to the laser beam offset. The performance of the presented model is evaluated offline by different welding samples. The results show that the proposed method can be used to guide post weld inspection and has the potential for on-line adaptive control.

The availability of accurate day-ahead energy prices forecasts is crucial to achieve a successful participation to liberalized electricity markets. Moreover, forecasting systems providing prediction intervals and densities (i.e. probabilistic forecasting) are fundamental to enable enhanced bidding and planning strategies considering uncertainty explicitly. Nonetheless, the vast majority of available approaches focus on point forecast. Therefore, we propose a novel methodology for probabilistic energy price forecast based on Bayesian deep learning techniques. A specific training method has been deployed to guarantee scalability to complex network architectures. Moreover, we developed a model originally supporting heteroscedasticity, thus avoiding the common homoscedastic assumption with related preprocessing effort. Experiments have been performed on two day ahead markets characterized by different behaviors. Then, we demonstrated the capability of the proposed method to achieve robust performances in out-of-sample conditions while providing forecast uncertainty indications.

The stability of the completely synchronous state in neural networks with electrical coupling is analytically investigated applying both the Master Stability Function approach (MSF), developed by Pecora and Carroll (1998), and the Connection Graph Stability method (CGS) proposed by Belykh et al. (2004). The local dynamics is described by Morris-Lecar model for spiking neurons and by Hindmarsh-Rose model in spike, burst, irregular spike and irregular burst regimes. The combined application of both CGS and MSF methods provides an efficient estimate of the synchronization thresholds, namely bounds for the coupling strength ranges in which the synchronous state is stable. In all the considered cases, we observe that high values of coupling strength tend to synchronize the system. Furthermore, we observe a correlation between the single node attractor and the local stability properties given by MSF. The analytical results are compared with numerical simulations on a sample network, with excellent agreement.

Simplicity is a basic principle of science and this implies that, if we want to explain the behaviour of animals by constructing robots that behave like real animals, one and the same robot should reproduce as many behaviours and as many behavioural phenomena as possible. In this paper we describe robots that both evolve and learn in their natural environment and, in addition, learn in the equivalent of an experimental laboratory and reproduce a variety of results of experiments on learning in animals. We introduce a new model of learning in which the weights of the connections that link the units of the robots neural network are genetically inherited and do not change during the robots life but what changes during life and makes the robots learn new behaviours is the synaptic receptivity of a special set of network units which we call learning units. The robots evolve in a variety of different environments and they learn in a variety of different ways including imprinting and learning by imitating the behaviour of others. Then we test the robots in the controlled conditions of an artificial laboratory and we reproduce a number of experimental results on both operant learning and classical conditioning, including learning and extinction curves, the role of the temporal interval between conditioned and unconditioned stimuli, and the influence of motivation on learning.

In recent years, a significant part of the studies on air pollutants has been devoted to improve statistical techniques for forecasting the values of their concentrations in the atmosphere. Reliable predictions of pollutant trends are essential not only for setting up preventive measures able to avoid risks for human health but also for helping stakeholders to take decision about traffic limitations. In this paper, we present an operating procedure, including both pollutant concentration measurements (CO, SO<inf>2</inf>, NO<inf>2</inf>, O<inf>3</inf>, PM10) and meteorological parameters (hourly data of atmospheric pressure, relative humidity, wind speed), which improves the simple use of neural network for the prediction of pollutant concentration trends by means of the integration of multivariate statistical analysis. In particular, we used principal component analysis in order to define an unconstrained mix of variables able to improve the performance of the model. The developed procedure is particularly suitable for characterizing the investigated phenomena at a local scale.

We propose the 3dSOBS+ algorithm, a newly designed approach for moving object detection based on a neural background model automatically generated by a self-organizing method. The algorithm is able to accurately handle scenes containing moving backgrounds, gradual illumination variations, and shadows cast by moving objects, and is robust against false detections for different types of videos taken with stationary cameras. Experimental results and comparisons conducted on the Background Models Challenge benchmark dataset demonstrate the improvements achieved by the proposed algorithm, that compares well with the state-of-the-art methods. (C) 2013 Elsevier Inc. All rights reserved.

[object Object]

Electronic engine controls based on real time diagnosis of combustion process can significantly help incompling with the stricter and stricter regulations on pollutants emissions and fuel consumption.Themost important parameter for hee valuation of combustion quality in internal combustion engine sisthein-cylinder pressure, butits direct measurement is very expensive and involves an intrusive approach to the cylinder.Previous researches demonstrated the direct relationship existing between in-cylinder pressure and engine crankshaft speed and several authors triedtore construct the pressure cycle on the basis of the engine speed signal. In this paper we propose the use of a Multi-Layer Percept ronneural network to model the relationship between the engine crank shaft speed and some parameters derived from the in-cylinder pressure cycle.This allows to have an on-intrusive estimation of cylinder pressure and are altime evaluation of combustion quality. Thestructureofthemodel and thetrainingprocedureisoutlinedinthepaper.Apossiblecombustioncontroller usingtheinformationextractedfromthecrankshaftspeedinformationisalsoproposed. The application of the neural network model is demonstrated on a single-cylinder spark ignition engine tested in a wider ange of speeds and loads.Results confirm tha ta good estimation of some combustion pressure parameters can be obtained by means of a suitable processing of crank shaft speed signal.

Acoustic methods are well established and widely used for the exploration of the seafloor and the sub-bottom sediments. However, the mapping and reconstruction of the sedimentary features revealed by acoustics can require a very long time because often large acoustic datasets need to be described and interpreted. To reduce the time of the geophysical visual interpretation, we implemented a new procedure for facies classification based on wavelet analysis and neural networks applied to the acoustic profiles. The optimized algorithm applied to a data set of the very shallow Lagoon of Venice classifies automatically the echo shape parameters to identify and map the main lagoon sedimentary features, such as palaeochannels and palaeosurfaces. The classification algorithm contains a set of wavelet transformation parameters as inputs to a neural network analysis based on the self-organizing map (SOM). The analysis was applied on 580 km of acoustic profiles acquired in a very shallow (less than 1 m) and turbid area of the lagoon with a sub-bottom penetration of about 6-7 m under the bottom. Without any special pre-requirement on the data, the algorithm was successfully tested against the results of the visual interpretation and allowed an automated and more efficient full 2D mapping of the sedimentary features of the area. We could distinguish and map different types of palaeochannels, buried creeks, palaeosurfaces as well as areas characterized by homogeneous mudflat facies. The results were validated by comparison with 5 cores sampled in the survey area corresponding with the main sedimentary features revealed by the acoustics.

Fuel cells based on polymer electrolyte membrane are considered as the most hopeful clean power technology. The operating principles of polymer electrolyte membrane fuel cells (PEMFC) system involve electrochemistry, thermodynamics and hydrodynamics theory for which it is difficult to establish a mathematical model. In this paper a nonlinear data driven model of a PEMFC stack is developed using Neural Networks (NNs). The model presented is a black-box model, based on a set of measurable exogenous inputs and is able to predict the output voltage and cathode temperature of a high power module working at the CNR- ITAE. The device in this study was a 5 kW NUVERA PEM fuel cell stack consisting of 50 cells with 500 cm2 of geometric area each (Fig. 1). The stack works with a Cathode Water Injection system that dissipate heat and humidifies the cathode side by mixing air stream and water. In this work the stack was operated in a test station that includes an electronic load, gas mass flow controllers and pressure sensors and actuators, and a control and monitor software. The station has allowed to modify a number of process variables and to obtain an experimental database about the PEMFC stack under different operating conditions. To obtain a significant stack voltage dynamic the load current was changed from 30 to 110 A, the air mass flow from 35 to 184 slpm, the hydrogen mass flow from 19 to 104 slpm and nitrogen mass flow from 7 to 39 slpm in a number of different combinations. To train the different neural models, a set of patterns with sampling time T=10 sec., covering different process working points, was considered. A trial and error approach was used to select the best model among possible candidates. All the networks were trained using the Levemberg-Marquardt algorithm, with the early stopping approach to prevent overfitting. A trial and error approach was used to select the best model among possible candidates. Several everal sub-optimal neural models were been obtained, corresponding to a different number of hidden neurons and/or different set of learning pattern. The neural model with the higher correlation coefficients has 12 hidden neurons and correlation coefficients (Acquired / Simulated Data) for the output voltage and cathode temperature are 99% and 98% respectively. The data driven obtained model performed quite satisfactory and stack voltage and cathode temperature dynamics were simulated with accuracy. In Fig.2 and Fig.3 the comparison of stack voltage and ccathode temperature acquired data and corresponding model estimation for a Test data subset are shown.The trained NN model is computationally fast and easy to use, especially in the case where physical models are not readily available.

We propose a 3D self organizing neural model for modeling both the background and the foreground in video, helping in distinguishing between moving and stopped objects in the scene. Our aim is to detect foreground objects in digital image sequences taken from stationary cameras and to distinguish them into moving and stopped objects by a model based approach. We show through experimental results that a good discrimination can be achieved for color video sequences that represent typical situations critical for vehicles stopped in no parking areas. © 2009 The authors and IOS Press. All rights reserved.