Multidisciplinary Approaches to Neural Computing

Artificial Neural Networks (ANN) are currently exploited in many scientific domains. They had shown to act as doable, practical, and fault tolerant computational methodologies. They are equipped with solid theoretical background and proved to be effective in many demanding tasks such as approximating complex functions, optimizing search procedures, detecting changes in behaviors, recognizing familiar patterns, identifying data structures. ANNs computational limitations, essentially related to the presence of strong nonlinearities in the data and their poor generalization ability when provided of fully connected architectures, have been hammered by more sophisticated models, such as Modular Neural Networks (MNNs), and more complex learning procedures, such as deep learning. Given the multidisciplinary nature of their use and the interdisciplinary characterization of the problems they approach, ranging from medicine to psychology, industrial and social robotics, computer vision, and signal processing (among many others) ANNs may provide the bases for a redefinition of the concept of information processing. These reflections are supported by theoretical models and applications presented in the chapters of this book.

Stromboli volcano is considered one of the most active volcanoes in the world. During its effusive phases, it is possible to record a particular typology of events named “hybrid events”, that rarely are observed in the daily volcano activity. These ones are often associated to fault failure in the volcanic edifice due to magma movement and/or pressurization. Their identification, analysis and location can improve the volcano eruptive process comprehension. However, it is not easy to distinguish them from the other usually recorded events, i.e. explosion-quakes, through a visual seismogram analysis. Thus, we present an automatic supervised procedure, based on a Multi-layer Perceptron (MLP) neural network, to identify and discriminate them from the explosions-quakes. The data are encoded by using LPC coefficients and then adding to this coding waveform features. The 99% of accuracy was reached when waveform features are coded together with LPC coefficients as input to the network, emphasizing the importance of temporal features for discriminating hybrid events from explosion-quakes. The results allow us to assert that the proposed neural strategy can be included in a more complex automatic system for the monitoring of Stromboli volcano and of other volcanoes in the world.

Prostate imaging is a very critical issue in the clinical practice, especially for diagnosis, therapy, and staging of prostate cancer. Magnetic Resonance Imaging (MRI) can provide both morphologic and complementary functional information of tumor region. Manual detection and segmentation of prostate gland and carcinoma on multispectral MRI data is not easily practicable in the clinical routine because of the long times required by experienced radiologists to analyze several types of imaging data. In this paper, a fully automatic image segmentation method, exploiting an unsupervised Fuzzy C-Means (FCM) clustering technique for multispectral T1-weighted and T2-weighted MRI data processing, is proposed. This approach enables prostate segmentation and automatic gland volume calculation. Segmentation trials have been performed on a dataset composed of 7 patients affected by prostate cancer, using both area-based and distance-based metrics for its evaluation. The achieved experimental results are encouraging, showing good segmentation accuracy.

This paper presents a novel approach aiming at improving the White Matter (WM) fiber-bundle extraction approach described in (Stamile C et al Brain Informatics and Health: 8th International Conference, BIH, 2015). This provides anatomically coherent fiber-bundles, but it is unable to distinguish symmetric fiber-bundles. The new approach we are proposing here overcomes this limitation by integrating QuickBundles (QB) into it. As a matter of fact, QB has features complementary to those of the approach of (Stamile C et al Brain Informatics and Health: 8th International Conference, BIH, 2015), because it is capable of distinguishing symmetric fiber-bundles but, often, it does not return anatomically coherent fiber-bundles. We also present some experiments showing that the Precision, the Recall and the F-Measure of this new approach improve by 9.76, 3.08 and 8.96%, compared to the corresponding ones of the approach of (Stamile C et al Brain Informatics and Health: 8th International Conference, BIH, 2015), which, in their turn, were shown to be better than the ones of QB.

Metamaterials are artificial materials having uncommon physical properties. For a fast and careful design of these structures, the development of simple and faithful models able to reproduce their electromagnetic behavior is a key factor. Very recently a quick method for the extraction of Drude-Lorentz models for electromagnetic metamaterials has been presented [1]. In this work we improve that approach, introducing a novel procedure exploiting a micro-genetic algorithm ($$\mu $$GA). Numerical results obtained for the case of a split ring resonator structure cleary show a better reconstruction behaviour for equivalent magnetic permittivity $$\mu _{eff}$$ than those provided by [1].

In this paper, we derive a modified InfoMax algorithm for the solution of Blind Signal Separation (BSS) problems by using advanced stochastic methods. The proposed approach is based on a novel stochastic optimization approach known as the Adaptive Moment Estimation (Adam) algorithm. The proposed BSS solution can benefit from the excellent properties of the Adam approach. In order to derive the new learning rule, the Adam algorithm is introduced in the derivation of the cost function maximization in the standard InfoMax algorithm. The natural gradient adaptation is also considered. Finally, some experimental results show the effectiveness of the proposed approach.

The paper describes a depth-based hand pose recognizer by means of a Learning Vector Quantization (LVQ) classifier. The hand pose recognizer is composed of three modules. The first module segments the scene isolating the hand. The second one carries out the feature extraction, representing the hand by a set of 8 features. The third module, the classifier, is a LVQ. The recognizer, tested on a dataset of 6500 hand poses, carried out by people of different sex and physical aspect, has shown an accuracy larger than 99% recognition rate. The hand pose recognizer accuracy is among highest presented in literature for hand pose recognition.

The last decade of technological development has given raise to a myriad of new sensing devices able to measure in many ways the movements of human arms. Consequently, the number of applications in human health, robotics, virtual reality and gaming, involving the automatic recognition of the arm movements, has notably increased. The aim of this paper is to recognise the arm movements performed by jugglers during their exercises with three and four balls, on the basis of few information on the arm orientation given by Euler Angles, measured with a cheap sensor. The recognition is obtained through a linear Support Vector Machine after a feature extraction phase in which the reconstruction of the system dynamics is performed, thus estimating three Correlation Dimensions, corresponding to Euler Angles. The effectiveness of the proposed system is assessed through several experimentations.

Understanding the emerging of cortical dynamical state, its functional role, and its relationship with network topology, is one of the most interesting open questions in computational neuroscience. Spontaneous cortical dynamics often shows spontaneous fluctuations with UP/DOWN alternations and critical avalanches which resemble the critical fluctuations of a system posed near a non-equilibrium noise-induced phase transition. A model with structured connectivity and dynamical attractors has been shown to sustain two different dynamic states and a phase transition with critical behaviour is observed. We investigate here which are the features of the connectivity which permit the emergence of the phase transition and the large fluctuations near the critical line. We start from the original connectivity, that comes from the learning of the spatiotemporal patterns, and we shuffle the presynaptic units, leaving unchanged both the postsynaptic units and the value of the connections. The original structured network has a large clustering coefficient, since it has more directed connections which cooperate to activate a precise order of neurons, respect to randomized network. When we shuffle the connections we reduce the clustering coefficient and we destroy the spatiotemporal pattern attractors. We observe that the phase transition is gradually destroyed when we increase the ratio of shuffled connections, and already at a shuffling ratio of 70% both the phase transition and its critical features disappear.

Supporting people in their homes is an important issue both for ethical and practical reasons. Indeed, in the recent years, the scientific community devoted particular attention to detecting human falls, since the first cause of death for elderly people is due to the consequences of a fall. In this paper, we propose a human fall classification system based on an innovative floor acoustic sensor able to capture the acoustic waves transmitted through the floor. The algorithm employed is able to discriminate human falls from non falls and it is based on Mel-Frequency Cepstral Coefficients and a two class Support Vector Machine. The dataset employed for performance evaluation is composed by falls of a human mimicking doll, everyday objects and everyday noises. The obtained results show that the proposed solution is suitable for human fall detection in realistic scenarios, allowing to guarantee a 0% miss probability at very low false positive rates.

Learning in non-stationary/evolving environments requires methods able to process and deal with non-stationary streams. In this paper we propose a novel algorithm providing a time-frequency decomposition of time-variant signals. Outcoming signals can be used to identify anomalous events/patterns or extract features associated with the time-variance of the signal, precious information for any consequent learning action. The paper extends the Hilbert-Huang Transform notoriously used to deal with time-variant signals by introducing (i) a new Empirical Mode Decomposition that identifies the number of frequency modes of the signal and (ii) an extension of the Hilbert Transform that eliminates negative frequency-values in the time-frequency spectrum. The effectiveness of the proposed Transform has been tested on both synthetic and real time-variant signals acquired by a real-world intelligent system for landslide monitoring.

In this paper, we consider the application of data mining methods in medical contexts, wherein the data to be analysed (e.g. records from different patients) is distributed among multiple clinical parties. Although inference procedures could provide meaningful medical information (such as optimal clustering of the subjects), each party is forbidden to disclose its local dataset to a centralized location, due to privacy concerns over sensible portions of the dataset. To this end, we propose a general framework enabling the parties involved to perform (in a decentralized fashion) any data mining procedure relying solely on the Euclidean distance among patterns, including kernel methods, spectral clustering, and so on. Specifically, the problem is recast as a decentralized matrix completion problem, whose proposed solution does not require the presence of a centralized coordinator, and full privacy of the original data can be ensured by the use of different strategies, including random multiplicative updates for secure computation of distances. Experimental results support our proposal as an efficient tool for performing clustering and classification in distributed medical contexts. As an example, on the known Pima Indians Diabetes dataset, we obtain a Rand-Index for clustering of 0.52 against 0.54 of the (unfeasible) centralized solution, while on the Parkinson speech database we increase from 0.45 to 0.50.

In this paper we present an innovative procedure to reduce the number of rules in a Mamdani rule-based fuzzy systems. First of all, we extend the similarity measure or degree between antecedent and consequent of two rules. Subsequently, we use the similarity degree to compute two new measures of conflicting and reinforcement between fuzzy rules. We apply these conflicting and reinforcement measures to suitably reduce the number of rules. Namely, we merge two rules together if they are redundant, i.e. if both antecedent and consequence are similar together, repeating this operation until no similar rules exist, obtaining a reduced set of rules. Again, we remove from the reduced set the rule with conflict with other, i.e. if antecedent are similar and consequence not; among the two, we remove the one characterized by higher average conflict with all the rules in the reduced set.

The advance knowledge of future traffic load is helpful for network service providers to optimize the network resource and to recover the demand criteria. This paper presents the task of internet traffic prediction with three different architectures of Deep Belief Network (DBN). The artificial neural network is created with the depth of 4 hidden layers in each model to learn the nonlinear hierarchal essence present in the time series of internet traffic data. The deep learning in the network is executed with unsupervised pretraining of the layers. The emphasis is given to the topology of DBN that achieves excellent prediction accuracy. The adopted approach provides accurate traffic predictions while simulating the traffic data patterns and stochastic elements, achieving 0.028 Root Mean Square Error (RMSE) value on the test data set. To validate our choice for hidden layer size selection, further more experiments were done for chaotic time series prediction.

Dealing with time-varying high dimensional data is a big problem for real time pattern recognition. Only linear projections, like principal component analysis, are used in real time while nonlinear techniques need the whole database (offline). Their incremental variants do no work properly. The onCCA neural network addresses this problem; it is incremental and performs simultaneously the data quantization and projection by using the Curvilinear Component Analysis (CCA), a distance-preserving reduction technique. However, onCCA requires an initial architecture, provided by a small offline CCA. This paper presents a variant of onCCA, called growing CCA (GCCA), which has a self-organized incremental architecture adapting to the nonstationary data distribution. This is achieved by introducing the ideas of “seeds”, pairs of neurons which colonize the input domain, and “bridge”, a different kind of edge in the manifold graph, which signal the data nonstationarity. Some examples from artificial problems and a real application are given.

This paper focuses on employing Convolutional Neural Networks (CNN) with 3-D kernels for Voice Activity Detectors in multi-room domestic scenarios (mVAD). This technology is compared with the Multi Layer Perceptron (MLP) and interesting advancements are observed with respect to previous works of the authors. In order to approximate real-life scenarios, the DIRHA dataset is exploited. It has been recorded in a home environment by means of several microphones arranged in various rooms. Our study is composed by a multi-stage analysis focusing on the selection of the network size and the input microphones in relation with their number and position. Results are evaluated in terms of Speech Activity Detection error rate (SAD). The CNN-mVAD outperforms the other method with a significant solidity in terms of performance statistics, achieving in the best overall case a SAD equal to 7.0%.

Variable selection is an important concept in data mining, which can improve both the performance of machine learning and the process knowledge by removing the irrelevant and redundant features. The paper presents a hybrid variable selection approach that merges a combination of filters with a wrapper in order to obtain an informative subset of variables in a reasonable time, improving the stability of the single approach of more than 36% in average, without decreasing the system performance. The proposed method is tested on datasets coming from the UCI repository and from industrial contexts.

Many industrial tasks are related to the problem of the classification of unbalanced datasets. In these cases rare patterns of interest for the particular applications have to be detected among a much larger amount of patterns. Since data unbalance strongly affects the performance of standard classifiers, several ad–hoc methods have been developed. In this work the main techniques for handling class unbalance are depicted and three methods developed by the authors and based on the use of neural networks are described and tested on industrial case studies.

The Production Scheduling is an important phase in a manufacturing system, where the aim is to improve the productivity of one or more factories. Finding an optimal solution to scheduling problems means to approach complex combinatorial optimization problems, and not all of them are solvable in a mathematical way, in fact a lot of them are part of the class of NP-hard combinatorial problems. In this paper a joint mixed approach based on a joint use of Evolutionary Algorithms and their quantum version is proposed. The context is ideally located inside two factories, partners and use cases of the white’R FP7 FOF MNP Project, with high manual activity for the production of optoelectronics products, switching with the use of the new robotic (re)configurable island, the white’R, to highly automated production. This is the first paper approaching the problem of the dynamic production scheduling for these types of production systems proposing a cooperative solving method. Results show this mixed method provide better answers and is faster in convergence than others.

This paper presents a Neural Network (NN) approach for steam turbines modelling. NN models can predict generated power as well as different steam features that cannot be directly monitored through sensors, such as pressures and temperatures at drums outlet and steam quality. The investigated models have been trained and validated on a dataset created through the internal sizing design tool and tested by exploiting field data coming from a real-world power plant, in which a High Pressure and a Low Pressure turbines are installed. The proposed approach is applied to identify the variation of the characteristics from data measurable on the operating field, by means of suitable monitoring and control algorithms that are implemented directly on the PLC.

This paper presents analyses and test results of engine management system’s operational architecture with an artificial neural network (ANN). The research involved several steps of investigation: theory, a stand test of the engine, training of ANN with test data, generated from the proposed engine control system to predict the future values of fuel consumption before calculating the engine speed. In our paper, we study a small size 1.5 L gasoline engine without direct fuel injection (injection in intake manifold). The purpose of this study is to simplify engine and vehicle integration processes, decrease exhaust gas volume, decrease fuel consumption by optimizing cam timing and spark timing, and improve engine mechatronic functioning. The method followed in this work is applicable to small/medium size gasoline/diesel engines. The results show that the developed model achieved good accuracy on predicting the future demand of fuel consumption for engine control unit (ECU). It yields with the error rate of 1.12e-6 measured as Mean Square Error (MSE) on unseen samples.

The paper describes the application of a Self-Organising Map for the analysis and the interpretation of measurements taken by a Non-Destructive Testing system named IMPOC$${}^\circledR $$ and related to the hardness of steel coils after the hot rolling mill. This work addresses the problem of understanding whether distinct process conditions may lead to non-uniform mechanical properties along the coil. The proposed approach allows to point out, for each specific steel grade, some process conditions that are more frequently associated to disuniformities.

This paper proposes a methodology to visual assist a robotic arm to accurately micro-localize different optoelectronic components by means of object detection and recognition techniques. The various image processing tasks performed for the effective guidance of the robotic arm are analyzed under the scope of implementing a specific production scenario with localization accuracy in the scale of a few microns, proposed by a laser diode manufacturer. In order to elaborate the necessary procedures for achieving the required functionality, the required algorithms engaged in every step are presented. The analysis of possible implementations of the vision algorithms for achieving the required image recognition tasks take into account the possible content of the camera data, the accuracy of the results based on predefined specifications, as well as the computational complexity of the available algorithmic solutions.

The aim of this work is to analyze, through artificial neural network models, cortical pattern of women with Intimate Partner Violence (IPV) to investigate representative models of sensitization or habituation to the emotional stimulus in IPV. We investigate the ability of high emotional impact images, during a recognition task, analyzing the electroencephalogram data and event related potentials. Neural network analysis highlights an impairment in IPV group in cortical arousal, during the emotional recognition task. The alteration of this capacity has obvious repercussions on people’s lives, because it involves chronic difficulties in interpersonal relationships.

Mild cognitive impairment (MCI) is a neurological disorder that degenerates into Alzheimer’s disease (AD) in 8–15% of cases. The MCI to AD conversion is due to a loss of connectivity between different areas of the brain. In this paper, a wavelet coherence approach is proposed for investigating how the brain connectivity evolves among cortical regions with the disease progression. We studied Electroencephalograph (EEG) recordings acquired from eight patients affected by MCI at time T0 and we also studied their follow up at time T1 (three months later): three of them converted to AD, five remained MCI. The EEGs were analyzed over delta, theta, alpha 1, alpha 2, beta 1 and beta 2 sub-bands. Differently from MCI stable subjects, MCI patients who converted to AD, showed a strong reduction of cortical connectivity in theta, alpha(s) and beta(s) sub-bands. Delta band showed high coherence values in each pair of electrodes in every patient.

In clinical practice, study of brain functions is fundamental to notice several diseases potentially dangerous for the health of the subject. Electroencephalography (EEG) can be used to detect cerebral disorders but EEG study is often difficult to implement, taking into account the multivariate and non-stationary nature of the signals and the invariable presence of noise. In the field of Signal Processing exist many algorithms and methods to analyze and classify signals reducing and extracting useful information. Support Vector Machine (SVM) based algorithms can be used as classification tool and allow to obtain an efficient discrimination between different pathology and to support physicians while studying patients. In this paper, we report an experience on designing and using an SVM based algorithm to study and classify EEG signals. We focus on Creutzfeldt-Jakob disease (CJD) EEG signals. To reduce the dimensionality of the dataset, principal component analysis (PCA) is used. These vectors are used as inputs for the SVM classifier with two classification classes: pathologic or healthy. The classification accuracy reaches 96.67% and a validation test has been performed, using unclassified EEG data.

Alzheimer’s disease (AD) is a neurological degenerative disorder that causes the impairment of memory, behaviour and cognitive abilities. AD is considered a cortical disease because it causes the loss of functional connections between the cortical regions. Electroencephalography (EEG) consists in recording, non-invasively, the electrical potentials produced by neuronal activity. EEG is used in the evaluation of AD patients because they show peculiar EEG features. The EEG traces of AD patients usually exhibit a shift of the power spectrum to lower frequencies as well as reduced coherence between the cortical areas. This is the reason why AD is defined as “disconnection disorder”. However, the correct interpretation of the EEG can be very challenging because of the presence of “artifacts”, undesired signals that overlap to the EEG signals generated by the brain. Removing artifacts is therefore crucial in EEG processing. Recently, the author contributed to develop an automatic EEG artifact rejection methodology called Enhanced Automatic Wavelet Independent Component Analysis (EAWICA) which achieved very good performance on both simulated and real EEG from healthy subjects (controls). The aim of the present paper is to test EAWICA on real EEG from AD patients. According to the expert physician’s feedback, EAWICA efficiently removed the artifacts while saving the diagnostic information embedded in the EEG and not affecting the segments that were originally artifact free.

The purpose of this work is to investigate the olfactory response to a neuter and a smell stimulation through Olfactory Event Related Potentials (OERP). We arranged an experiment of olfactory stimulation by analyzing Event Related Potential during perception of 2 odor stimuli: pleasant (Rose, 2-phenyl ethanol C₂H₄O₂) and neuter (Neuter, Vaseline Oil CH₂). We recruited 15 adult safe non-smokers volunteers. In order to record OERP, we used VOS EEG, a new device dedicated to odorous stimulation in EEG. After the OERP task, the subject filled a visual analogic scale, regarding the administered smell, on three dimensions: pleasantness (P), arousing (A) and familiarity (F). We performed an artificial neural network analysis that highlighted three groups of significant features, one for each amplitude component. Three neural network classifiers were evaluated in terms of accuracy on both full and restricted datasets, showing the best performance with the latter. The improvement of the accuracy rate in all VAS classifications was: 13.93% (A), 64.81% (F), 9.8% (P) for P300 amplitude (Fz); 16.28% (A), 49.46% (F), 24% (P) for N400 amplitude (Cz, Fz, O2, P8); 110.42% (A), 21.19% (F), 24.1% (P) for N600 amplitude (Cz, Fz). Main results suggested that in smell presentation we can observe the involvement of slow Event-Related-Potentials, like N400 and N600, ERP involved in stimulus encoding.

This paper investigates the acoustical and perceptual speech features that differentiate a depressed individual from a healthy one. The speech data gathered was a collection from both healthy and depressed subjects in the Italian language, each comprising of a read and spontaneous narrative. The pre-processing of this dataset was done using Mel Frequency Cepstral Coefficient (MFCC). The speech samples were further processed using Principal Component Analysis (PCA) for correlation and dimensionality reduction. It was found that both groups differed with respect to the extracted speech features. To distinguish the depressed group from the healthy one on the basis the proposed speech processing algorithm the Self Organizing Map (SOM) algorithm was used. The clustering accuracy given by SOM’s was 80.67%.

In this study we challenged the universal view of facial emotion perception evaluating the effects of gender and different luminance backgrounds on the recognition accuracy of neutral facial expressions. To this aim, we applied the Ekman standard paradigm for assessing the human ability to decode neutral facial expressions reproduced on black, white and grey backgrounds and portrayed by male and female actors. The exploited stimuli consisted of 10 different neutral faces (5 females) selected from the Dutch Radboud database (Langner et al. Cogn Emot, 2010 [21]) where luminance backgrounds were changed in black, grey and white. The resulted 30 stimuli were assessed by 31 subjects (16 females) who were asked to tag each of them with one of the six primary emotion labels. The data analysis demonstrates a significant gender effect where neutral male faces are less accurately decoded than females ones. On the other hand, no effects of luminance backgrounds have been identified.

Twitter is an irreplaceable source of data for opinion mining, emergency communications, or fact sharing, whose readability is severely limited by the sheer volume of tweets published every day. A method to represent and synthesize the information content of conversations on Twitter in form of semantic maps, from which the main topics and the main orientations of tweeters may easily be read, is proposed hereafter. After a preliminary grouping of tweets in conversations, relevant keywords and Named Entities are extracted, disambiguated and clustered. Annotations are made using extensive knowledge bases and state-of-the-art techniques from Natural Language Processing and Machine Learning. The results are in form of coloured graphs, to be easily interpretable. Several experiments confirm the high understandability and the good adherence to tackled topics of the mapped conversations.

The aim of this study is to describe a new database acquired in two different elementary schools of Barcelona province. The study assessed the effect of the type of handwriting learning in general writing performance. In the first school, classical cursive handwriting is learnt while the second one substitutes this skill for keyboarding and print-script handwriting. Analyses in two different groups of age (8–9 and 11–12 years old) for both schools have been performed. A set of 14 different handwriting tasks has been acquired for each student using an Intuos Wacom series 4 tablet plus ink pen and specific software to conduct the analysis. The results revealed that cursive handwriting might improve the handwriting performance by increasing the speed of writing and drawing.

Two studies were conducted during the Italian General Elections of 2006 (N = 179) and 2008 (N = 607), to investigate the relationships among implicit and explicit attitudes, voting intention and voting behaviour. Several structural equation models that included direct and indirect effect of implicit and explicit attitudes toward political objects (coalitions and leaders) on voting intention and behaviour were executed to test a prediction model of political preferences and voting behaviour. Notwithstanding some differences, the results of the two studies showed that (i) the implicit evaluations of political objects are more differentiated than the explicit ones; (ii) that implicit attitudes contribute in a specific and additive way to determine the voting intention and behaviour, and (iii) that the effect of the implicit attitude is also mediated by the explicit attitudes. Findings are discussed in the frame of dual cognition models and in the light of the peculiar political scenarios of the considered electoral process.

Through two experiments we investigated, in a laboratory setting, whether a series of identical outcomes in a supposed random game would induce the gambler’s fallacy or the hot-hand fallacy. By using two indices of fallacy, the choice of a card on which to bet and the probability estimate of the occurrence of a given outcome, we tested explicitly the widely accepted hypothesis that the two fallacies were based on erroneous probability estimates. Moreover, we investigated whether fallacies increase the proneness to bet. Our results support the occurrence of the gambler’s fallacy rather than the hot-hand fallacy but suggest that choice and probability estimates are two reciprocally independent processes. Finally, probability estimates predict the amount bet.

In the last 30 years, the systematic analysis of human thought has provided new evidences on intuition’s nature. It has been observed in experimental level that in front of decision-making problems, most people unknowingly adopt adaptive solutions that are different from logical inferences of normative rationality. To cope with the temporal and cognitive limitations, humans always use heuristic strategies that allow them to gather quickly useful information for survival. Naturally formal logic can lead to adequate choices, but its processes are slow and cognitively expensive. In this paper we intend to show how, in specific situations and contexts, the paths of formal logic and of natural logic (heuristics, intuitions and so on) diverge dramatically.

A large amount of theoretical and experimental research—from dynamic systems to computational neurosciences, from statistical learning to psychobiology of development—indicates that the encounter between Humans and very powerful AI will lead, in the near future, to organisms capable of going over the simulation of brain functions: hybrids that will learn from their internal states, will interpret the facts of reality, establish their goals, talk with humans and, especially, will decide according to their own ‘system of values’. Soon the traditional symbolic-formal domain could be overcome by the construction of systems with central control functions, with cognition similar to the biological brain. This requires a clarification on how they will act and, mostly, how they will decide. But what do we know today about decision-making processes and what is their relationship with the emotional spheres? If, traditionally, emotions were considered separate from the logical and rational thought, in recent years we have begun to understand that emotions have deep influence on human decisions. In this paper, we intend to show how emotions are crucial in moral decisions and that their understanding may help us to avoid mistakes in the construction of hybrid organisms capable of autonomous behavior.