Tesi di Dottorato

Permanent URI for this communityTesi di Dottorato

Browse

Has files: YesSubject: search.filters.subject.Machine learning

Search Results

Now showing 1 - 6 of 6
  • No Thumbnail Available
    Item
    Arsenic Ore Mixture Froth Image Generation with Neural Networks and a Language for Declarative Data Validation
    (Università della Calabria, 2022-04-14) Zamayla, Arnel; Greco, Gianluigi; Alviano, Mario; Dodaro, Carmine
    Computer vision systems that measure froth flow velocities and stability designed for flotation froth image analysis are well established in industry, as they are used to control material recovery. However flotation systems that has limited data has not been explored in the same fashion bearing the fact that big data tools like deep convolutional neural networks require huge amounts of data. This lead to the motivation of the research reported in the first part of this thesis, which is to generate synthetic images from limited data in order to create a froth image dataset. The image synthesis is possible through the use of generative adversarial network. The performance of human experts in this domain in identifying the original and synthesized froth images were then compared with the performance of the models. The models exhibited better accuracy levels by average on the tests that were performed. The trained classifier was also compared with some of the established neural network models in the literature like the AlexNet, VGG16 ang ResNet34. Transfer learning was used as a method for this purpose. It also showed that these pretrained networks that are readily available have better accuracy by average comapared to trained experts. The second part of this thesis reports on a language designed for data validation in the context of knowledge representation and reasoning. Specifically, the target language is Answer Set Programming (ASP), a logic-based programming language widely adopted for combinatorial search and optimization, which however lacks constructs for data validation. The language presented in this thesis fulfills this gap by introducing specific constructs for common validation criteria, and also supports the integration of consolidated validation libraries written in Python. Moreover, the language is designed so to inject data validation in ordinary ASP programs, so to promote fail-fast techniques at coding time without imposing any lag on the deployed system if data are pretended to be valid.
  • No Thumbnail Available
    Item
    A comparative study of reconstruction methods for neutron tomography
    (Università della Calabria, 2020-01-24) Micieli, Davide; Carbone, Vincenzo; Gorini, Giuseppe; Tassi, Enrico
    Neutron tomography is a well established technique to non-destructively investigate the inner structure of a wide range of objects. The main disadvantages of this technique are the time-consuming data acquisition, which generally requires several hours, and the low signal to noise ratio of the acquired images. One way for decreasing the total scan time is to reduce the number of radiographs. However, the Filtered Back-Projection, which is the most widely used reconstruction method in neutron tomography, generates low quality images affected by artifacts when the number of projections is limited or the signal to noise ratio of the radiographs is low. This doctoral thesis is focused on the comparative analysis of different reconstruction techniques, aimed at finding the data processing procedures suitable for neutron tomography that shorten the scan time without reduction of the reconstructed image quality. At first the performance of the algebraic reconstruction methods were tested using experimental neutron data and studied as a function of the number of projections and for different setups of the imaging system. The reconstructed images were quantitatively compared in terms of image quality indexes. Subsequently, the recently introduced Neural Network Filtered Back-Projection method was proposed in order to reduce the acquisition time during a neutron tomography experiment. This is the first study which proposes and tests a machine learning based reconstruction method for neutron tomography. The Neural Network Filtered Back-Projection method was quantitatively compared to conventional reconstruction algorithms used in neutron tomography. Finally, we present NeuTomPy, a new Python package for tomographic data processing and reconstruction. NeuTomPy is a cross-platform toolbox ready to work with neutron data. The first release of NeuTomPy includes pre-processing algorithms, a wide range of classical and state-of-the-art reconstruction methods and several image quality indexes, in order to evaluate the reconstruction quality. This software is free and open-source, hence researchers can freely use it and contribute to the project.
  • No Thumbnail Available
    Item
    Technologies and IoT Protocols applied to Energy Management in Smart Home Environment
    (Università della Calabria, 2021-09-13) Serianni, Abdon; De Ranco, Floriano
    This thesis presents the studies during this period of my PhD course. In the first research period, I focused the activities principally on the study and analysis of the protocols and technologies used for the IoT solutions in Smart Home environment. It was analyzed the MQTT protocol and its possible applications. The MQTT protocol uses the event-driven publish/subscribe pattern. In our tests, MQTT usage was compared with a classic HTTP request/response paradigm, used in REST and CoAP approaches. A layered IoT communication architecture will be proposed and described. The usage of proposed IoT communication architecture was analyzed in Smart Home context and in other application contexts such as e-Health and Internet of Vehicles (IoV). After an analysis of Data Mining and Machine learning concepts, the focus of the activities was on Neural Networks. The use of LSTM networks was analyzed for time-series forecasting and prediction of consumption in two different environments (home and office). In the smart home environment, smart objects are characterized by limited resources. Our proposal to increase the computational capabilities of these smart devices is a hidden cognitive object that uses pre-trained NN and continuous learning for anomaly detection and suggested action prediction tasks. The Cognitive Smart Object is the joining of a smart device and a hidden cognitive object. The Cognitive Smart Object was used in thermal comfort control application and manage better energy consumption. The concepts introduced have been used for an assisted comfort solution and the neural network results were used to suggest to the user conventional management of the climatic comfort levels. A Continuous Learning mechanism was been implemented with the usage of user feedback to shape the neural network and obtain a neural network that follows user behaviours that diverge from behaviour compliant with the ASHRAE standard. From the analysis of the results obtained, it was possible to highlight how NN has given results closer to the user’s habits and at the same time the user has been educated to use the right levels of thermal comfort.
  • No Thumbnail Available
    Item
    Ensemble learning techniques for cyber security applications
    (2017-07-13) Pisani, Francesco Sergio; Crupi, Felice; Folino, Gianluigi
    Cyber security involves protecting information and systems from major cyber threats; frequently, some high-level techniques, such as for instance data mining techniques, are be used to efficiently fight, alleviate the effect or to prevent the action of the cybercriminals. In particular, classification can be efficiently used for many cyber security application, i.e. in intrusion detection systems, in the analysis of the user behavior, risk and attack analysis, etc. However, the complexity and the diversity of modern systems opened a wide range of new issues difficult to address. In fact, security softwares have to deal with missing data, privacy limitation and heterogeneous sources. Therefore, it would be really unlikely a single classification algorithm will perform well for all the types of data, especially in presence of changes and with constraints of real time and scalability. To this aim, this thesis proposes a framework based on the ensemble paradigm to cope with these problems. Ensemble is a learning paradigm where multiple learners are trained for the same task by a learning algorithm, and the predictions of the learners are combined for dealing with new unseen instances. The ensemble method helps to reduce the variance of the error, the bias, and the dependence from a single dataset; furthermore, it can be build in an incremental way and it is apt to distributed implementations. It is also particularly suitable for distributed intrusion detection, because it permits to build a network profile by combining different classifiers that together provide complementary information. However, the phase of building of the ensemble could be computationally expensive as when new data arrives, it is necessary to restart the training phase. For this reason, the framework is based on Genetic Programming to evolve a function for combining the classifiers composing the ensemble, having some attractive characteristics. First, the models composing the ensemble can be trained only on a portion of the training set, and then they can be combined and used without any extra phase of training. Moreover the models can be specialized for a single class and they can be designed to handle the difficult problems of unbalanced classes and missing data. In case of changes in the data, the function can be recomputed in an incrementally way, with a moderate computational effort and, in a streaming environment, drift strategies can be used to update the models. In addition, all the phases of the algorithm are distributed and can exploits the advantages of running on parallel/ distributed architectures to cope with real time constraints. The framework is oriented and specialized towards cyber security applications. For this reason, the algorithm is designed to work with missing data, unbalanced classes, models specialized on some tasks and model working with streaming data. Two typical scenarios in the cyber security domain are provided and some experiment are conducted on artificial and real datasets to test the effectiveness of the approach. The first scenario deals with user behavior. The actions taken by users could lead to data breaches and the damages could have a very high cost. The second scenario deals with intrusion detection system. In this research area, the ensemble paradigm is a very new technique and the researcher must completely understand the advantages of this solution.
  • No Thumbnail Available
    Item
    Seamless acceleratin numerical regular grid methods on manycore systems
    (2018-01-19) Spataro, Davide; Leone, Nicola; Spataro, William; D'Ambrosio, Donato
    Over the last two decades, a lot has changed regarding the way modern scientific applications are designed, written and executed, especially in the field of data-analytics, scientific computing and visualization. Dedicated computing machines are nowadays large, powerful agglomerates of hundreds or thousands of multi-core computing nodes interconnected via network each coupled with multiple accelerators. Those kinds of parallel machines are very complex and their efficient programming is hard, bug-prone and time-consuming. In the field of scientific computing, and of modeling and simulation especially, parallel machines are used to obtain approximate numerical solutions to differential equations for which the classical approach often fails to solve them analytically making a numerical computer-based approach absolutely necessary. An approximate numerical solution of a partial differential equation can be obtained by applying a number of methods, as the finite element or finite difference method which yields approximate values of the unknowns at a discrete number of points over the domain. When large domains are considered, big parallel machines are required in order to process the resulting huge amount of mesh nodes. Parallel programming is notoriously complex, often requiring great programming efforts in order to obtain efficient solvers targeting large computing cluster. This is especially true since heterogeneous hardware and GPGPU has become mainstream. The main thrust of this work is the creation of a programming abstraction and a runtime library for seamless implementation of numerical methods on regular grids targeting different computer architecture: from commodity single-core laptops to large clusters of heterogeneous accelerators. A framework, OpenCAL had been developed, which exposes a domain specific language for the definition of a large class of numerical models and their subsequent deployment on the targeted machines. Architecture programming details are abstracted from the programmer that with little or no intervention at all can obtain a serial, multi-core, single-GPU, multi- GPUs and cluster of GPUs OpenCAL application. Results show that the framework is effective in reducing programmer effort in producing efficient parallel numerical solvers.
  • No Thumbnail Available
    Item
    User behavioral problems in complex social networks
    (2019-06-20) Perna, Diego; Tagarelli, Andrea; Crupi, Felice
    Over the past two decades, we witnessed the advent and the rapid growth of numerous social networking platforms. Their pervasive diffusion dramatically changed the way we communicate and socialize with each other. They introduce new paradigms and impose new constraints within their scope. On the other hand, online social networks (OSNs) provide scientists an unprecedented opportunity to observe, in a controlled way, human behaviors. The goal of the research project described in this thesis is to design and develop tools in the context of network science and machine learning, to analyze, characterize and ultimately describe user behaviors in OSNs. After a brief review of network-science centrality measures and ranking algorithms, we examine the role of trust in OSNs, by proposing a new inference method for controversial situations. Afterward, we delve into social boundary spanning theory and define a ranking algorithm to rank and consequently identify users characterized by alternate behavior across OSNs. The second part of this thesis deals with machine-learning-based approaches to solve problems of learning a ranking function to identify lurkers and bots in OSNs. In the last part of this thesis, we discuss methods and techniques on how to learn a new representational space of entities in a multilayer social network.