Search

U ovom radu opisan je razvoj računalnog agenta za igru zmija, čija pravila su objašnjena na početku rada. Agent je razvijen uz pomoć neuronskih mreža i genetskih algoritama, a arhitektura mreže i način odvijanja selekcije, križanja i mutacije opisan je u odgovarajućim poglavljima. Izrađeno je grafičko korisničko sučelje koje omogućuje igranje od strane čovjeka ili praćenje procesa treniranja agenta. Grafički su prikazani dobiveni rezultati u procesu treniranja i testiranja. This paper describes the development of a computer agent for the snake game, the rules of which are explained at the beginning of the paper. The agent is developed from neural networks and genetic algorithms, the architecture of the network and the methods of selection, crossover and mutation are described in the corresponding chapters. A graphical user interface has been created that allows playing by a human or monitoring the process of training an agent. The results obtained in the process of training and testing are shown graphically.

Kroz ovaj rad je prikazan razvoj konvolucijske neuronske mreže primjenjujući VGG arhitekturu u problemu rješavanja klasifikacije slika. Slike pasa i mačaka su zadane iz poznate baze podataka zadane u svrhu Kaggle-ovog natjecanja iz strojnog učenja iz 2013. godine. U uvodu rada se upoznajemo s osnovnom gradivnom jedinicom neuronske mreže perceptorom i osnovnom principima njezinog rada. U nastavku će se predstaviti opća arhitektura konvolucijskih neuronskih mreža. Naglasak ovog rada je na savladavanju principa razvoja osnovnog modela VGG arhitekture neuronske mreže s kojom se želi prikazati rast preciznosti klasifikacije mreže dodavanjem novih blokova. Prikazat će se razvoj osnovnog modela kroz korištenje cloud servisa i online sučelja Google Colab koristeći se otvorenim softverskim bibliotekama za strojno učenje (Tenserflow). Potom se na istoj bazi podataka testira već unaprijed istrenirana arhitektura neuronske mreže VGG-16 te se uspoređuju rezultati s osnovnim modelom. This paper presents the development of a convolutional neural network applying the VGG architecture to the problem of image classification. The images of dogs and cats were taken from a well-known database provided for Kaggle's 2013 machine learning competition. In first part of the paper, we are introduced to the basic building block of the neural network, the preceptor, and the basic principles of its operation. Furthermore, the general architecture of convolutional neural networks will be presented. The emphasis of this paper is on mastering the principles of developing the basic model of the VGG architecture of the neural network, with which we want to show the growth of the accuracy of the network classification by stacking new blocks. The development of the basic model will be presented through the use of cloud services and the Google Colab online interface using open software libraries for machine learning (TensorFlow). Then, the pre-trained VGG-16 neural network architecture is tested on the same database and the results are compared with the basic model.

Zadatak ovog rada je saznati na koji način bi se oblik 3D objekta izmijenio kako bi se poboljšala preciznost određivanja njegove orijentacije. Prikazivanje slika, manipuliranje 3D objektima i potreban okvir za povezivanje 2D slika i 3D objekata je ostvaren bibliotekom PyTorch3D. Opisan je diferencijabilni prikazivač, ekstraktor značajki EfficientNet, konvolucijska neuronska mreža i parametri i funkcije korišteni u funkciji gubitka neuronske mreže. Na kraju su prikazani rezultati testiranja modela i ovisnost izgleda 3D objekta o raznim kombinacijama parametara s naglaskom na funkciju gubitka. The task of this paper is to find out how the shape of a 3D object would change in order to improve the accuracy of determining its orientation. The image generation, manipulation of 3D objects and the necessary framework for connecting 2D images and 3D objects is realized through the PyTorch3D library. A description of the differentiable renderer, EfficientNet feature extractor, convolutional neural network and the parameters and functions used in the neural network loss function is provided. In the end, the paper shows the results of the model and the dependence of the appearance of the 3D object on various combinations of parameters with emphasis on the loss function.

Problem klasifikacije fragmenata datoteka jedan je od ključnih problema kojim se bavi računalna forenzika. Tehnike koje rješavaju ovaj problem zahtijevaju puno računalnih resursa i vremena kako bi proizvele dobre rezultate. U ovom radu proučava se utjecaj koeficijenata Fourierove transformacije na uspješnost klasifikacije neuronskim mrežama s fokusom na umjetne neuronske mreže. Implementirano je više arhitektura i oformljeno više različitih ulaznih skupova podataka. Dane su usporedbe i komentari rezultata za konstruirane mreže. File fragment classification problem is one of the main problems in computer forensics. The techniques that produce good results are very computationally expensive and time-consuming. This paper studies the impact of Fourier transform coefficients on classification of neural networks, focusing on ANNs. Multiple network architectures were implemented, and multiple sets of input data were formed. The paper compares and comments on the results for each neural network model.

Posljednjih nekoliko godina tehnologije zasnovane na umjetnoj inteligenciji brzo su napredovale, tijekom prethodna dva desetljeća svjedoci smo velikog napretka u umjetnoj inteligenciji i njenoj primjeni. Neki od najranijih radova u uspješnoj primjeni umjetne inteligencije dogodili su se upravo u medicinskom kontekstu. U radu su objašnjene osnove umjetne inteligencije, strojnog i dubokog učenja. Dat je pregled umjetne inteligencije u medicini te prednosti njenog korištenja u raznim područjima medicine. Objašnjeno je područje računalnog vida s naglaskom na detekciju objekta te su opisana neka istraživanja u tom području. Opisani su algoritmi strojnog učenja, neuronske mreže, umjetne neuronske mreže te konvolucijske neuronske mreže. U sklopu rada izrađena je aplikacija koja prepoznaje kirurške instrumente. U izradi aplikacije korišteni su razvojni alati: Cloud Annotations, Colaboratory i Node-RED. Za trening modela poslužila je biblioteka strojnog učenja TensorFlow i programski jezik Python. Model strojnog učenja u TensorFlow.js formatu, ukomponiran je s grafičkim sučeljem, izrađenim koristeći razvojni alat Node-RED i programski jezik JavaScript. Aplikacija korisniku omogućuje izravno fotografiranje slike te učitavanje slike s računala za njeno prepoznavanje. Aplikacija prepoznaje pet različitih kirurških instrumenata: zakrivljene kirurške škare, ravne kirurške škare, skalpel, kiruršku pincetu te hvataljke. In the last few years, technologies based on artificial intelligence have advanced rapidly, and over the past two decades, we have witnessed great progress in artificial intelligence and its application. Some of the earliest work in the successful application of artificial intelligence occurred precisely in the medical context. This paper explains the basics of artificial intelligence, machine, and deep learning. An overview of artificial intelligence in the medical field and the advantages of its use in various fields of medicine is given. The area of computer vision is explained with an emphasis on object detection. The paper describes machine learning algorithms, neural networks, artificial neural networks, and convolutional neural networks. As part of this thesis, was developed an application that classifies surgical instruments. The following development tools were used in the development of the application: Cloud Annotations, Colaboratory, and Node-RED. The TensorFlow machine learning library was used for model training and the Python programming language. The machine learning model in TensorFlow.js format is integrated with a graphical interface, that is created using the development tool Node-RED and the JavaScript programming language. The application allows the user to take a photo directly or load the image from a computer to recognize it. The application recognizes five different surgical instruments: curved surgical scissors, flat surgical scissors, scalpel, surgical tweezers, and forceps.

Ovaj rad objašnjava osnovne koncepte potrebne da se načelno shvate konvolucijske neuronske mreže počevši od osnove umjetne inteligencije pa sve do konkretne implementacije. Rad kao motivaciju obrađuje detekciju registracijskih tablica vozila te se usredotočuje na rješavanje dotičnog problema. Primjenjujemo konvolucijske neuronske mreže u tom naumu, specifično YOLO (You Only Look Once) arhitekturu te službenu darknet implementaciju. Uz YOLO arhitekturu koja je detaljno objašnjena, navedene i uspoređene su ostale poznate arhitekture konvolucijskih neuronskih mreža. Konačno, izgrađen je generaliziran model za registracijske tablice koji nije specifičan nijednoj državi niti obliku tablice. This paper attempts to explain basic concepts required to understand essentials of convolutional neural networks starting from the basics of artificial intelligence up to actual implemenation. As motivation, this paper explains license plate detection and focuses on solving said problem. We use convolutional neural networks in attempt to solve the issue, specifically YOLO (You Only Look Once) arhitecture and its official darknet implementation. In addition to the YOLO arrhitecture which is thoroughly explained, other frequently used convolutional neural network architectures are mentioned and compared. Finally, a generalized model for detecting license plates is constructed which is not specific to any country nor shape.

Cilj rada je istražiti primjenjivost umjetnih neuronskih mreža u cilju implementacije funkcija previđanja u kontekstu kompresije slika bez gubitaka. Opisane su osnove kompresije slika bez gubitaka i umjetnih neuronskih mreža. Opisane su i argumentirane metode korištene u radu. Posebno je obrađen pristup optimizaciji arhitekture mreže i hiperparametara koji je temeljen na tehnici Parzenovih stabala. Opisana je implementacija četiri različite funkcije predviđanja temeljene na umjetnim neuronskim mrežama, te su analizirani rezultati. The goal of this paper is to determine the feasibility of using artificial neural networks as prediction functions in the context of lossless image compression. The paper explains the basics of lossless image compression and artificial neural networks. Used methods are described and analyzed, with special focus on optimization of network architecture and hyperparameters using Tree-structured Parzen estimators. The paper covers implementation of four different prediction functions based on artificial neural networks, as well as the analysis of the results.

Duboke umjetne neuronske mreže pokazale su se kao dobro sredstvo za kvalitetnu strojnu tvorbu govora. Rad daje općenit uvod u umjetne neuronske mreže i načine strojne tvorbe govora. U fokusu je model WaveGlow, koji zajedno s modelom Tacotron2 daje novu perspektivu na kvalitetno sintetiziranje govora iz teksta. Model je primijenjen na setu podataka za hrvatski jezik i prikazani su njegovi rezultati. Deep neural networks showed great results in TTS (Text To Speech) synthesis. This paper gives general introduction to articifial neural networks and speech synthesis. The focus is on model WaveGlow, which together with model Tacotron2 gives a new perspective on quality text to speech synthesis. The model is trained on a croatian language dataset and its results are presented in the paper.

U ovom završnom radu razmatran je prijedlog vizijskog sustava za prepoznavanje nota iz notnog zapisa baziran na računalnom vidu te primjeni umjetnih neuronskih mreža. Rad započinje osnovama glazbene teorije radi boljeg razumijevanja ovog rada, nakon koje slijedi konceptualna razrada problema. Opisan je i razrađen princip rada modela računalnog vida te obrade slika kreiran za potrebe ovog rada. Objašnjen je princip rada osnovnih umjetnih neuronskih mreža te je detaljnije razrađena struktura te primjena konvolucijskih neuronskih mreža. This final thesis considers the use of artificial neural networks in visual recognition systems. The paper begins with the basics of music theory for a better understanding of this work, followed by a conceptual elaboration of the problem. The principle of computer vision and image processing model created for the purpose of this paper is described and elaborated. The principle of operation of basic artificial neural networks is explained and the structure and application of convolutional neural networks are elaborated in more detail.

U radu je opisana primjena dubokog učenja za pretragu Python izvornih kodova na osnovu korisničkog upita. Objašnjeni su osnovni principi dubokog učenja i predložena je arhitektura za rješavanje zadanog problema. Ostvarena je implementacija s kojom je demonstrirana funkcionalnost dubokog modela. Na kraju rada predložene su mjere za analizu implementacije. Analizom mjera pokazana je uspješnost izvođenja implementiranog modela dubokog učenja. This paper discribes the application of deep learning to demonstrate search engine for Python source codes based on user query. Basic principles of deep learning have been explained and problem solving architecture is proposed. An implementation of search engine was implemented demonstrating the functionality of the deep model. At the end of this paper, measures for analysis of this implementation were presented. The analysis of the measures showed the success performace of implemented deep learning model.

Ovaj rad se bavi izradom ekonomskih predviđanja cijene na temelju podataka iz prošlosti putem umjetnih neuronskih mreža. Pojašnjeni su osnovni pojmovi teorije iz područja strojnog učenja, najviše umjetnih neuronskih mreža i ekonomije. Rad obuhvaća izradu vlastite umjetne neuronske mreže i testiranja s pravim povijesnim podacima s burze uz dokumentiran proces implementacije., The paper deals with the description of the technologies that were needed to implement simple financial forecasting system that uses artificial neural networks. Theories and basic terms needed for understanding this paper from fields of economics, machine learning, mostly neural networks are explained. The paper involves projecting own artificial neural network model that uses real historical dana for making financial forecast. Process of implementing artificial neural network is documented and explained.

Vitamin C (askorbinska kiselina) je u vodi topljivi vitamin važan za održavanje bioloških funkcija u tijelu. Ljudi ga ne mogu sami sintetizirati pa ga je potrebno unositi kao dodatak prehrani. Najrašireniji oblik oralnih dodataka prehrani jest tableta zbog svoje lake primjene i brze apsorpcije. Cilj ovog rada bio je procijeniti sposobnost primjene blisko infracrvene spektroskopije (NIR), kemometrije i umjetnih neuronskih mreža (ANN) za predviđanje ukupnih aktivnih tvari u tabletiranim dodacima prehrani (vitamin C). Uzorci su pripremljeni s različitim koncentracijama laktoze i vitamina C. U radu su se izravnim metodama pratile promjene električne vodljivosti, ukupno otopljenih čvrstih tvari i pH u ovisnosti o koncentraciji laktoze i vitamina C u uzorcima otopina i tabletama. Koncentracija vitamina C određena je spektrofotometrijski pomoću dikromata i permanganata. Osim izravnog mjerenja, korištena je i blisko infracrvena spektroskopija kako bi se utvrdile sličnosti i razlike u uzorcima. Za obradu podataka dobivenih NIR spektroskopijom korištena je analiza glavnih komponenata (PCA analiza) koja je pokazala dobro razdvajanje na temelju koncentracije laktoze i vitamina C. Kako bi se ustanovila povezanost između podataka i koncentracije kreirani su PLS regresijski modeli koji su ustanovili linearnu ovisnost između apsorbancije i koncentracije vitamina C. Primjena ANN također je pokazala visoku pouzdanost i mogućnost primjene za predviđanje točne koncentracije vitamina C u tabletama. Vitamin C (ascorbic acid) is a water-soluble vitamin important for maintaining biological functions in the body. People cannot synthesize vitamin C, so it needs to be intake as a dietary supplement. The most widespread form of oral nutritional supplements is the tablet because of its easy use and rapid absorption. The aim of this paper was to evaluate the applicability of near infrared spectroscopy (NIR), chemometrics and artificial neural networks (ANN) to predict the content of active substances in dietary supplements (vitamin C). Samples were prepared with different concentrations of lactose and vitamin C. In the paper, changes in electrical conductivity, total dissolved solids and pH were monitored by direct methods, depending on the concentration of lactose and vitamin C in the solutions and tablets. Vitamin C concentration was also directly determined spectrophotometrically using dichromate and permanganate. In addition to direct measurement, near-infrared spectroscopy was also used to determine the similarities and differences in the samples. For the processing of the data obtained by NIR spectroscopy, principal component analysis (PCA analysis) was used, which showed good separation based on lactose and vitamin C concentration. To establish a relationship between data and concentration, PLS regression models were developed which established linear dependence between absorbance and concentration of vitamin C. The ANN also proved to be highly precise and reliable for prediction of vitamin C concentrations in tablets.

U radu je opisan razvoj umjetne inteligencije na primjerima računalnih igara. Proučeno je korištenje genetskih algoritama za razvijanje težina i topologije umjetnih neuronskih mreža. Naučene mreže koriste se za stvaranje strategije igranja jednostavnih videoigara. Ostvaren je algoritam u okviru rada koji je onda ispitan na jednostavnim problemima. Analizirana je djelotvornost sustava i razmotreni su sljedeći mogući koraci za unaprjeđenje algoritma. In this paper we describe the influence of computer games on the field of artificial intelligence. We cover application of genetic algorithms on optimizing weights and developing structure of artificial neural networks. Trained networks are used to define playing strategies for simple videogames. We implemented the algorithm within the scope of this paper which is then tested on simple problems. We then analyzed it's efficiency and proposed some further improvements.

Ovaj rad bavi se detekcijom umora vozača uporabom dubokih umjetnih neuronskih mreža. Opisana je motivacija za rad na sprječavanju posljedica od umora vozača i pregled metoda koje su trenutno korištene u postojećim vozilima. Objašnjeno je što je to duboko učenje i osnove konvolucijskih neuronskih mreža. Opisana je metoda koja koristi duboko učenje i treptanje kao indikator umora. Prikazana je i implementacija programa temeljenog na toj metodi i osnovni alati koji su bili nužni da se ona ostvari. Na kraju su prikazani rezultati primjene implementacije na skupu podataka i predložena su poboljšanja i naputci za daljnji rad. This paper discusses drowsiness detection in drivers using deep artificial neural networks. It presents the motivation for preventing drowsy driving and an overview of methods being currently used in existing vehicles. In this work, an explanation is given on what deep learning is and the basics of convolutional neural networks. It then describes a method which uses deep learning and blinking as a drowsiness indicator. It also shows a software implementation for the described method and tools which were necessary for that implementation. At the end of the paper results of implementation are shown on a dataset and suggestions are given for further improvements.

Tema ovog rada je prilagodba ponašanja inteligentnih agenata u zadanom okruženju. Opisano je što su to agenti, njihove karakteristike, te načini djelovanja. Uz to je također definirano više vrsta agenata, te način na koji oni procesuiraju informacije, te djeluju na okolinu. Osim samog koncepta neuronskih mreža, opisan je i način na koji agenti mogu učiti informacije, što će im u konačnici uvjetovati kvalitetu ponašanja, odnosno zadovoljavanje definiranih uvjeta. Budući da agenti ne moraju nužno uspješno naučiti zadovoljavajuće ponašanje, potreban je proces uvježbavanje. Za to su korišteni evolucijski algoritmi, koju vuku inspiraciju iz prirode, te se pomoću njih odabiru samo oni agenti koji zadovoljavaju zadane uvjete, te se prema njima stvaraju novi, njima slični. Cilj projekta je napraviti program koji implementira opisano, te prikazati kako se agenti ponašaju, odnosno njihovu sposobnost učenja., The topic of this paper is how to adjust the behavior of intelligent agents in the given environment. First, multi-agent systems are described, because they implement the described system. It describes what the agents are, their features and ways of action. Furthermore, more types of agents are defined, the ways they process information and act in the environment. Their function is defined through their ability to process information, for which purpose in this paper neural networks are used. Apart from the concept of neural networks, it is also introduced and explained the way in which agents can learn information, which will ultimately be the condition for the quality of behavior or the fulfillment of defined parameters. Since agents do not necessarily need to learn successful behavior, it is necessary to introduce something that improves their behavior. For this purpose, evolutionary algorithms are used, inspired by nature, and only those agents that meet the given conditions are chosen to create new ones, similar to them. The goal of the project is to create a program that implements the described, and show how the agents behave, that is, their ability to learn is observed.

U radu je obrađena umjetna inteligencija i njezina primjena danas. U svijetu velikih podataka i poplave informacija sustavi koji primjenjuju umjetnu inteligenciju postaju sve značajniji. Krajem dvadesetog stoljeća, usvajanjem bottom-up pristupa u razvoju umjetne inteligencije započinju novi pravci istraživanja. Obilježeni su značajnim porastom ulaganja u razvoj inteligentnih sustava, koje traje i danas. U radu su opisane različite metode kojima se ostvaruje umjetna inteligencija. One uključuju ekspertne sustave, strojno učenje, umjetne neuronske mreže, evolucijsko računalstvo te koncept inteligentnog agenta. Prikazani su primjeri uporabe sustava temeljenih na umjetnoj inteligenciji poput virtualnog asistenta, sustava za preporuke, autonomnih vozila. Posebno je istaknut Google, najveći komercijalni sustav umjetne inteligencije ikada izgrađen. Opisana su nova istraživanja umjetne inteligencije primjenom računalne arhitekture inspirirane radom mozga. Spomenuta su očekivanja primjene umjetne inteligencije u personaliziranim pametnim domovima povezivanjem predmeta u “internet stvari”. Na kraju se navodi kako konačni cilj, stvaranje strojeva koji misle, još nije postignut. Nastojanja da se taj cilj ostvari ne izazivaju samo podsmijeh, već zabrinutost javnosti za budućnost čovječanstva. Poticaj za brigu je mogućnost razvoja generalne umjetne inteligencije u desetljećima koja slijede., This paper deals with artificial intelligence and its application today. In the world of big data and the flood of information systems using artificial intelligence are becoming increasingly important. At the end of the twentieth century a bottom-up approach to the development of artificial intelligence started new directions in research. They were marked by a significant increase in investment in the development of intelligent systems, which continues today. The paper describes various methods achieved in field of artificial intelligence. Those methods include expert systems, machine learning, artificial neural networks, evolutionary computing, and the concept of an intelligent agent. Given are examples of systems based on artificial intelligence such as a virtual assistant, customers recommendation system, self-driving car. Particularly emphasized is Google, the largest commercial system of artificial intelligence ever built. Described are current research projects in artificial intelligence which are using computer architecture inspired by the brain. There is also expectation of wide usage of artificial intelligence in the personalized smart homes by linking objects in the "internet of things". In the end it is stated that the final goal, the creation of machines that thin, has not yet been reached. Yet today, these efforts do not cause laughs, but the concern about aftermath of humanity if we reach that goal. Boost for concern is the possibility of the development of general artificial intelligence in the decades that follow.

Prepoznavanje vrste i lokacije kvara nužan je preduvjet za povratak elektroenergetskog sustava (EES-a) u normalno stanje, kao i sprečavanje mogućeg oštećenja elemenata elektroenergetske mreže. U doktorskom radu analizirane su mogućnosti određivanja lokacije i vrste kvara u prijenosnoj elektroenergetskoj mreži upotrebom umjetnih neuronskih mreža. Rad se posebno fokusira na prepoznavanje visokoomskih kvarova (eng. High Impedance Fault ili HIF) u prijenosnoj elektroenergetskoj mreži, gdje su mjerene električne veličine bliske nazivnim vrijednostima koje su prisutne u normalom pogonu. Za rješavanje navedenog problema odabran je pristup upotrebom umjetnih neuronskih mreža (eng. Artificial Neural Networks ili ANN). U radu je korišten trofazni model elektroenergetske mreže, a kao temelj raspoznavanja stanja kvara od normalnog pogona u slučaju pogona nelinearnih trošila, predložena je metoda koja objedinjuje određivanje frekvencijskog spektra valnog oblika električnih veličina i umjetnih neuronskih mreža. Da bi se to postiglo korištena je brza Fourierova transformacija (FFT) za dobivanje udjela pojedinih harmonika proračunatih električnih veličina koje su zatim korištene kao ulazni podaci za učenje i testiranje umjetne neuronske mreže. Na isti način korištenjem frekvencijskog spektra omogućeno je i prepoznavanje lokacije i vrste kratkog spoja na nadzemnom vodu. Identifying the types and locations of faults is a necessary precondition for the restoration of the power system in a normal state, and preventing possible damage to the electricity network elements. The doctoral thesis analysed the possibility of determining the location and type of fault in the power system transmission network by using artificial neural networks. The doctoral thesis is focused on recognition of high impedance faults (HIFs) in the power system transmission network, where the measured electrical parameters have values close to the nominal values that are present in the normal operation. To solve the above problems approach using artificial neural networks has been chosen. The doctoral thesis used three-phase models of the electric power network, as a basis of fault recognition apart from normal operation condition in the case of non-linear loads, a method proposed combines the determination of electrical quantities waveform frequency spectrum and artificial neural networks. In order to achieve this fast Fourier transform FFT was used to obtain a harmonics spectrum of calculated electrical quantities that are then used as input data for learning and testing artificial neural networks. In the same way the use of the frequency spectrum allowed the identification of the location and type of short circuit on the transmission line. The overhead line is an integral part of the power system and its protection is important to ensure system stability and to reduce the damage to equipment that may occur due to short circuits on the overhead lines. Short-circuiting leads to mechanical and thermal stresses that are potentially harmful to high-voltage equipment. Overhead protection includes three main tasks: detecting, classifying and defining a location on overhead line. Fast detection of overhead line failure enables quick isolation of faulted overhead line from the system to protect it from further damaging effects. Properly determining the fault location is necessary to allow reparation of faulted overhead line in the shortest possible time and return it to normal operation, increasing availability and power supply. Accurate determination of the fault location, the amount of time required by the maintenance crew in search of a fault can be reduced to a minimum. However, fault identification is not always an easy task. If there is a fault, it is necessary to isolate it as quickly as possible to preserve the stability of the rest of the system. Overhead protection relays normally use voltages and currents as input signals to detect, classify and locate a fault in overhead lines. In the event of a fault, the relay will send a switch-off signal to switch off the overhead line, so the rest of the system can continue working in normal or near normal conditions. Different types of algorithms and methods for finding faults on overhead lines have been developed and proposed over the years and can be classified into the following groups: a) Calculation of the phase current and voltage phases from which the fault location can be determined by line impedance, b) Using differential overhead line equations, c) A method of traveling waves using data from one or both ends of the overhead line. Some protective relays may have difficulties detecting a fault due to the high impedance and the DC component of the short circuit current. The approach used by travelling waves has problems with the detection of faults that are very close to the substation and if the angle at the time of the fault falls close to zero (or equal to zero). One of the tools recently introduced in the power system is the Artificial Neural Network (ANN). ANN is a powerful pattern recognition, classification and generalization tool that is useful for energy system analysis because ANN can learn with the off-line data. The ANN also has excellent features such as resistance to interference, robustness and error insensitivity. This doctoral thesis describes the application of the ANN to detect faults in the transmission system network of the power system, the application of which is described in a large number of papers. The fault itself depends on the impedance of the surrounding network and the impedance of the fault itself, so protection relays are most often set for topology with the highest value of fault current. Advantages of ANN such as the ability of continuous learning, associative memory, nonlinear mapping and parallelism allow its use in a wide range of engineering research. HIF is generated when the phase line makes unwanted contact with low conductivity objects such as road surface, pavement, lawn or tree branches, limiting the fault current to very low values. These faults can not be easily detected by conventional overcurrent protection devices and often cause an electrical arc fault if there is no available return path for the current, leading to presence of high frequency components in fault current. Different ways of detecting HIF were previously suggested by other researchers. These methods based on algorithms used can be divided into time domain and frequency domain. In the time domain, the relay method with the ground ratio, proportional relay algorithm, smart relay, flicker failure and semi-peripheral asymmetry were suggested. In the frequency domain several articles are published based on Kalman's filtering, fracture theory and neural networks. Recently the proposed wavelet transformation has been used to achieve better results. The techniques for fault locating can be classified into two categories: 1) method of using data from one end of the transmission line and 2) method of using data from both ends of the transmission line. It is well known that the techniques based on measurement at both ends require communication links and synchronized sampling equipment. In one paper it was proposed to protect the system of discrimination against HIF and the normal operating system events in a distribution system based on wavelet packet transformation and ANN. However, much research has not been carried out on the differences between HIFs and nonlinear load in high-voltage transmission networks. Doctoral thesis consists of nine chapters. Following the introduction of the research objectives and the literature review, in the second chapter, a transmission network model was developed, using three-phase models, used to perform simulations necessary for further calculations. The third chapter briefly described types of short circuit in power system using three-phase models of network elements. The fourth chapter describes the working principle of artificial neural networks and their possible application in power system analyses. The fifth chapter describes the Fourier transformation process and its basic features, and its application in the analysis of periodic functions. The sixth chapter describes a method that uses ANN for the fault type of detection and its verification. The seventh chapter describes a method that, using ANN, calculates the location of a fault on overhead line and its verification. The eighth chapter describes a method that performs the recognition of HIF from a normal nonlinear load and its verification through ANN. Concluding considerations are presented in Chapter 9. Short circuits in the transmission network pose a danger both to the equipment and to the staff that can be found near the fault. High fault current cause dynamic forces that can cause mechanical damage to equipment and thermal damage caused by Jule's heat. A special problem is the HIF, where in cases when fault currents are not high enough to make the overcurrent protection work properly. Such failures are most often caused by a low conductive element such as a branch, a ground conductor drop or an electric arc. Unlike low-impedance faults, HIFs pose a major risk to electrical power equipment and are accompanied by an electric arc at the site of failure. The doctoral thesis proposed a methodology for detecting the type and location of low-impedance failt and HIF on the overhead line and the distinction of HIF from a normal non-linear load operation. The proposed methodology uses the ANN to solve the aforementioned problems by using voltage and current waveforms at the end of the overhead line for input quantities and switching from the time domain to the frequency domain. Calculation of voltage and current harmonics make input sets for ANN to provide satisfactory results. The tested network is three-phase modelled according to the actual part of the transmission system. The calculations were carried out in such a way as to take into account variable load flow, fault locations, fault inception and impedance at the fault location. In the reviewed literature, cases with less variable parameters were processed. The first part of the method deals with identifying the type of fault on overhead line. All types of short circuit are made except for a three-phase short circuit without a simultaneous earth connection since the possibility of such failure is negligible. Since detection of the type of fault is a classification problem, ANN has been used to identify patterns. ANN provided the best results in the case of network with 100 neurons in the hidden layer, and the lowest accuracy of the phase B to phase C with ground fault classification, whose classification accuracy was 93.52%, which was expected to overlap part of HIF's performance with the diode rectifier locomotive drive in normal drive condition. Verification of this method was performed on the example of a single-phase short circuit in the real network whose voltage and current parameters are translated from the actual network to the model. ANN successfully recognised the entry as a phase B with ground fault. In this section, the first scientific contribution of the ANN model to detect a failure has been realised. The second part of the method deals with identifying the location of the failure on overhead line. To solve this problem, ANN is used that uses feedforward network with backpropagation algorithm and the least squares error function. The best results were achieved by ANN with 10 neurons in the hidden layer where the average error was 1.13%, while the maximum error value was 2.48% with respect to the total length of the overhead line. Verification was also carried out on a real short circuit whose parameters were translated to the model while at the same time knowing the location of the failure detected by the fault locator in the relay. For ANN with 10 neurons in the hidden layer, the error was 1.23%, which is a satisfactory level for team teams on the field who must repair the failure in the shortest possible time. The third part of the method deals with the discovery of HIF from a nonlinear load. In this paper, nonlinear load was simulated by electric railway substation. The diode and thyristor locomotives are also modelled in detail. This is achieved by increasing the security of determining the type of fault, since the HIF is undetected by overcurrent protection and the location of the fault places it close to electric railway substation. And in this case, the ANN was used for pattern recognition. In this case, ANN had a 100% success in not only detecting electric railway substation load from fault but also discerning non-linear loads due to its specificity in waveforms. The verification was carried out on the example of the waveform of the diode locomotive load in the real network and its waveform was transmitted to the model. ANN correctly classified this form as a diode locomotive load. This part of the method has accomplished another scientific contribution that defines the differentiation of the high-impedance fault from the normal load state of the transmission network with a large share of unbalanced and non-linear loads. Throughout all three parts of the method in the final sub-headings, as a third scientific contribution, a model was evaluated to determine the type and location of the fault based on verified computational models and measurements. A set of proposed methods ensures increased accuracy and reliability of existing protective devices as well as additional support in the analysis after the protection process. The accuracy of determining the location of the overhead line fault can be considered satisfactory, although the amount of error depends largely on the amount of different topological network states. By selecting just realistic possible overhead line conditions the fault location accuracy would be further enhanced.

Selekcijski intervju najčešće je korištena metoda procjene ličnosti kandidata za različite poslove, usprkos tome što kandidati često tijekom njega iskrivljavaju svoje odgovore. Dodatni problem predstavlja to što stručnjaci u području odabira osoblja nisu vješti u otkrivanju lažiranja. Neki su istraživači predložili korištenje većeg broja ponašajnih pokazatelja kako bi se odgovorilo na ovaj izazov, iako su istraživanja u ovom području još dosta rijetka. Cilj ovog rada bio je istražiti mogućnost otkrivanja lažiranja odgovora u selekcijskom intervjuu putem multimodalnog pristupa, odnosno putem ponašajnih indikatora iz četiri šire kategorije: paraverbalnog, verbalnog i neverbalnog ponašanja te izraza lica. Dodatno, istražili smo i usporedili u kojoj se mjeri dijelovi intervjua u kojima su se sudionici predstavljali iskreno, odnosno kao idealan kandidat, mogu uspješno klasificirati logističkom regresijom i umjetnim neuronskim mrežama. U ovom istraživanju sudjelovao je prigodan uzorak od 102 studenta, odnosno osobe koje su diplomirale unatrag godinu dana. Oni su sudjelovali u zamišljenom selekcijskom intervjuu za poziciju voditelja pozivnog centra, pri čemu su snimani kamerom. Intervju se sastojao od dva bloka: u jednom zadatak sudionika bio je predstaviti se iskreno, a u drugom kao idealni kandidat za radno mjesto. Redoslijed blokova i pitanja bio je uravnotežen. Kako bismo dodatno motivirali sudionike na što uvjerljivije lažiranje njihovih odgovora, naveli smo da će 10 % najuvjerljivijih dobiti nagradni bon od 500 kn u jednoj vele-drogeriji. Na kraju intervjua prikupili smo podatke o nekim kontrolnim varijablama te varijablama vezanim uz učinkovitost eksperimentalne manipulacije. Parovi nezavisnih opažača kodirali su pojedine kategorije ponašanja na videosnimkama. Dodatno, za analizu izraza lica koristili smo i aplikaciju OpenFace. Provjera eksperimentalne manipulacije pokazala je da je ona bila uspješna, ali ne u potpunosti. Što se tiče prosječnih razlika u ponašajnim pokazateljima između dvije situacije odgovaranja, dobili smo značajne multivarijatne efekte za paraverbalno i verbalno ponašanje te za izraze lica, neovisno o tome je li se radilo o ljudskim opažačima ili o OpenFaceu. Logistička regresija i neuronske mreže pokazale su se podjednako uspješnima u klasifikaciji blokova intervjua, pri čemu je stopa točnih klasifikacija iznosila 65% za oba modela. Introduction The selection interview is the most frequently used method of personnel assessment for various jobs, despite its susceptibility to response distortion. Unfortunately, existing research indicates that experts in personnel selection are no better than chance in their detection of response distortion, irrespective of their interviewing experience. As one of the potential solutions for this challenge, some authors have proposed investigating the behavioural indicators of faking during the interview. According to various general theories and models of deception (i.e., Buller & Burgoon, 1996; DePaulo et al, 2003; Ekman, 2002; Walczyk et al, 2014), differences in behavioural cues should mirror the differences in psychological processes between honesty and deception. More precisely, deception should, at least under some circumstances, be more cognitively taxing, accompanied by more rigid behaviour and by less authentic facial expressions. Although there is an abundance of papers whose authors investigated the possibility of deception detection via behavioral cues in the context of forensic and cognitive psychology, research on this topic is still very scarce when it comes to the selection interview. Therefore, the aim of this study was to explore the possibility of detection of faking in the selection interview through a multimodal approach, based on paraverbal, verbal and nonverbal cues, and facial expressions. Additionally, it was examined whether it is possible to detect response distortion with two different algorithms: one based on logistic regression and the other on artificial neural networks. Methodology: In total, a convenience sample of 102 students/recent graduates (71% female, Mage=24.2, SDage=3.08) participated in a video-recorded mock structured selection interview for the position of call centre manager. There was a total of 16 interview questions that measured the dimensions of extraversion and honesty/humility. The interview consisted of two blocks: in one block participants had to answer honestly, while in another block their task was to present themselves as an ideal candidate for the position. The order of interview questions and blocks was balanced across participants and every question appeared just in one block. As an incentive to distort their answers effectively, participants were told that 10% of those with the most convincing self-presentations in faking condition would receive an award coupon of approximately 70 euros. At the end of the interview, some control variables and variables related to the experimental manipulation check were measured. Two independent coders analysed cues from every behavioural category. Additionally, facial expressions were analysed with the OpenFace application. Results and discussion: The manipulation check indicated that participants answered less honestly, used more faking strategies, experienced more cognitive load and fear of detection, and controlled their behaviour more when presenting themselves as ideal candidates. They achieved higher scores on extraversion and honesty/humility while faking as well. Interestingly, participants reported higher levels of motivation to present themselves convincingly in the honest condition. Regarding the differences in behavioural cues between honest vs. ideal candidate conditions, significant multivariate effects were obtained for paraverbal, verbal and facial expressions (both for human coders and OpenFace) categories. In the context of the paraverbal behaviour, participants demonstrated more filled pauses and fewer speech errors while presenting themselves as ideal candidates. Regarding the verbal behavior, faking was accompanied by more first-person pronouns and fewer terms indicating uncertainty. In the facial expressions’ category, both the human coders and the OpenFace data indicate that, overall, the face was more rigid and “frozen” during faking. Additionally, an interesting pattern emerged: during the faking condition action units implicated in eyebrow raising (AU1 and AU2) were more active, while during the honest responding, there was more activation in action units related to smiling in general (AU10, AU12, and AU14), and especially to its honest, authentic variant (AU6). There was no significant multivariate effect for the nonverbal behavior in differentiating honest and faking conditions. The pattern of results described above is, at least to some degree, in line with some of the general models and theories of deception. A higher frequency of filled pauses and less activity in the face during the faking condition could indeed indicate that it was more cognitively taxing for participants to present themselves as ideal candidates, compared to responding honestly. Alternatively, this pattern of results could also have emerged as an unplanned consequence of overcontrolling one’s behaviour during the faking, in accordance with hypotheses derived from a self-presentational perspective (DePaulo et al., 2003) and interpersonal deception theory (Buller & Burgoon, 1996). On the other hand, as evidenced by their results in the experimental manipulation check, participants didn’t experience intense emotions during the interview, limiting their utility as an exploratory mechanism beyond behavioral differences between faking and honest responding. As for the classification of the interview conditions (honest vs. fake responding), the same levels of performance were achieved both with logistic regression and artificial neural networks, with 65% sensitivity, specificity, and overall accuracy rates on the test data. These rates are only slightly lower than the ones achieved on the training data, indicating a good generalizability of our models’ performances to a new, previously unseen dataset. Additionally, these results are comparable to some previous findings in the field (e.g., Culbertson et al., 2016). Conclusion: It was demonstrated in this study that response distortion in the selection interview could be detected by its unique behavioural signature, but not all modalities and their cues were equally useful. Although statistically large effect sizes were obtained for various measures of the experimental manipulation, in an absolute and practical sense the psychological differences between conditions could have been more pronounced. This could have at least partly contributed to the absence of a larger number of significant behavioural cues. Regarding the classification of faking/honest responding, when investigating different algorithms, future researchers should remember that more powerful models are not necessarily more accurate.

Ovaj diplomski rad se bavi primjenom umjetne inteligencije za rješavanje problema u području tehnološke pripreme proizvodnje. U radu su objašnjene metode umjetne inteligencije, navedene su prednosti i nedostatci umjetnih neuronskih mreža te je provedeno i predstavljeno istraživanje primjenom softvera NeuroIntelligence. Cilj istraživanja je na temelju ulaznih parametara: posmak po zubu, posmak radnog stola, broj okretaja, dubina rezanja primjenom umjetnih neuronskih mreža procijeniti karakteristike obrađene površine kod obrade odvajanjem čestica. Naime, na ovaj se način unaprijed može procijeniti ispunjavaju li izabrani parametri obrade unaprijed zadanu kvalitetu obrađivane površine., This master’s thesis deals with the application of artificial intelligence to solve problems in the field of technological preparation of production. In this paper methods of artificial intelligence are explained, advantages and disadvantages of artificial neural networks are mentioned and research in NeuroIntelligence software is performed and presented. The aim of the research is to evaluate the characteristics of the processed surface by mechanical machining process based on the input parameters: feed per tooth, bench feed, spindle speed and depth of cut using artificial neural networks. Namely, in this way, it is possible to assess in advance whether the selected processing parameters meet the predefined quality of the processed surface.

Duboke neuronske mreže su standardni alat koji pomaže u rješavanju problema kao što su detekcija objekata, raspoznavanje oblika i klasificiranje podataka. Potrebna je određena količina znanja kako bi se duboka neuronska mreže istrenirala za željenu radnju. Precizno označeni podatkovni skupovi uvelike pomažu pri učenju neuronske mreže. Što više podataka za učenje proslijedimo neuronskoj mreži to će ona bolje raditi svoj zadatak za koji je istrenirana. U ovom radu opisane su osnove umjetne inteligencije i strojnog učenja. Istražene su vrste neuronskih mreža i razlika između dubokih i plitkih neuronskih mreža. U svrhu istraživanja razlike učinkovitosti između plitkih i dubokih konvolucijskih neuronskih mreža odabrana je mala konvolucijska neuronska mreža koja se sastoji od dva konvolucijska sloja te tri potpuno povezana sloja i VGG-11 konvolucijska neuronska mreža koja se sastoji od osam konvolucijskih i tri potpuno povezana sloja. Deep neural networks are a standard tool to help solve problems such as object detection, shape recognition, and data classification. A certain amount of knowledge is required to train a deep neural network for a desired action. Precisely labeled datasets greatly aid neural network learning. The more training data we feed the neural network, the better it will perform at the task for which it was researched. This paper describes the basics of artificial intelligence and machine learning. The types of neural networks and the difference between deep and shallow neural networks are explored. In order to investigate the difference in efficiency between shallow and deep convolutional neural networks, a small convolutional neural network consisting of two convolutional layers and three fully connected layers and a VGG-11 convolutional neural network consisting of eight convolutional and three fully connected layers were selected.

Računalima je dostižan računalni vid uz pomoć umjetne inteligencije kakav je svojstven ljudima, da prepoznaju i procesiraju razne objekte sa slika i videa u dvodimenzionalnom i trodimenzionalnom sustavu. Razvojem novih algoritama i većom dostupnošću računalne snage, te pristupa internetu, dnevno se generiraju velike količine podataka. Iz tih razloga dolazi do iznimnog napretka računalnog vida što omogućuje rješavanje sve više novih i raznolikih zadataka koji uključuju prepoznavanje. Ovim radom su analizirani model i segmentacijska svojstva umjetne neuronske mreže za segmentaciju osoba na slikama. Model je razvijen u Python programskom jeziku. Provedena evaluacija je uključivala ljudske subjekte. Computers achieve computer vision with the help of artificial intelligence, with characteristics of human vision, to recognize and process various objects from images and videos in two- and three-dimensions. With the development of new algorithms and the availability of greater computing power, as well as access to the Internet, large amounts of data are generated daily. For these reasons, there is an extraordinary advance in computer vision that allows solving new and diverse problems that involve recognition. This paper analyzes the model and segmentation properties of an artificial neural network for the segmentation of people in images. The model was developed in the Python programming language. The conducted evaluation involves human subjects.

Strojno učenje se može koristiti u svrhu brže opće dijagnostike u medicini. U ovom radu strojno učenje korišteno je u analizi slika tumora dojke. Primjenom odgovarajućih algoritama na skup podataka moguće je stvoriti model koji će moći brže i preciznije dati dijagnozu. Slike tkiva dobivenih pod mikroskopom klasificirane su u odgovarajuće klase malignih ili benignih tumora. Za potrebe rada stvorena su dva modela različite složenosti, kao i hibridni model koji je kombinacija složenijeg s referentnim modelom. Modeli su stvoreni i trenirani uz pomoć Pythonove knjižnice TensorFlow. Dobiveni rezultati ukazuju na potencijal složenijeg modela i odlične rezultate hibridnog modela, iako su svi modeli dobar pokušaj rješenja zadanog problema. Machine learning can be used for the purpose of faster general diagnostics in medicine. In this paper, machine learning was used in the analysis of breast tumor images. By applying appropriate algorithms to a set of data, it is possible to create a model that will be able to make a diagnosis faster and more accurately. Microscopic images of tissues obtained under the microscope are classified into appropriate classes of malignant or benign tumors. For the needs of the work, two models of different complexity were created, as well as a hybrid model that is a combination of the more complex model with the referenced model. Models were created and trained with the help of Python's TensorFlow library. The obtained results indicate the potential of a more complex model and the excellent results of the hybrid model, although all models can be considered as a good attempt to solve a given problem.

Rad se bavi metodama objašnjivosti koje se koriste na nekoliko često korištenih vrsta umjetnih neuronskih mreža.Na studijskom primjeru procjene starosti školjaka se prikazuje rad nekoliko agnostičnih metoda i onih specifičnih za umjetne neuronske mreže.Razmatraju se i specifični postupci za konvolucijske i povratne umjetne neuronske mreže te neke od vrsta podataka na kojima su one često primijenjene. The paper deals with explanatory methods used on several commonly used types of artificial neural networks. On the case study of estimating the age of shellfish, it shows the work of several agnostic methods and those specific to artificial neural networks. Specific methods for convolutional and recurrent artificial neural networks are also described, as well as some of the types of data on which they are often applied.

U ovom radu dokazujemo da pomoću duboke neuronske mreže je moguće dobiti aproksimativno rješenje za običnu diferencijalnu jednadžbu. Definiramo funkciju greške koja se koristi za treniranje neuronske mreže te implementiramo propagaciju pomoću vektorskog algoritma. Obavljamo eksperimente na mreži s različitim parametrima i uspoređujemo rezultate s iterativnom metodom Runge–Kutta reda 4. In this paper we prove that deep neural networks can be used to approximate solution of ordinary differential equation. We defined custom loss function used to train neural network and implement vectorized algorithm for forward propagation. Finally, we conduct experiments on the network and compare results with iterative method for differential approximation Runge–Kutta order four.

U sklopu ovog rada opisan je rad umjetnih neuronskih mreža s naglaskom na konvolucijske neuronske mreže. Također su obrađeni pojmovi iz teme semantičke segmentacije. U konačnici su implementirani i testirani modeli semantičke segmentacije biomedicinskih videozapisa i prikazani su rezultati. This paper describes artificial neural networks with an emphasis on convolutional neural networks. The main points of semantic segmentation are also described. In the end, 4 semantic segmentation models were implemented and evaluated and the results were shown.

Vidjeli smo povijesni razvoj i mnoge algoritme strojnog učenja, njihovu implementaciju te na koji način se sprječava njihova prenaučenost. Također je opisano na koji način su nastale umjetne neuronske mreže, kako ih projektirati i implementirati te na posljetku istrenirati i testirati. U sklopu rada implementirani su navedi algoritmi koji se mogu primijeniti na problem klasifikacije te je projektirana umjetna neuronska mreža. We have seen the historical development and many algorithms of machine learning, their implementation and how to prevent overfitting of those algorithms. It is also described how artificial neural networks were created, how to design and implement them, and ultimately train and test them. As part of the paper, the listed algorithms that can be applied to the classification problem were implemented and an artificial neural network was designed.

Zbog svoje široke primjene, gotovo da se danas može reći da je ovo vrijeme prijelaza na tehnologiju umjetnih neuronskih mreža. Stoga je cilj ovog rada predstaviti tu tehnologiju i njezinu primjenu u daljinskim istraživanjima. U tu svrhu korišten je digitalni ortofoto izrađen iz crno-bijelih aerofotosnimaka, približnog mjerila 1:20 000. Istraživanjem je obuhvaćena gospodarska jedinica “Jamaričko brdo”, šu- marije Lipovljani. Procjena sastojinskog obrasta provedena je primjenom višeslojnog percep trona, kao najkorištenijeg modela umjetnih neuronskih mreža u daljinskim istraživanjima.Također je korištena samoorganizirajuća neuronska mreža sa svrhom kontrole utvrđenih obrasta u Osnovi gospodarenja, pre ma njihovoj raspodjeli u tri kategorije (normalan, manji od normalnog, slab). Provedenim istraživanjem dobivena su dobra generalizacijska svojstva više slojnog perceptrona u procjeni obrasta, kao i da se samoorganizirajuća neu ronska mreža može primijeniti u kontroli i raspodjeli sastojinskih obrasta. Kako se u šumarstvu svakodnevno provodi velik broj različitih mjerenja, upravo umjetne neuronske mreže predstavljaju model temeljen na teoriji učenja, kojim bi se značajnije moglo unaprijediti korištenje ovako velikog broja podataka, koji su se do sada rješavali isključivo statističkim metodama i metodama operacijskih istraživanja., In the field of remote sensing the results of research undertaken with the purpose of determining quantitative and qualitative stand parameters showed the usefulness of artificial neural networks (Ardö et al. 1997, Skidmore et al. 1997, Wang & Dong 1997, Moisen & Frescino 2002, Ingram et al. 2005, Joshi et al. 2006, Kuplich 2006, Verbeke et al. 2006, Klobučar et al. 2008) as an alternative approach to classical statistical methods. This paper explores the possibility of estimating and distributing stand density using methods of artificial neural networks. These methods involve particular textural features of first and second order histograms on a digital ortophoto compiled from black and white aerial photographs at an approximate scale of 1:20,000. The paper is also aimed at collecting data with an acceptable accuracy, which will reduce material investments. Research encompassed the area of the MU “Jamaričko Brdo”, Lipovljani forest administration. Cyclic surveying was conducted in 2000. In order to determine textural features of first and second order histograms, a sample was cut out from a digital ortophoto for 80 stand scenes (compartments/subcompartments) in management classes of pedunculate oak, sessile oak and common beech of the fourth (the most common), fifth and sixth age class. A multi-layer perceptron was used to solve the problem of stand density estimation. A multi-layer perceptron is a neural network without feedback connections, where supervised learning is carried out with the error back propagation algorithm. An early stopping method was applied to improve generalization. The early stopping method is a statistical cross-validation method in which the available data are divided into three sets: training, validation and testing set. Of the overall dataset, 50 % (or 40 compartments/subcompartments) relates to the training set, whereas the two remaining datasets were divided equally: 25 % (20 compartments/subcompartments) relate to the validation set and 25 % (20 compartments/ subcompartments) to the testing set. There are numerous variations of error back propagation algorithms. As for the early stopping method, it is not advisable to use an algorithm which converges too rapidly (Xiangcheng et al. 2005, Demuth et al. 2006). Consequently, two algorithms were used: resilient back-propagation and scaled conjugate gradient algorithm. Prior to training the neural network itself, the data were preprocessed. In this sense, two operations were performed using MATLAB functions: normalization of input-output values and analysis of the main components of input values. Training encompassed a total of seven algorithm models with error back propagation with one or two hidden layers containing a different number of hidden neurons. Different activation functions were also applied in hidden and output layers. Self-organizing neural network was used to control densities according to their distribution into three categories (normal, less than normal, poor). To study the applicability of this neural network, 80 compartments/subcompartments were divided into two sets: training set and testing set, each consisting of 40 compartments/subcompartments. The data were preprocessed before the neural network was trained, just as was the case with the multilayer perceptron. Textural features of first order histograms (arithmetic means, standard deviation, smoothness, third moment, evenness and entropy) and second order histograms (absolute value of difference, inertia, covariance, entropy and energy) were used as input data for the neural network, whereas output density values were taken from the Management plan. Output values may also be represented as the number of trees, basal area or volume per hectare or as some other quantitative and qualitative stand values. Stand density was used as an output value for two reasons: a) poorer spectral features of the applied photographs, and b) the fact that, from the aspect of the forestry profession, the photographs were obtained in the unfavorable period (time of the year in which the ground is the least covered with vegetation). To test the difference in stand density values between the data from the Management plan and the optimal model of artificial neural network, the analysis of variance for repeated measurements was used. Research confirmed good generalization characteristics of a multilayer perceptron in density estimation, as well as the fact that a self-organizing neural network can be used to control and distribute stand densities. The applied procedure of density estimation achieves an acceptable accuracy and a high degree of automatism, which removes the subjective nature of classical remote sensing methods. This research confirmed the advantages and disadvantages of artificial neural networks. The advantages are as follows: it is not necessary to know data models, the networks can be used to analyze new conditions, and they tolerate imperfect data. The disadvantages are: the need to determine optimal architecture and the impossibility of estimation outside the scope of learning data values. However, despite their numerous advantages, artificial neural networks will not completely replace classical statistical methods. Instead, a dual approach and integration of these two techniques in decision making processes will be a very useful tool in forest resource management of the 21st century. They are currently broadly applied, so we could say that this is a time of transition to the technology of artificial neural networks. Consequently, forestry of the Republic of Croatia should make broader use of this new technology.

U radu se opisuje proces obrade radarskih slika i njihovo korištenje za prepoznavanje objekata. Koriste se konvolucijske neuronske mreže (YOLOv4 arhitektura) koje se treniraju na bazi podataka koja je snimljena pomoću radara. Cilj projekta je stvoriti model koji će dobro generalizirati i moću prepoznati objekte poput automobila, bicikala ili čovjeka iz radarskih slika. The paper describes the process of processing radar images and their use for object recognition. Convolutional neural networks (YOLOv4 architecture) are used, which are trained on a radar-recorded database. The goal of the project is to create a model that will generalize well and be able to recognize objects such as cars, bicycles, or people from radar images.

U ovom diplomskom radu opisani su napadi koji koriste sporedna svojstva kriptografskih uređaja te su izvršeni profilirajući napadi nad tragovima iz baze podataka ASCAD. Rezultati ́ su uspoređivani s onima iz rada [19]. Implementirani su modeli višeslojnog perceptrona i konvolucijske mreže u svrhu SCA napada. Osim klasifikatora implementiran je i generativni model, CGAN, u svrhu generiranja novih tragova. Pokazalo se da duboki modeli za učinkovit profilirajući napad trebaju veliku količinu tragova te su oba implementirana klasifikatora u tom slučaju dala jako dobre rezultate. Tragovi generirani CGAN mrežom u scenariju malog broja dostupnih tragova nisu doprinijeli kvaliteti napada. This master thesis describes side-channel attacks (SCA) and the conducted profiling at- tacks on traces from the ASCAD database. The results were compared with those from the paper [19]. Multilayer perceptron and convolutional neural network were implemented for the purpose of SCA attack. Apart from the classifiers, a generative model, CGAN, has been implemented for the purpose of generating new traces. It turned out that deep models need a large amount of traces for an effective profiling attack and both implemented classifiers gave very good results in that case. The traces generated by the CGAN network in the scenario where the attacker is in lack of traces did not contribute to the quality of the attack.

Podržano učenje jedan je od glavnih pristupa strojnom učenju i primjenjuje se za rješavanje mnogih zanimljivih problema. Jedna od osobito zanimljivih je primjena na automatsko igranje igara, što podrazumijeva izradu računalnih igrača koji samostalno znaju igrati neku igru. Postoje brojni algoritmi koji se primjenjuju u svrhu učenja takvih igrača, a među popularne i široko primjenjive tu se ubraja i algoritam poznat pod nazivom Q-učenje. U sklopu ovog rada izrađena je kopija poznate igre Space Invaders. Za treniranje igrača korištena je varijanta Q-učenja koja koristi umjetnu neuronsku mrežu kao aproksimacijski model. Isprobane su različite konfiguracije neprijatelja te je istražen utjecaj hiperparametara. Komentirani su konačni rezultati i dana je njihova interpretacija u kontekstu korištenih hiperparametara. Reinforcement learning is one of the main approaches to machine learning and it is applied to solve many interesting problems. One of the particularly interesting ones is the application in automatic game playing, which involves the creation of computer players that know how to play a certain game on their own. There are a number of algorithms used to teach such players and among the popular and widely used ones is the Q-learning algorithm. A copy of the famous game Space Invaders was made as part of this paper. A variant of Q-learning, which uses an artificial neural network as an approximation model, was used to train the player. Different enemy configurations were tested. The influence of hyperparameters was also examined. Final results are commented and their interpretation is given in the context of used hyperparameters.

Računalni vid je područje umjetne inteligencije koje se bavi omogućavanjem računala da prepoznaju i procesiraju razne objekte sa slika i videa na način kako je to svojstveno ljudima. Zahvaljujući ogromnoj količini vizualnih podataka koja se dnevno generira, razvojem novih algoritama i većom dostupnošću potrebne računalne snage, u posljednje vrijeme došlo je do značajnog napretka računalnog vida i sukladno tome primjenjuje se kod sve više zadataka koji uključuju prepoznavanje. Uzimajući ovo u obzir u ovom radu je razvijena softverska aplikacija za prepoznavanje maski na licu koja se bazira na umjetnim neuronskim mrežama i računalnom vidu. Ova softverska aplikacija izrađena je u programskom jeziku Python. Na kraju ovog rada provedena je evaluacija koja je uključivala ljudske subjekte. Computer vision is the field of artificial intelligence that deals with enabling computers to recognize and process various objects from images and videos in the way that is inherent to humans. Thanks to the huge amount of visual data generated daily, the development of new algorithms and greater availability of the necessary computing power, in recent times there has been a significant improvement in computer vision and accordingly it is applied to more and more tasks involving recognition. Taking this into account, a software application for face mask detection based on artificial neural network and computer vision was developed. This software application was developed in the Python programming language. At the end of this paper, evaluation was carried out which involved human subjects.

Govor je primarni način komuniciranja između ljudi, a izražene emocije igraju veliku ulogu u razumijevanju konteksta poruke prenesene govorom. Kako se razvija komunikacija između ljudi i računala potrebno je omogućiti računalima prepoznavanje ljudskih emocija. U ovome radu izrađen je softver koji na temelju umjetnih neuronskih mreža iz govora prepoznaje izražene emocije. Umjetne neuronske mreže obrađuju akustičke značajke iz govora kako bi prepoznale emocije. Izrađena su dva različita modela umjetnih neuronskih mreža, te su oba modela trenirana na dostupnim bazama podataka. Dva izrađena modela su eksperimentalno evaluirana na ljudskim subjektima. Speech is the primary means of communication between people, and expressed emotions play an important role in understanding the context of the message sent through it. As the communication between humans and computers develops it is necessary to enable machines to recognize human emotions. In this paper, a software has been developed which is based on artificial neural networks to recognize expressed emotions from speech. Artificial neural networks process acoustic features of speech to recognize emotions. Two different models of artificial neural networks have been developed, and both models have been trained on available databases. The two developed models were experimentally evaluated by human subjects.

Odjeli za ljudske resurse velikih tvrtki svakodnevno moraju pročitati veliki broj životopisa iz kojih je potrebno provjeriti i usporediti kompetencije kandidata kako bi odabrali prave ljude za tražene pozicije. U radu se ostvaruje programsko rješenje koje će za predane životopise automatski izvući kompetencije kandidata. Načelo rada programskog ostvarenja te potrebno teorijsko znanje koje pokriva programsko ostvarenje detaljno su objašnjeni. Programsko ostvarenje predstavlja problem izvlačenja informacija koji je riješen korištenjem modela neuronske mreže sastavljenog od LSTM i standardnih gustih slojeva, dok se obrada teksta vrši standardnim postupcima korištenih prilikom obrade prirodnog jezika. The human resources departments of large companies have to read substantial number of resumes on daily basis from which it is necessary to check and compare the competencies of the candidates in order to select the right people for the required job. The paper presents an implementation of a solution that will automatically extract the competencies of candidates for the given resume. Realization of the implementation and the necessary theoretical knowledge required for understanding the workings of the implementation are explained in detail. The implementation itself is an information extraction problem that is solved using an artificial neural network model consisting of LSTM and standard dense layers, while word processing is performed using standard procedures in natural language processing.

Razvojem metoda strojnoga učenja, uključivo sa umjetnim neuronskim mrežama, otvorile su se prilike za velikim skokovima u strojnoj tvorbi govora, posebice u sustavima koji to čine u objedinjenome ciklusu, end-to-end Text-To-Speech sustavima. Upravo zbog generalne prirode takvih modela, pristup korištenjem umjetnih neuronskih mreža omogućuje primjenu na razne jezike bez značajnih promjena u arhitekturi te pristupu samoga sustava. Sustav Tacotron 2, Mozilla TTS implementacijom, pokazao se kao idealan odabir primjenjiv na korpus hrvatskoga jezika. Ta je pretpostavka u ovome radu i dokazana, uspješnom tvorbom hrvatskoga govora uz minorne ili pak nepostojeće promjene u načinu izvođenja sustava. No, kao preduvjet kvalitetnim rezultatima, objašnjeni su pristupi obradi te pripremi podataka, te pokazana ovisnost kvalitete u kontekstu razumljivosti o količini podataka. With the appearance and development of machine learning models, and an emphasis on artificial neural networks, large leaps in the machine generation of human speech, or otherwise, have been made, especially when considering an end-to-end approach. It is because of the learning nature of the models that artificial neural networks exceed expectations when applied to different sets of data, in this case different languages. Tacotron 2, and the Mozilla TTS implementation, has proven to be the ideal choice for the synthesis of croatian speech, due to the earlier mentioned advantages. That hypothesis has been proven in this paper by successfully synthesizing croatian speech with little to no adjustments to the implementation itself, especially when it comes to the architecture. As a prerequisite for a working system, we have shown that the quality of generated speech, within the context of understandability, is dependant on the size of the dataset on which the model is being trained, while on the other hand predicting that the naturalness of speech rather depends more on the quality of the recordings themselves.

Rad se bavi umjetnim neuronskim mrežama te njihovom optimizacijom u kontekstu regresijskih mreža za predviđanje svojstava silicenskih nanovrpci. Dan je pregled umjetnih neuronskih mreža, njihovih mehanizama te raznih metoda za poboljšanje njihove funkcije. Spomenute su silicenske nanovrpce, njihova primjena u elektronici te potencijal umjetnih neuronskih mreža u predviđanju njihovih svojstava. Sprovedena je optimizacija regresijske neuronske mreže čiji je cilj predviđati jedno od transportnih svojstava silicenskih nanovrpci. The paper deals with artificial neural networks and their optimization in the context of regression networks for the prediction of silicene nanoribbon properties. An overview of artificial neural networks, their mechanisms, and diverse methods for the improvement of their function is given. Silicene nanoribbons, their usage in electronics, as well as the potential of neural networks in the prediction of their properties is mentioned. An optimization of a regression neural network whose goal is the prediction of a silicene nanoribbon transport property is conducted.

Govor je ljudima primarni način komunikacije, a emocije koje se njime prenose pomažu boljem međusobnom razumijevanju. Zbog sve učestalije komunikacije između ljudi i strojeva, postaje bitno da i strojevi budu u mogućnosti prepoznati ljudske emocije. U ovom radu napravljeno je softversko rješenje virtualnog agenta koji prepoznaje emocije na temelju govora. U tu svrhu korištene su akustičke i lingvističke značajke govora. Najvažniji dio softvera čine umjetne neuronske mreže koje su učene na dostupnim bazama podataka. Izrađeni softverski model eksperimentalno je evaluiran na ljudskim subjektima. Speech is primary means of communication between people, and the emotions it conveys help them to better understand each other. Due to more frequent communication between humans and machines, it is becoming important for machines to be able to recognize human emotions as well. In this paper, a software solution of a virtual agent which recognizes emotions based on speech is developed. For this purpose, acoustic and linguistic speech features were used. The most important part of the software consists of artificial neural networks which have been trained on available databases. The developed software model was experimentally evaluated on human subjects.

Poznavanje okoline temeljna je pretpostavka autonomnog djelovanja robota. Vizualna percepcija korištenjem dubinskih vizualnih senzora jedan je od načina dobivanja informacije o okolini robota. Detekcija i praćenje objekata u mjerenjima senzora ključni su zadaci percepcijskog sustava. U mnogim je praktičnim primjenama autonomne manipulacije potrebno ostvariti zadatke poput hvatanja, blokiranja ili udaranja gibajućih objekata. Zbog dinamičkih ograničenja robotske ruke, za uspješno ostvarivanje navedenih zadataka, osim detekcije i praćenja, nužno je i predviđati trajektoriju gibajućeg objekta. U ovom završnom radu razvijen je algoritam detekcije, praćenja i predviđanja trajektorije loptice dubinskim vizualnim senzorom. Knowing the robotic system's environment is one of the basic assumptions when using these systems. Visual perception using a visual-depth sensor is one of the ways of getting information about the robot's environment. Detection and object tracking is a crucial task of every perception system. In many practical applications of autonomous manipulation, it is necessary to accomplish tasks like catching, blocking, or hitting moving objects. Because of the dynamic limitations of the robotic arm, for the successful accomplishment of these tasks, besides detection and tracking, it is necessary to predict the trajectory of the moving object. In this paper, an algorithm for ball detection, tracking, and trajectory prediction using a depth-visual sensor is developed.

U ovom radu provedena je klasifikacija uzoraka opojnih droga upotrebom umjetnih neuronskih mreža. Za potrebe istog, najprije je provedena optimizacija mreža: određen je tip mreže, broj neurona u skrivenom sloju, te aktivacijska i funkcija za treniranje. Najboljim se pokazao višeslojni linearni perceptron s 12 neurona u skrivenom sloju, tangens hiperbolnom aktivacijom, treniran funkcijom konjugiranog gradijenta uz visku točnost od 84,93% ispravnog predviđanja na nezavisnom (vanjskom) setu varijabli. Umjetne neuronske mreže pokazale su se kao brz i pouzdan alat kojim može ostvariti kvalitetna klasifikacija uzoraka opojnih droga te kao takve sigurno imaju svijetlu perspektivu u rasvjetljavanju kriminalnih djelatnosti vezanih uz transfer opojnih droga. The classification of samples of narcotic drugs using artificial neural networks is shown in this paper. For the same purpose, network optimization was first implemented: a network type was determined, a number of neurons within the hidden layer, also an activation and training functions were determined. The best solution is multilayer linear perceptron with 12 neurons within the hidden layer, with tangent hyperbolic activation function, trained with conjugated gradient function in the high accuracy of 84.93% of the correct prediction on the independent (external) set of variables. Artificial neural networks have proven to be a fast and reliable tool that can achieve a high-quality classification of narcotic drugs samples and certainly have a bright perspective on the illicit trafficking of narcotic drugs.

Reducing green urban infrastructure pose a huge threat to cities sustainability. It is important to monitor and track health of vegetation. For efficient planning of urban development and maintenance of green urban infrastructure key is to have up to date maps. Using satellite imagery is the easiest way to cover large city areas. In order to map vegetation, there are many possible solutions, however most effective way is using machine learning methods. Machine learning is divided into supervised and unsupervised classification and each can be divided into several different methods. Many authors have considered different methods and they use them to access accuracy on satellite image information extraction. They have used different satellite images and naturally higher resolution imagery results in better classification. However, there is still lack of comprehensive analysis of more supervised machine learning methods on similar cities. This paper aims to provide analysis of four different machine learning methods: support vector machine, artificial neural network, naïve Bayes and random forest. The objective of the support vector machine algorithm is to find a hyperplane in an N-dimensional space that distinctly classifies the data points where hyperplanes are decision boundaries that help classify the data points and support vectors data points that are closer to the hyperplane and influence the position and orientation of the hyperplane. Artificial neural network are brain-inspired systems which are intended to replicate the way that humans learn. Neural networks consist of input and output layers, and hidden layers. They are excellent tools for finding patterns which are far too complex or numerous for a human programmer to extract and teach the machine to recognize. A Naive Bayes classifier is a probabilistic machine learning model that’s used for classification task and the crux of the classifier is based on the Bayes theorem. Random Forest creates a forest and makes it random and is an ensemble of Decision Trees, most of the time trained with the bagging method which general idea is that a combination of learning models increases the overall result. All of the mentioned methods will be tested on Sentinel-2A imagery. Sentinel-2A multispectral imager has thirteen sensors which is useful in vegetation extraction. Methods will be compared using error matrix and kappa statistics.

Razvojem blisko infracrvene spektroskopije (NIRs) kroz povijest dolazi do njene sve veće primjene u prehrambenoj industriji. Povećanje tržišnih i potrošačkih želja za kvalitetom hrane s pozitivnim zdravstvenim prednostima stvorilo je potrebu za učinkovitim i točnim analitičkim metodama gjde se NIR pokazala kao uspješna metoda. NIR se koristi za praćenje fizikalnih i kemijskih svojstava prehrambnih proizvoda, te se zbog specifičnosti i osjetljivosti na promjene u kemijskim i fizikalnim svojstvima uzorka koji se analizira može upotrijebiti za praćenje autentičnosti i podrijetla različitih prehrambenih proizvoda. Cilj ovog rada je bio ispitati mogućnost razlikovanja dvije sorte jabuka (granny smith i idared) i njihovih sokova te utjecaj dodatka različitih koncentracija šećera u sokove jabuka pomoću NIR spektroskopije. Rezultati su pokazali da je NIR spektroskopija uspješna prilikom razlikovanja sorti jabuka, njihovih sokova i različitih koncentracija šećera u sokovima. Development of near infrared spectroscopy (NIRs) through history is increasing its application in food industry. Increasing market and consumer desires for quality food with positive health benefits created the need for efficient and accurate analytical methods where NIR proved to be a successful method. NIR is used to monitor physical and chemical properties of food products, and due to the specificity and susceptibility to changes in the chemical and physical properties of the sample being analyzed, it can be used to track the authenticity and origin of different food products. The aim of this paper was to examine the possibility of distinguishing two apple varieties (granny smith and idared) and their juices and the influence of the addition of different concentrations of sugar in apple juices using NIR spectroscopy. The results showed that NIR spectroscopy was successful in distinguishing apple varieties, their juices, and varying sugar concentrations in juices.

Industrijska konoplja je biljna vrsta za koju se unatoč ozbiljnom zanemarivanju tijekom prošlog stoljeća, u posljednjih par desetljeća ponovno probudilo zanimanje. Uz ostale grane industrije i znanosti i u prehrambenom sektoru se također istražuju njene namjene. Jedan od načina je priprema čaja od industrijske konoplje. U ovom radu praćene su promjene električne vodljivosti, ukupno otopljenih čvrstih tvari i pH u ovisnosti o volumnom udjelu dodanog čaja konoplje u čajeve kamilice, lipe i mente. Osim izravnog mjerenja ovih fizikalnih svojstava, korištena je i blisko-infracrvena (NIR) spektroskopija. Za obradu podataka NIR spektara upotrijebljena je analiza glavnih komponenti, na temelju čega se pomoću umjetnih neuronskih mreža predviđao volumni udio i spomenuta fizikalna svojstva u uzorcima. Rezultati dobiveni na ovaj način pokazali su se pouzdaniji od izravne metode kod sva tri tipa čajnih mješavina. Također provedena je i senzorska procjena koja je pokazala najveći afinitet ispitanika za čajne mješavine s 5% dodanog čaja konoplje. Hemp is a herbaceous species that has recently seen a resurgence of interest, despite the serious neglect in the previous century. In addition to other branches of industry and science, in food sector its uses are also being explored. One of them is preparation of hemp tea. In this paper changes in electrical conductivity, total dissolved solids and pH were monitored depending on the volume of added hemp tea in chamomile, linden and peppermint tea. In addition, near infrared spectroscopy was used as an alternative method. Principal component analysis was used to process the spectrum data, based on which, by means of artificial neural networks, the volume fraction and physical properties in the samples were predicted. The results obtained in this way were more reliable for all three tea blends. A sensory evaluation was also performed that showed the highest liking for tea mixes with 5% added hemp tea.

U ovome su radu ukratko opisane osnovne značajke i arhitektura općenite neuronske mreže. Zatim je opisan rad dubokih konvolucijskih neuronskih mreža te pripadnih načina i metoda učenja. Paralelno uz rad implementiran je program koji omogućava detekciju lopte u snimkama sportskih utakmica, konkretno nogometnim snimkama. U završnom dijelu rada prokomentiran je dobiveni rezultat te je provedena analiza dali se što moglo i kako bolje napraviti. This paper briefly describes the basic features and architecture of the general neural network. Then, the work of deep convolutional neural networks and related ways and methods of learning are described. Parallel with the work, a program has been implemented that allows the detection of balls in sport videos, specifically football videos. In the final part of the work, the obtained result was commented and the analysis of what and how something could have been better was conducted.