Construction of deep neutral networks using swarm intelligence to detect anomalies : master's thesisSašo Pavlič
, 2021, magistrsko delo
Opis: The design of neural network architecture is becoming more difficult as the complexity of the problems we tackle using machine learning increases. Many variables influence the performance of a neural model, and those variables are often limited by the researcher's prior knowledge and experience. In our master's thesis, we will focus on becoming familiar with evolutionary neural network design, anomaly detection techniques, and a deeper knowledge of autoencoders and their potential for application in unsupervised learning. Our practical objective will be to build a neural architecture search based on swarm intelligence, and construct an autoencoder architecture for anomaly detection in the MNIST dataset.
Ključne besede: neural architecture search, machine learning, swarm intelligence
Objavljeno v DKUM: 18.10.2021; Ogledov: 550; Prenosov: 82
Celotno besedilo (3,18 MB)
Gradivo ima več datotek! Več...
Implementation of a new reporting process in a group xSara Črešnik
, 2021, magistrsko delo
Opis: Reporting is present in every company. Whether it is small or big, it cannot be avoided. It plays a crucial role in the process and progress of business. The quality of reporting affects the development of the work environment and the company. Since business report is a document that contains business information, which supports the decisions about the future-oriented business decisions, it is very important for it to be designed in such a way that it contains the key information for the recipient and provides support for business decisions. The reporting process can take place horizontally upwards or downwards. Content and structure vary depending on the recipient of the report. We live in an age when our every step is accompanied by digitization, computerization, artificial intelligence, mass data, the Internet of Things, machine learning, and robotics. These changes have also affected the reporting process as well as its processes. The processes of data acquisition, processing and sharing have changed. Furthermore, the data quantity has increased, whereas the speed of the time in which to prepare the reports has decreased. We can have data without information, but we cannot have information without data. There is never enough time, especially nowadays when we are used to having everything at our fingertips. These are two conflicting factors – having more data and less time to prepare quality reports. The systems are developed to optimize the process, increase efficiency and quality and, what is nowadays most important, they have been created to obtain mass data in the shortest possible time. Therefore, it is important to adapt and implement software that can help achieve our daily tasks. We must know how to process huge amounts of real-time data and deliver the information they contain. It is crucial for companies to keep up with the environment and implement changes and innovations into their business process. A company is like a living organism for it must constantly evolve and grow. As soon as it stops growing and evolving, it can fail because it starts lagging and is therefore no longer competitive to others. To deliver faster feedback, companies need data of better quality. There are tools that can improve the business process, better facilitating the capacity of the human agents. The goal is to harness the employees’ full potential and knowledge for important tasks, such as analyzing, reviewing, and understanding data and acting upon them, invoking information technology to automate repetitive processes and facilitate better communication.
The focus in this master’s thesis is on the reporting process in Group X. Group X is one of the world leaders in the automotive industry, a multinational corporation based in Canada with subsidiaries around the world. The complexity of the business reporting that is implemented for the Headquarters in Canada has to address the complexity of the multinational corporation to support the decision process.
The aim of the thesis is to propose a reporting process for preparing and producing reports with a huge amount of data in a very time-efficient manner. We start by examining the existing processes and upon that, identifying the processes required for the reports to reach the final recipients. Our goal is to identify the toolset, which would increase efficiency, accuracy, credibility, and reduce errors in the fastest possible time. We investigate a short-term and a long-term solution. By a short-term solution, we mean a system, program, or a tool that can help us increase our potential by using digital resources, which are already existing in the organization. By a long-term solution, we mean a solution, which requires employment of specialized future tools in the field of reporting and in repetitive processes, which we can identify with current knowledge and expectations for development. This includes machine learning, robotic process automatization, artificial intelligence.
Ključne besede: Consolidated reporting, reporting process, robotic process automatization, business intelligence, artificial intelligence, machine learning, SharePoint, Big Data, digital transformation, electronic data interchange.
Objavljeno v DKUM: 01.09.2021; Ogledov: 427; Prenosov: 3
Celotno besedilo (1,71 MB)
Energy efficient system for detection of elephants with Machine Learning : master's thesisMarko Sagadin
, 2020, magistrsko delo
Opis: Human-Elephant Conflicts are a major problem in terms of elephant conservation.
According to WILDLABS, an average of 400 people and 100 elephants are killed every year in India alone because of them.
Early warning systems replace the role of human watchers and warn local communities of nearby, potentially life threatening, elephants, thus minimising the Human-Elephant Conflicts.
In this Master's thesis we present the structure of an early warning system, which consists of several low-power embedded systems equipped with thermal cameras and a single gateway.
To detect elephants from captured thermal images we used Machine Learning methods, specifically Convolutional Neural Networks.
The main focus of this thesis was the design, implementation and evaluation of Machine Learning models running on microcontrollers under low-power conditions.
We designed and trained several accurate image classification models, optimised them for on-device deployment and compared them against models trained with commercial software in terms of accuracy, inference speed and size.
While writing firmware, we ported a part of the TensorFlow library and created our own build system, suitable for the libopencm3 platform.
We also implemented reporting of inference results over the LoRaWAN network and described a possible server-size solution.
We finally a constructed fully functional embedded system from various development and evaluation boards, and evaluated its performance in terms of power consumption.
We show that embedded systems with Machine Learning capabilities are a viable solution to many real life problems.
Ključne besede: machine learning, microcontroller, on-device inference, thermal camera, low-power system
Objavljeno v DKUM: 06.01.2021; Ogledov: 638; Prenosov: 121
Celotno besedilo (13,35 MB)
Development of a Model for Predicting Brake Torque Using LSTM and TCN Models : magistrsko deloTomaž Roškar
, 2020, magistrsko delo
Opis: The main purpose of this thesis is to compare two state-of-the-art machine learning models, LSTM (Long Short-Term Memory) and TCN (Temporal Convolutional Network), on an AVL List GmbH case use, where the goal is to predict vehicle brake torque. Dataset used for model testing consists of multiple features which are preprocessed using several preprocessing methods. For model implementation Python’s libraries Keras and TensorFlow are used. Results from this thesis show that TCN is able to outperform LSTM. TCN achieves lower RMSE on the test dataset and is significantly faster in training and evaluation.
Ključne besede: brake torque, machine learning, neural network, LSTM, TCN, RNN, CNN
Objavljeno v DKUM: 24.09.2020; Ogledov: 696; Prenosov: 0
The Digital Pig: Automatic Systems for Behavior Detection in Weaned PigsAnja Lešnik
, 2020, diplomsko delo
Opis: In this bachelor's thesis, we used machine learning techniques to detect pigs in group pens, which would help to improve the welfare and comfort of pigs. Mask-RCNN was used for object segmentation. The implementation was based on Resnet101. The goal was to achieve the highest possible precision in detection of the pig's body, head, and tail. We predicted that the accuracy will be the highest for body detection and lower for head and tail detection. We also concluded that the difference in precision and recall will be less than 10% between hand-labeled bounding boxes and the predicted bounding boxes from our model. As predicted, body detection represented the highest results, as the accuracy of head and tail detection was lower. The difference between precision and recall was 10% for body detection and higher than 10% for head and tail detection. Precision of the body detection was 96%, as the whole body is easier to detect. The head detection precision score was 66%. Tail detection precision was 77%, which is a large difference compared to the percentage of head detection. The use of machine learning in livestock farming could be a potentially useful tool for detecting welfare in pigs, as it would reduce the frequency of aggressive behaviors and the number of injuries. In the future, we want to refine our model to achieve higher precision for head and tail detection. Once the algorithm has clearly detected all the pigs in the image, we will try to refine the model to detect different forms of behavior. This technology would help us to evaluate welfare, which would be improved if necessary.
Ključne besede: pig, pig annotation, behavior, welfare, machine learning
Objavljeno v DKUM: 08.09.2020; Ogledov: 560; Prenosov: 104
Celotno besedilo (1,40 MB)
Nature-inspired algorithms for hyperparameter optimization : magistrsko deloFilip Glojnarić
, 2019, magistrsko delo
Opis: This master thesis is focusing on the utilization of nature-inspired algorithms for hyperparameter optimization, how they work and how to use them. We present some existing methods for hyperparameter optimization as well as propose a novel method that is based on six different nature-inspired algorithms: Firefly algorithm, Grey Wolf Optimizer, Particle Swarm Optimization, Genetic algorithm, Differential Evolution, and Hybrid Bat algorithm. We also show the optimization results (set of hyperparameters) for each algorithm and we present the plots of the accuracy for each combination and handpicked one. In discussion of the results, we provide the answers on our research questions as well as propose ideas for future work.
Ključne besede: artificial intelligence, artificial neural networks, machine learning, nature-inspired algorithms, evolutionary algorithms
Objavljeno v DKUM: 09.12.2019; Ogledov: 1162; Prenosov: 93
Celotno besedilo (969,13 KB)
Classification of perimetric data for supporting glaucoma diagnosisJanja Belinc
, 2018, magistrsko delo
Opis: The aim of the study: Glaucoma is a chronic, progressive and asymptomatic retinal disease which results in an irreversible visual field loss. The main objective of this Master’s thesis work was to study the applicability of classification techniques for supporting glaucoma diagnosis.
Research Methodology: In this study perimetric data was obtained by SPARK strategy implemented in Oculus perimeters and provided by medical experts from the Hospital Universitario de Canarias (HUC). This data was used for constructing the feature vectors for the classification problem. Feature vectors of 66 values and feature vectors of 6 values were tested in the experiments. The proposed classification study attempted to: a) demonstrate that the studied classifiers were able to distinguish between “healthy” and “glaucomatous” eyes using only perimetric data, and b) analyse which feature vector design was the most suitable to accomplish this task.
Results: The classification results showed that classifiers performed better on 6 than on 66 perimetry values, which demonstrated the suitability of the 6 points selected by the SPARK strategy and supported its use in medical field.
Conclusion: In this study two remarkable findings for pattern recognition in perimetric data were obtained. Firstly, that reducing the dataset improved the efficiency of the studied classifier, and secondly, that simple pattern recognition models types were more efficient than complex ones.
Ključne besede: Eye disease, visual field, SPARK perimetry, pattern recognition, machine learning, supervised learning, ROC analysis
Objavljeno v DKUM: 27.08.2018; Ogledov: 836; Prenosov: 81
Celotno besedilo (4,25 MB)
Organizational learning supported by machine learning models coupled with general explanation methods : a case of B2B sales forecastingMarko Bohanec
, Marko Robnik Šikonja
, Mirjana Kljajić Borštnar
, 2017, izvirni znanstveni članek
Opis: Background and Purpose: The process of business to business (B2B) sales forecasting is a complex decision-making process. There are many approaches to support this process, but mainly it is still based on the subjective judgment of a decision-maker. The problem of B2B sales forecasting can be modeled as a classification problem. However, top performing machine learning (ML) models are black boxes and do not support transparent reasoning. The purpose of this research is to develop an organizational model using ML model coupled with general explanation methods. The goal is to support the decision-maker in the process of B2B sales forecasting.
Design/Methodology/Approach: Participatory approach of action design research was used to promote acceptance of the model among users. ML model was built following CRISP-DM methodology and utilizes R software environment.
Results: ML model was developed in several design cycles involving users. It was evaluated in the company for several months. Results suggest that based on the explanations of the ML model predictions the users’ forecasts improved. Furthermore, when the users embrace the proposed ML model and its explanations, they change their initial beliefs, make more accurate B2B sales predictions and detect other features of the process, not included in the ML model.
Conclusions: The proposed model promotes understanding, foster debate and validation of existing beliefs, and thus contributes to single and double-loop learning. Active participation of the users in the process of development, validation, and implementation has shown to be beneficial in creating trust and promotes acceptance in practice.
Ključne besede: decision support, organizational learning, machine learning, explanations
Objavljeno v DKUM: 01.09.2017; Ogledov: 1250; Prenosov: 211
Celotno besedilo (1,31 MB)
Gradivo ima več datotek! Več...
Link prediction in multiplex online social networksMahdi Jalili
, Yasin Orouskhani
, Milad Asgari
, Nazanin Alipourfard
, Matjaž Perc
, 2017, izvirni znanstveni članek
Opis: Online social networks play a major role in modern societies, and they have shaped the way social relationships evolve. Link prediction in social networks has many potential applications such as recommending new items to users, friendship suggestion and discovering spurious connections. Many real social networks evolve the connections in multiple layers (e.g. multiple social networking platforms). In this article, we study the link prediction problem in multiplex networks. As an example, we consider a multiplex network of Twitter (as a microblogging service) and Foursquare (as a location-based social network). We consider social networks of the same users in these two platforms and develop a meta-path-based algorithm for predicting the links. The connectivity information of the two layers is used to predict the links in Foursquare network. Three classical classifiers (naive Bayes, support vector machines (SVM) and K-nearest neighbour) are used for the classification task. Although the networks are not highly correlated in the layers, our experiments show that including the cross-layer information significantly improves the prediction performance. The SVM classifier results in the best performance with an average accuracy of 89%.
Ključne besede: social networks, complex networks, signed networks, link prediction, machine learning
Objavljeno v DKUM: 08.08.2017; Ogledov: 1042; Prenosov: 401
Celotno besedilo (940,17 KB)
Gradivo ima več datotek! Več...