No description available.

No description available.

Title: Predicting the Behaviour of Streaming Services Users Autor: Bc. Šimon Stachura Department: Katedra teoretické informatiky a matematické logiky Supervisor: Mgr. Martin Pilát, Ph. D. Abstract: Streaming services are one of the phenomena of the last decade, allowing online legal access to media for a large number of users. The media is usually provided to the users as an automatically generated sequence, created by some form of a recommender system. The preferences of individual users are usually estimated based on historical data from their previous usage of the service. Skipping behaviour on individual elements of the generated sequence (songs, for instance) is one of the basic signals expressing these preferences. Goal of this work is to predict users' behaviour based on their previous experience with the service. We chose a large dataset consisting of real data from usage of the Spotify service, and considered options for preprocessing and representing them. We decided to use recurrent neural networks with the Encoder-Decoder architecture for modelling the behaviour of the users. These models encode the information about users' historical behaviour into a compact inner representation of the session, and based on that representation they generate expected behaviour in the next time steps. We created...

Perennial warm-season grasses constitute the backbone of many forage production systems, whether for grazing or harvested feed. North American native plants, specifically grasses, forbs and legumes offer unique ecosystem benefits along with forage quality and digestibility that are unmatched by introduced species. The disparity in breeding and research focused on improvement of introduced species as opposed to native genera has led to inflated use of introduced species as forage types in lieu of native options, due to their unimproved nature. Eastern gamagrass [Tripsacum dactyloides (L.) L.] is proven to be a widely adapted, highly productive forage species in the southeast, Great Plains and northeast United States. A major limitation to more widespread use of eastern gamagrass is high seed dormancy, which leads to increased seed cost. Here, research used recurrent phenotypic selection breeding methods to reduce seed dormancy, with the ultimate goal of developing a population of individuals that produce non-dormant eastern gamagrass seed.

In the present work, we investigate the capacity of machine learning to reconstruct simulated large scale surface temperature anomalies given a sparse observation field. Several methods are combined: self-organizing maps and recurrent neural networks of the temporal trajectory. To evaluate our global scale reconstruction, we base our validation on global climate indices time series and EOF analysis. In our experiments, the obtained reconstructions of the global surface temperature anomalies provide a good correlation (over 90%) with the target values when considering scarce available observations sampling about 0.5% of the globe. We reconstruct the surface temperature anomaly fields from 0.05% of total number of data points. We obtain an RMSE of 0.39°C. We further validate the quality of the results calculating a correlation of 0.92, 0.97 and 0.98 between the reconstructed and target indices of AMO, ENSO and IPO. / Klimatsystemet består av olika komponenter inklusive atmosfären, havet och jorden. Som ett öppet system utbyter det hela tiden energi med resten av universum. Det är också ett dynamiskt system vars utveckling kan förutsägas av kända fysiska lagar. Interaktionen mellan dess olika komponenter leder till en så kallad naturlig variation. Denna variabilitet återspeglas i form av svängningslägen, inklusive AMO, ENSO och IPO. För att studera dessa variationer har vi klimatmodeller som representerar de olika krafterna och deras effekt på klimatförändringar på lång sikt. I detta sammanhang är variationerna i det förflutna klimatet särskilt intressanta och tillåter oss en bättre förståelse av klimatförändringar och bättre förutsäga den framtida utvecklingen. Men för att studera det förflutna klimatet eller paleoklimat är den enda tillgängliga informationen endast fullständig under de senaste 150 åren. Innan dess är de enda tillgängliga indikatorerna naturliga, kallad klimatproxy, som trädringar eller iskärnor. Vi kan härleda tidsserier med klimatdata, till exempel temperatur. Denna information är emellertid knappast tillfälligt såväl som över hela världen. Återskapa det globala klimatet från sådana data hanteras fortfarande dåligt. Länken mellan lokal information och global klimat studeras här med hjälp av statistiska metoder, inklusive neurala nätverk. Det långsiktiga målet med denna studie är att bygga en metod för att rekonstruera paleoklimatet från data om klimatproxy, vi fokuserar inledningsvis på rekonstruktionen av ett så kallat perfekt klimat, det vill säga en modell som endast tar hänsyn till naturlig variation, från rumsligt sällsynta tidsserier. De studerade uppgifterna är de från globala yttemperaturutgångar från den havsatmosfärkopplade IPSL-modellen. Uppgifterna förbehandlas för att ta bort säsongens genomsnittliga cykel och omvandlas till temperaturavvikelser. Dessutom väljs rutnätpunkter som representerar information om proxyer pseudo-slumpmässigt, med respekt för den verkliga dispositionen av dessa, övervägande i norr på kontinenterna. Uppgifterna delas upp i träningsdata (150 år), validering (30 år) och testdata (120 år). De metoder som används kombinerar (1) självorganiserande kartor och hierarkisk stigande klassificering, användbara för att producera en reducerad storlek av inmatningsdata, här baserat på tidskorrelationen mellan temperaturutvecklingen under 150 år, (2) ItCompSOM använder korrelationen mellan klasser erhållna genom självorganiserande kartor för att rekonstruera obevakad data, (3) återkommande nervnätverk för att förklara den temporära komponenten i data och förbättra den tidigare rekonstruktionen. Slutligen är definitionen av nya mätvärden nödvändig för att validera de föreslagna modellerna. Utvärderingen av produkterna görs således genom temporär rekonstruktion av AMO, ENSO, IPO klimatlägen samt genom projicering av huvudkomponenterna i analysen av huvudkomponenterna i inputdata. Således konstrueras en reducerad modell av globala temperaturdata baserad på 150 års fullständiga data först, vilket reducerar den rumsliga informationen från 9216 rutnätpunkter till 191 regioner associerade med 1 medelvärde vardera. För att ansluta denna modell till tidssekvenser av sällsynta temperaturer i världen antas det att varje klass som innefattar minst en observerad proxy-data är känd. Rekonstruktionen av globala yttemperaturutvecklingar med ItCompSOM ger en korrelation till indexen på mer än 90% för endast 0,5% av de initiala observationerna. Detta resultat förbättras kraftigt tack vare återkommande nervnätverk, vilket leder till en korrelation av 0,92, 0,97 respektive 0,98 för AMO, ENSO och IPO med endast 0,05% av observationerna. Dessa poäng förklaras med den använda metoden, regionaliseringen hjälper till att koncentrera informationen. Medan 0,5% av rutpunkterna är lika med 43 poäng, om de är korrekt fördelade, representerar de 22% av informationen om regionerna (43 av 191). Dessa mycket uppmuntrande resultat återstår att tillämpas på verkliga klimatproblem, det vill säga med hänsyn till å ena sidan den externa och antropologiska kraften, osäkerheterna relaterade till de verkliga uppgifterna om ombud å andrasidan.

Supervised machine learning (ML) projects require data for model training, validation, and testing. However, the confidential nature of field and well production data often hinders the progress of ML projects. To address this issue, we developed a well simulator that generates realistic well production data based on physical, governing differential equations. The simulation models the reservoir, wellbore, flowline, and choke coupled using transient nodal analysis to solve for transient flow rate, pressure, and temperature as a function of variable choke opening over time in addition to a wide range of static parameters for each component. The simulator’s output is then perturbed using the gauge transfer function to introduce systematic and random errors, creating a dataset for ML projects without the need for confidential production data. We then generated a simulated dataset to train a recurrent neural network (RNN) on the task of classifying well on/off times. This task typically requires a significant number of manhours to manually filter and verify data for hundreds or thousands of wells. Our RNN model achieves high accuracy in classifying the correct on/off labels, representing a promising step towards a fully-automated rate allocation process. Our simulator for well production data can be used for other ML projects, circumventing the need for confidential data, and enabling the study and development of different ML models to streamline and automate various oil and gas work processes. Overall, the success of our RNN model demonstrates the potential of ML to improve the operational efficiency of various oil and gas work processes.

No description available.

<div>Unmanned aerial vehicles (UAVs) have gained more attention in recent years because of their ability to execute various missions. However, recent works have identified vulnerabilities in UAV systems that make them more readily prone to cyberattacks. In this work, the vulnerabilities in the communication channel between the UAV and ground control station are exploited to implement cyberattacks, specifically, the denial of service and false data injection attacks. Unlike other related studies that implemented attacks in simulations, we demonstrate the actual implementation of these attacks on a Holybro S500 quadrotor with PX4 autopilot firmware and MAVLink communication protocol.</div><div><br></div><div>The goal was to create a lightweight intrusion detection system (IDS) that leverages recurrent neural networks (RNNs) to accurately detect cyberattacks, even when implemented on a resource-constrained platform. Different types of RNNs, including simple RNNs, long short-term memory, gated recurrent units, and simple recurrent units, were trained and tested on actual experimental data. A recursive feature elimination approach was carried out on selected features to remove redundant features and to create a lighter RNN IDS model. We also studied the resource consumption of these RNNs on an Arduino Uno board, the lowest-cost companion computer that can be implemented with PX4 autopilot firmware and Pixhawk autopilot boards. The results show that a simple RNN has the best accuracy while also satisfying the constraints of the selected computer.<br></div>

No description available.

By organically fusing the methods of convolutional neural network (CNN) and recurrent neural network (RNN), this dissertation focuses on the application of optical character recognition and image classification processing. The first part of this dissertation presents an end-to-end novel receipt recognition system for capturing effective information from receipts (CEIR). The main contributions of this research part are divided into three parts. First, this research develops a preprocessing method for receipt images. Second, the modified connectionist text proposal network is introduced to execute text detection. Third, the CEIR combines the convolutional recurrent neural network with the connectionist temporal classification with maximum entropy regularization as a loss function to update the weights in networks and extract the characters from receipt. The CEIR system is validated with the scanned receipts optical character recognition and information extraction (SROIE) database. Furthermore, the CEIR system has strong robustness and can be extended to a variety of different scenarios beyond receipts. For the convolutional recurrent structure application of land use image classification, this dissertation comes up with a novel deep learning model for land use classification, the convolutional recurrent land use classifier (CRLUC), which further improves the accuracy in classifying remote sensing land use images. Besides, the convolutional fully-connected neural networks with hard sample memory pool structure (CFMP) is invented to tackle the remote sensing land use image classification tasks. The CRLUC and CFMP algorithm performances are tested in popular datasets. Experimental studies show the proposed algorithms can classify images with higher accuracy and fewer training episodes compared to popular image classification algorithms.