Master Degree / Yüksek Lisans Tezleri
Permanent URI for this collectionhttps://hdl.handle.net/11147/3008
Browse
14 results
Search Results
Master Thesis Hardware acceleration with fpga Based electronic boards for machine learning(01. Izmir Institute of Technology, 2024) Akkuş, Batuhan; Gümüş, Abdurrahman; Apaydın, Mehmet SerkanSon yıllardaki makine ög˘renmesi algoritmalarındaki gelis¸meler uç cihazlardaki kullanımını da arttırmıs¸tır (Merenda et al., 2020). Makine ög˘renimi algoritmaları genel- likle GPU tabanlı bilgisayarlarda çalıs¸tırılmaktadır, bu da yüksek enerji tüketimi (De- sislavov et al., 2021), yog˘un donanım kaynag˘ı gereksinimleri ve büyük fiziksel boyutları (Liu et al., 2022) nedeniyle uç cihazlar için uygun olmamaktadır. Bu tez, donanım hızlandırıcısı olarak FPGA platformlarında makine ög˘renmesi algoritmalarının, özellikle derin sinir ag˘larının uygulanması ve çıkarım yapılmasını aras¸tırarak, düs¸ük güç tüke- timi, verimli donanım kullanımı ve yüksek çıkarım performansı elde etmeyi hedefle- mektedir. Bu sistemlerin uç cihazlara adaptasyonu için esneklig˘i ve verimlilig˘i artırmak amacıyla, CNV ag˘ının (Umuroglu et al., 2017b) daha hafif bir varyasyonu olan CNV light gelis¸tirilmis¸. Bu ag˘, PyTorch tabanlı bir araç olan Brevitas (Pappalardo et al., 2019) ile nicemleme-farkında-eg˘itim yöntemi, kullanılarak 1, 2, 4 ve 8-bit seviyelerine nicemleme yapılmıs¸tır. CNV light ag˘ı CIFAR-10, SVHN, GTSRB ve MNIST veri setleri üzerinde Brevitas ile eg˘itilmis¸tir. Modeller FINN çerçevesi (Umuroglu et al., 2017a) kullanılarak FPGA'ya sentezlenmis¸tir. Modeller en fazla, en az ve sabit FPS seviye donanım kul- lanımına göre ayarlanmıs¸tır. Xilinx XC7Z020-1CLG400C FPGA, modelin metriklerini deg˘erlendirmek ve raporlamak için kullanılmıs¸tır. GTSRB veri setinde, ikili (W1A1) nicemleme yapılmıs¸ CNV light ag˘ı, tüm donanım kullanımları için %95.12 dog˘ruluk ve en fazla donanım kullanımında 12,191 FPS performansı ve 3.20W güç tüketimi elde etti, minimum donanım kullanımı için ise 6 FPS ve 1.62W güç tüketti. Sonuçlar, FPGA'ların uç cihazlarda makine ög˘renmesi modellerini verimli ve ölçeklenebilir platformlar olarak kullanılabileceg˘ini göstermektedir.Master Thesis Analysis of fingerprint matching performance with deep neural networks(01. Izmir Institute of Technology, 2022) Göçen, Alper; Erdoğmuş, NesliFingerprints are unique biometric properties for each person. In the literature and industry, they are widely used for identification purposes. Collecting biometric datasets is a tedious work since it is not possible without the owners’ consent, and existing fingerprint datasets are either not sufficient to use in deep learning tasks by means of size or most of them are kept private to the collectors’ use. This increases the need of synthetic fingerprint images and their use in a variety of tasks especially for training deep learning models. In this study, the performance of a CNN architecture named Finger ConvNet[1] is compared to well-known networks and the question of whether a mixed dataset consisting of synthetically generated and real fingerprint images can reach a performance close or equal to ones having only real images is discussed. As a result of experiments, it is shown that the number of real images in the dataset is an important factor and that the performance of the mixed dataset was less than the one having only real images proposed in the referred study.Master Thesis Application of Graph Neural Networks on Software Modeling(01. Izmir Institute of Technology, 2020) Leblebici, Onur Yusuf; Tuğlular, Tuğkan; Belli, FevziDeficiencies and inconsistencies introduced during the modeling of software systems can cause undesirable consequences that may result in high costs and negatively affect the quality of all developments made using these models. Therefore, creating better models will help the software engineers to build better software systems that meet expectations. One of the software modelling methods used for analysis of graphical user interfaces is Event Sequence Graphs (ESG). The goal of this thesis is to propose a method that predicts missing or forgotten links between events defined in an ESG via Graph Neural Networks (GNN). A five-step process consisting of the following steps is proposed: (i) data collection from ESG model, (ii) dataset transformation, (iii) GNN model training, (iv) validation of trained model and (v) testing the model on unseen data. Three performance metrics, namely cross entropy loss, area under curve and accuracy, were used to measure the performance of the GNN models. Examining the results of the experiments performed on different datasets and different variations of GNN, shows that even with relatively small datasets prepared from ESG models, predicts missing or forgotten links between events defined in an ESG can be achieved.Master Thesis Trajectory Prediction of Moving Objects by Means of Neural Networks(Izmir Institute of Technology, 1997) Barışık, Hakan; Aytaç, İsmail Sıtkı; Aytaç, İsmail SıtkıEstimating the three-dimensional motion of an object from a sequence of object positions and orientation is of significant importance in variety of applications in control and robotics. For instance, autonomous navigation, manipulation, servo, tracking, planning and surveillance needs prediction of motion parameters. Although "motion estimation" is an old problem (the formulations date back to the beginning of the century), only recently scientists have provided with the tools from nonlinear system estimation theory to solve this problem eural Networks are the ones which have recently been used in many nonlinear dynamic system parameter estimation context. The approximating ability of the neural network is used to identifY the relation between system variables and parameters of a dynamic system. The position, velocity and acceleration of the object are estimated by several neural networks using the II most recent measurements of the object coordinates as input to the system Several neural network topologies with different configurations are introduced and utilized in the solution of the problem. Training schemes for each configuration are given in detail. Simulation results for prediction of motion having different characteristics via different architectures with alternative configurations are presented comparatively.Master Thesis Image Classifcation by Means of Pattern Recognition Techniques(Izmir Institute of Technology, 1997) Güzel, Cumhur; Püskülcü, HalisImage classification plays an important role in many computer vision tasks such as surface inspection, shape determination etc. Various 2-D image classification techniques are investigated, assessed and a computational method to classifY the 2-D X-ray images is developed and evaluated in this study. Various pattern recognition techniques are devised for the solution of the image classification. Those techniques may be divided into mainly two groups. First one, is mathematical and statistical model based, second one, is the artificial neural network based techniques. We have concentrated on artificial neural network techniques. In the experiments, both techniques were applied for the classification of the VUR (vesico ureteral reflux) images, in this study. However, according to the experiments performed on VUR case study, neural network technique was more successful than others, in terms of classifier. A hybrid method is proposed in this study, rather than pure artificial neural network solution. Representation of images is performed via transformation invariant mathematical structure called Fourier Descriptors and these structures are used as input to train the neural network for the classification part.The application is performed as follows: Feature extraction is performed first, then extracted features are used as pattern vectors for training the neural network. Representation of the shapes in X-ray images is performed by using Fourier Descriptors. Usage of Fourier descriptors as a method of representation of the shapes, provides the transformation invariant' (translation, rotation, scaling invariant structure) representation of X-ray images. These new vector representation is fed to the neural network. Backpropagation is used as a training algorithm. After training is finished, system is readyfor questioning. The minimum-mean-distance and nearest neighbor rules are also applied for the pattern vectors generated for the experiments. But the multilayer perceptron trained by backpropagation outperforms both of these statistical classifiers.Master Thesis Prediction of Energy Consumption of Residential Buildings by Artificial Neural Networks and Fuzzy Logic(Izmir Institute of Technology, 2012) Turhan, Cihan; Gökçen Akkurt, GüldenThere are several ways to attempt to forecast building energy consumption. Different techniques, varying from simple regression to dynamic models that are based on physical principles, can be used for simulation. A frequent hypothesis for all these models is that the input variables should be based on realistic data when they are available, otherwise the evaluation of energy consumption might be under or over estimated. The aim of this thesis is to create simple models based on artificial intelligence methods (artificial neural networks and fuzzy logic) as predicting tools and to compare these methods with a building energy performance software (KEP-IYTE ESS). Architectural projects and heat load calculation reports of 148 apartment buildings (5-13 storey) from three municipalities in Ġzmir provide the input data for the models and software. Building energy consumption is modeled as a function of zoning status, heating system type, number of floors, wall overall heat transfer coefficient, glass type, area/volume ratio, existence of insulation, total external surface area, orientation, number of flats, total external surface area/total useful area, total windows area/total external surface area, width/length, total wall area/total useful floor area, total lighting requirement/total useful floor area and total wall area. Four different artificial neural network models and one fuzzy logic model were constructed, trained, tested and the results were compared with the software outcomes. The lowest mean absolute percentage error (MAPE) and mean absolute deviation (MAD) of ANN models appeared to be 4.1% and 6.57, respectively, which shows that ANN can make accurate predictions. On the other hand, fuzzy model gave an 4.86% and 7.59 of MAPE and MAD, respectively, which can be considered as sufficient accuracy.Master Thesis Data Driven Modeling Using Reinforcement Learning in Autonomous Agents(Izmir Institute of Technology, 2003) Karakurt, Murat; Özdemir, SerhanThis research has aspired to build a system which is capable of solving problems by means of its past experience, especially an autonomous agent that can learn from trial and error sequences. To achieve this, connectionist neural network architectures are combined with the reinforcement learning methods. And the credit assignment problem in multi layer perceptron (MLP) architectures is altered. In classical credit assignment problems, actual output of the system and the previously known data in which the system tries to approximate are compared and the discrepancy between them is attempted to be minimized. However, temporal difference credit assignment depends on the temporary successive outputs. By this new method, it is more feasible to find the relation between each event rather than their consequences.Also in this thesis k-means algorithm is modified. Moreover MLP architectures is written in C++ environment, like Backpropagation, Radial Basis Function Networks, Radial Basis Function Link Net, Self-organized neural network, k-means algorithm.And with their combination for the Reinforcement learning, temporal difference learning, and Q-learning architectures were realized, all these algorithms are simulated, and these simulations are created in C++ environment.As a result, reinforcement learning methods used have two main disadvantages during the process of creating autonomous agent. Firstly its training time is too long, and too many input parameters are needed to train the system. Hence it is seen that hardware implementation is not feasible yet. Further research is considered necessary.Master Thesis The Control of a Manipulator Using Cerebellar Model Articulation Controllers(Izmir Institute of Technology, 2003) Darka, Murat; Özdemir, SerhanThe emergence of the theory of artificial neural networks has made it possible to develop neural learning schemes that can be used to obtain alternative solutions to complex problems such as inverse kinematic control for robotic systems. The cerebellar model articulation controller (CMAC) is a neural network topology commonly used in the field of robotic control which was formulated in the 1970s by Albus. In this thesis, CMAC neural networks are analyzed in detail. Optimum network parameters and training techniques are discussed. The relationship between CMAC network parameters and training techniques are presented. An appropriate CMAC network is designed for the inverse kinematic control of a two-link robot manipulator.Master Thesis Artificial Neural Networks and Fuzzy Logic Applications in Modeling the Compressive Strength of Portland Cement(Izmir Institute of Technology, 2004) Can, Sever; Akkurt, SedatPortland cement production is a complex process that involves the effect of several processing parameters on the quality control of 28-day cement compressive strength (CCS). There are some chemical parameters like the C3S, C2S, C3A, C4AF, and SO3 contents in addition to the physical parameters like Blaine (surface area) and particle size distribution. These factors are all effective in producing a single quantity of 28-day CCS. The long duration of 28 day CCS test provided the motivation for research on predictive models. The purpose for these studies was to be able to predict the strength instead of waiting for 28 days for the test to be complete. In this thesis, artificial intelligence (AI) methods like artificial neural networks (ANNs) and fuzzy logic were used in the modeling of the 28-day CCS. The two models were compared for their quality of fit and for the ease of application.Quality control data from a local cement plant were used in the modeling studies. The data were separated randomly into two parts: the first one contained 100 data points to be used in training and the second part had 50 data points to be used in testing stages of the models. In this study, four different AI models were created and tested (3 ANN, 1 fuzzy logic). One of the ANN models (Model A) had 20 input parameters in 20x20x1 architecture with testing average absolute percentage error (AAPE) of 2.24%. The other ANN model (Model B) had four input parameters (SO3, C3S, Blaine and total alkali amount) in 4x4x1 architecture with AAPE of 2.41%. Both of the Model A and the Model B were created in the MatLAB® environment by writinga custom computer code. The last ANN model (Model C) actually refers to 72 differentANN models created in the MatLAB® neural networks toolbox. In order to obtain a model with the lowest error, different learning algorithms, training functions and architectures in combinations were tested. The lowest AAPE among these models appeared to be 2.31%. The fuzzy logic model (Model D) which had four input parameters (SO3, C3S, Blaine and total alkali amount) was created in the MatLAB fuzzy logic toolbox. In order to write the fuzzy rules, the sensitivity analysis of the Model B was utilized. The AAPE of the Model D was 2.69%. The model was compared with the ANN models for its error levels and ease of application. The results indicated that through the application of fuzzy logic algorithm, a more user friendly andmore explicit model than the ANNs could be produced within successfully low error margins.Master Thesis "ann" Artifical Neural Networks and Fuzzy Logic Models for Cooling Load Prediction(Izmir Institute of Technology, 2005) Bozokalfa, Gökhan; Akkurt, SedatIn this thesis Artificial Neural Networks (ANN) and fuzzy logic models of the building energy use predictions were created. Data collected from a Hawaian 42 storey commercial building chiller plant power consumption and independent hourly climate data were obtained from the National Climate Data Center of the USA. These data were used in both ANN and the fuzzy model setting up and testing. The tropical climate data consisted of dry bulb temperature, wet bulb temperature, dew point temperature, relative humidity percentage, wind speed and wind direction.Both input variables and the output variable of the central chiller plant power consumption were fuzzified, and fuzzy membership functions were employed. The Mamdani fuzzy rules (32 rule) in If .Then format with the centre of gravity (COG; centroid) defuzzification were employed. The average percentage error levels in the fuzzy model and the ANN model were end up with 11.6% (R2.0.88) and 10.3% (R2.0.87), respectively. The fuzzy model is successfully presented for predicting chiller plant energy use in tropical climates with small seasonal and daily variations that makes this fuzzy model.