Teleiletişim ve Enformatik Teknolojileri Uygulama ve Araştırma Merkezi
Yüklə 453,68 Kb.
|
| Alper Denli, Cemil Gurbuz, James Mechie and Michael Weber, Preliminary Results of 2-D Modelling Studies From Cyprus Arc ProjectExplosion Seismic Data, EGU General Assembly 2013, held 7-12 April, 2013 in Vienna, Austria, id. EGU2013-188
Alper Denli, Cemil Gurbuz, James Mechie and Michael Weber, Crustal Structure of the Central Anatolia – Eastern Mediterrenean, Turkey and Cyprus from Wide-Angle Seismic Data, EGU General Assembly 2016, held 17-22 April, 2016 in Vienna, Austria, id. EGU2016-392 Defense Jury Members
Defense Date: 28.03.2016 Erinç Dikici Thesis Supervisor: Assoc. Prof. Murat Saraçlar SUPERVISED, SEMI-SUPERVISED AND UNSUPERVISED METHODS IN DISCRIMINATIVE LANGUAGE MODELING FOR AUTOMATIC SPEECH RECOGNITION Abstract Discriminative language modeling aims to reduce the error rates by rescoring the output of an automatic speech recognition (ASR) system. Discriminative language model (DLM) training conventionally follows a supervised approach, using acoustic recordings together with their manual transcriptions (reference) as training examples, and the recognition performance is improved with increasing amount of such matched data. In this thesis we investigate the case where matched data for DLM training is limited or not available at all, and explore methods to improve ASR accuracy by incorporating unmatched acoustic and text data that come from separate sources. For semi-supervised training, we utilize weighted finite-state transducer and machine translation based confusion models to generate artificial hypotheses in addition to the real ASR hypotheses. For unsupervised training, we explore target output selection methods to replace the missing reference. We handle discriminative language modeling both as a structured prediction and a reranking problem and employ variants of the perceptron, MIRA and SVM algorithms adapted for both problems. We propose several hypothesis sampling approaches to decrease the complexity of algorithms and to increase the diversity of artificial hypotheses. We obtain significant improvements over baseline ASR accuracy even when there is no transcribed acoustic data available to train the DLM. PUBLICATIONS Journals
Conferences - International
Conferences - National
Defense Jury Members
Defense Date: 21.06.2016 İskender Haydaroğlu Thesis Supervisor: Assoc. Prof. Şenol Mutlu ENERGY HARVESTING WIRELESS OPTICAL MICROSYSTEMS
This thesis covers a novel approach to photovoltaic energy harvesting and optical data transmission in the context of millimeter-scaled smart autonomous microsystems through the use of a single light emitting diode (LED) to both efficiently harvest optical energy and transmit data to enable wireless, batteryless operation. A proof of concept design for demonstrating the viability of the use of a LED in the proposed manner, harvesting optical energy and transmitting a fixed device ID optically through the same LED using a transmitter based on a continually running charge pump is presented. Next, a low voltage temperature sensor design is integrated into the existing design, to prove by example that the harvested voltage from the LED is high enough that it requires no voltage boosting to power essential analog blocks such as sub-bandgap references, oscillators, and comparators, as opposed to integrated CMOS photovoltaic harvesting. Finally, an alternative, energy efficient optical transmitter architecture and a new ultra low power, ultra low energy temperature sensor are designed and integrated into a single chip. The scalable, inverter based switched capacitor boosting transmitter uses the trickle current from the LED to charge its capacitors directly with minimized losses in efficiency, transmitting data with 1 nJ/bit to a receiver designed and built in-house for up to 10 cm distance. The temperature sensor consumes less than 3 µW, features digital offset correction and an adaptive full-partial conversion algorithm to minimize the conversion time, effectively reducing energy per conversion from 0.6 nJ-3 nJ to 0.15 nJ-0.75 nJ. Total power consumption is in the order of 6 µW, harvested by a 0.1 mm2 LED, making the system viable for millimeter-scaled outdoor solar harvesting applications. All three designs were fabricated in UMC 0.18 µm CMOS process and tested in-house. PUBLICATIONS Journals
Conferences
Defense Jury Members
Defense Date: 13.05.2016 Kübra Kalkan Thesis Supervisor: Prof. Fatih Alagöz FILTERING BASED DEFENSE MECHANISMS AGAINST DDOS ATTACKS FOR CORE NETWORKS Abstract In this thesis, we present filtering based defense mechanisms against Distributed Denial of Service (DDoS) attacks for core networks. Initially, several filtering techniques are analyzed and their advantages and disadvantages are presented. A comparative classification of these methods is provided for security analysts. Classification results suggest that there are a few filtering methods that are both proactive and collaborative. Proactivity provides prevention of attacks before it spreads whereas collaboration enables getting knowledge about different points of the network and deciding filters together. Thus, we proposed a proactive and collaborative model called ScoreForCore. It is a statistical packet based defense mechanism that selects the most appropriate attributes for current attack traffic. Our results suggest that the success of the system’s behavior on legal and attack packets increased considerably. This strategy is also convenient for current emerging technology for core networks, called Software Defined Networking (SDN). It has several problems related to security that are largely induced by the centralized control paradigm. In that regard, DDoS attacks are specifically valid for SDN environment. Several defense mechanisms in SDN environment are analyzed and comparative classification is provided for rendering the current state of the art in the literature. Then, our defense strategy is applied on SDN environment with capable switches. This mechanism’s(SDNScore) results suggest that it gives perfect results for several known attacks and 84% success for an unknown attack. Since there is a trade-off between SDN paradigm and capable switches in SDNScore, we improved it and proposed another model called Joint Entropy based Scoring for SDN (JESS) that carries all burden to the controller and does not need capable switches. The results suggest that it is an elegant defense method for SDN environment. PUBLICATIONS Journals
Conferences
Defense Jury Members
Defense Date: 07.10.2016 Özlem Özmen Okur
DETAILED ANALYSIS OF VOXEL BASED MORPHOMETRY Abstract Voxel Based Morphometry, VBM, is one of the most widely used brain morphometry methods which aims to reveal the structural differences between the brain MR images of different populations. It is a whole brain and fully automatic approach in which all the images are registered onto a common template and then segmented into grey matter, white matter and cerebrospinal fluid. After an optional modulation step (regaining the original volume which is shrinked or enlarged during the registration), smoothing takes place in order to make the data more normally distributed and to diminish the inexact nature of the nonlinear registration. Finally, voxel-wise statistical operations are performed between the groups of the images. As revealed in several studies, changes in these steps and changes in their parameters might influence the resulting statistics. Although some short guidelines exist for conducting the processing stages, this thesis tries to explain each main step and gathers the discussions in the literature to make the VBM users aware of some pitfalls and limitations of VBM; and also gives brief descriptions about the other brain morphometry methods to give a view for where VBM stands at. In this thesis, the effect of modulation and masking strategy at the statistical stage were studied and concluded that not using the modulation and using average-based masking for the statistical part increased the detection power of VBM. Additionally, within the scope of this thesis, three clinical applications of VBM are performed and presented: Comparisons of the brain images of mathematicians, SSPE patients, and solvent abusers vs healthy controls. PUBLICATIONS Journals
Conferences
Defense Jury Members
Defense Date: 31.08.2016 Seda Yelkenci Thesis Supervisor: Prof. Dr. Mustafa Aktar Modelling 3D Seismic Wave Propagation in Marmara Region Abstract This study focuses on the modeling of 3D seismic wave propagation in the east of the Marmara Sea in particular for the city of Istanbul, which is identified as one of the megacities with the highest seismic risk in the world. For the first time, an attempt is made for creating a 3D seismic model and for testing the new model with real data. In the frame of constructing 3D velocity model, previous crustal studies of Marmara region and all other available field data, including surface and borehole measurements, are compiled to form a collection of 1D models. Each 1D model relates to a specific location point inside the study area. We have used interpolation methods, in particular Delaunay triangles approach, in order to fill in the no-data zones, which separate the 1D observation points. Elastic wave propagation is simulated inside the newly created 3D model using finite difference approach. An open source code called Wave Propagation Program (WPP), which operates on parallel processing environment, is used for that purpose. We have tested the performance of the 3D model with real data using the earthquake of September 29, 2004 (Ml=4.1) occurred in Çınarcık Basin, which was recorded by 18 permanent broadband stations and 100 strong motion stations. A detailed analysis of the source properties of the event is done, both for the location and the fault plane solution. Real and synthetic waveforms are compared both in time and frequency domains. Matching of the waveform shapes are studied in detail. In each case improvement of 3D model over 1D counterpart is discussed. A more quantitative evaluation of 1D and 3D performances is carried out using waveform correlation. The final result shows that a considerable improvement is achieved with 3D model both in terms of amplitudes and P and S arrival times. The finite difference method is also applied to specified basin structures filled with soft sediments of low shear velocities. Sabiha Gökçen Airport area in Pendik, is studied in detail because its basement geometry and sedimentary cover are well-known. The analysis, performed both in the time and frequency domain, helps to understand the characteristics of the 3D wave propagation inside the basin and the site effects related to it. PUBLICATIONS Conferences
Defense Jury Members
Defense Date: 25.05.2016 Aykut Yiğitel
GREEN NETWORKING: FROM CONVENTIONAL TO NEXT GENERATION HETEROGENEOUS CELLULAR NETWORKS Abstract
Increasing energy costs drive the telecommunication service providers to become highly interested in energy efficient operations. The exponential growth in mobile dataexchange which is further augmented by the rapid proliferation of smart phones increases the operational expenses of the cellular network operators significantly. Also, ecologists state that the primary triggering factor of the global warming is adding excessive amounts of greenhouse gases to the atmosphere and 72% of the totally emitted greenhouse gases is carbon dioxide (CO2). Increasing environmental awareness combined with the high energy prices has driven the network operators to reduce their CO2 footprint by adopting energy efficient green methods. In this thesis, our main focus is to save energy in three types of wireless cellular networks (i) Conventional Cellular Networks (ii) Packet-switched Cellular Networks and (iii) Next Generation Multi-tier Cellular Networks. We formulate novel mathematical optimization problems for each of the listed cellular networks to find the best possible topology which minimizes the overall power consumption of the network while satisfying a certain quality of service level. Our decision variables in the optimization models are switching base stations on/off and adaptively adjusting their transmission power levels as well as deploying additional pico base stations as a remedy according to the present traffic conditions. Although the optimization tools provide the optimum solutions for smaller instances of the problem, we propose novel heuristics to solve large-scale realistic instances due to their prohibitive complexity. Results of extensive simulations, which are designed as close to real life conditions as possible, show that the proposed green methods help to maintain an energy-aware network and save significant amount of energy by adjusting the network topology to the current traffic conditions adaptively. PUBLICATIONS Journals
International Conferences
Local Conferences
Defense Jury Members Prof. Cem Ersoy Boğaziçi University Assist Prof. Özlem Durmaz İncel Galatasaray University Prof. Necati Aras Boğaziçi University Prof. Can Özturan Boğaziçi University Prof. Tuna Tuğcu Boğaziçi University Assoc. Prof. Ayşegül Gençata Yayımlı İstanbul Teknik University Defense Date: 19.01.2016 Yüklə 453,68 Kb. Dostları ilə paylaş:
|