Search

Convolutional Neural Network (CNN) has undergone tremendous advancements in recent years, but visual reasoning tasks are still a huge undertaking, particularly in few-shot learning cases. Little is known, especially in solving the Same-Different (SD) task, which is a type of visual reasoning task that requires seeking pattern repetitions in a single image. In this thesis, we propose a patch-as-filter method focusing on solving the SD tasks with few-shot learning. Firstly, a patch in an individual image is detected. Then, transformations are learned to create sample-specific convolutional filters. After applying these filters on the original input images, we, lastly, acquire feature maps indicating the duplicate segments. We show experimentally that our approach achieves the state-of-the-art few-shot performance on the Synthetic Visual Reasoning Test (SVRT) SD tasks by accuracy going up above 30% on average, with only ten training samples. Besides that, to further evaluate the effectiveness of our approach, SVRT-like tasks are generated with more difficult visual reasoning concepts. The results suggest that the average accuracy is increased by approximately 10% compared to several popular few-shot algorithms. The method we suggest here has shed new light upon new CNN approaches in solving the SD tasks with few-shot learning.

In this paper, a new type of trajectory, based on an exponential jerk, is presented along with filters for their online generation. The goal is to generalize constant jerk trajectories, widely used in industrial applications, in order to reduce vibrations of motion systems. As a matter of fact, constant jerk trajectories do not assure a complete vibration suppression when the damping of the resonant modes is not negligible. The values of the parameters (decay rate and duration) of the jerk impulses that allow residual vibration cancellation are derived in an analytical way as a function of the dynamic characteristics of the plant. Comparisons with the well-known input shaping techniques and with system-inversion-based filters show the advantages of the proposed method in terms of robustness with respect to modeling errors, smoothness of the resulting trajectory, and time duration of the motion under velocity and acceleration constraints.

Ankara : The Materials Science and Nanotechnology Program of the Graduate School of Engineering and Science of Bilkent University, 2014. Thesis (Master's) -- Bilkent University, 2014. Includes bibliographical references leaves 54-57. In this study, we design and produce grating coupled surface plasmon surfaces which are switched by electrochemistry. Grating structures are fabricated using digital versatile discs (DVDs) which are commercially available. According to atomic force microscopy (AFM) results, we categorize the different grating structures in two groups, namely shallow and deep gratings. Plasmonic properties of the surfaces are investigated using numerical simulations. Gold and silver are used as plasmon supporting metallic layers on gratings. Refractive index sensitivity of the plasmon resonances are studied using deionized water, air and glycerol solutions as the dielectric media and results are compared with simulations. Experimental results are coherent with the simulations in terms of reflection spectra. Electrochemical switching of plasmonic properties may have applications in tunable and switchable filters, as well as enhanced colorimetric sensing. We deposit ultrathin films of copper on plasmonic surfaces and investigate reversible changes in the plasmonic properties. Copper sulfate is selected as the electrolyte. Cyclic voltammetry is performed on plasmonic surfaces while monitoring optical reflectance. Copper is observed to deposit in the form of nanoislands on silver and gold films rather than uniform thin films. The effect of copper deposition on the plasmonic properties of the grating structure is simulated by Lumerical software and is seen to be two fold. For small effective thickness of copper nanoislands, the plasmon resonance condition shifts, whereas for thicker copper deposition plasmonic resonances are eliminated. Finally, copper's oxidation and reduction reactions are controlled by changing applied voltage thus shifting the resonance wavelength. Resonances are switched reversibly multiple times not only for different molarities but also for different grating sructures and plasmon supporting metallic layers . In summary, we demonstrate that plasmonic properties of nanostructured metallic surfaces can be controlled by electrochemistry. Switchable resonance surfaces can be used as dynamic filters or may enhanced contrast in plasmon resonance imaging applications. Karayalçın, Nihat Serkan M.S.

In this work, several procedures for the fault detection and isolation (FDI) on general aviation aircraft sensors are presented. In order to provide a comprehensive wide-spectrum treatment, both linear and nonlinear, model-based and data-driven methodologies are considered. The main contributions of the paper are related to the development of both FDI polynomial method (PM) and FDI scheme based on the nonLinear geometric approach (NLGA). As to the PM, the obtained results highlight a good trade-off between solution complexity and resulting performances. Moreover, the proposed PM is especially useful when robust solutions are required for minimising the effects of modelling errors and noise, while maximising fault sensitivity. As to the NLGA, the proposed work is the first development and robust application of the NLGA to an aircraft model in flight conditions characterised by tight-coupled longitudinal and lateral dynamics. In order to verify the robustness of the residual generators related to the previous FDI techniques, the simulation results adopt a typical aircraft reference trajectory embedding several steady-state flight conditions, such as straight flight phases and coordinated turns. Moreover, the simulations are performed in the presence of both measurement and modelling errors. Finally, extensive simulations are used for assessing the overall capabilities of the developed FDI schemes and a comparison with neural networks (NN) and unknown input Kalman filter (UIKF) diagnosis methods is performed.

International audience; We discuss interval techniques for speeding up the exact evaluation of geometric predicates and describe a C++ implementation of interval arithmetic that is strongly influenced by the rounding modes of the widely used IEEE 754 standard. Using this approach we engineer an efficient floating point filter for the computation of geometric predicates. We validate our approach experimentally, comparing it with other static, dynamic and semi-static filters.