Search

Deep neural networks have been successful in various domains, such as computer vision and natural language processing. On the other hand, researchers have discovered a vulnerability of convolutional neural networks to the samples with imperceptible perturbations, also known as, adversarial perturbations. It has been observed that adversarial perturbations can alter the predictions of a deep model. One of the most common approaches to increase the robustness of deep models is adversarial training. However, adversarial training often suffers from a degradation in generalization performance. In this study, orthogonal regularization is used along with adversarial training to facilitate both generalizability and adversarial robustness in deep models. According to the experiments on MNIST and CIFAR10 datasets, imposing orthogonality in weights improves both the generalization performance and adversarial robustness.

Increasing demand in SATCOM systems has led to the development of Satcom on the Move (SOTM) antennas. Performance, light weight, low cost, and high data rate are the required parameters in the designing of SOTM antenna. The designer has to make a suitable system with respect to these design parameters. With this aim, this paper proposes a 4 axes antenna capable of 3 axes stabilization to compensate for the disturbance. With adding the position controller, the antenna can point and track the satellite on the mobile platforms. Due to the singularity problem in this type of antenna, the 3 axes stabilized antenna system is used in this paper. The kinematic equations and Jacobian operator are derived for calculation of the necessary axis feedback and look angles. The Jacobian operator with MEMS sensor which is a cost-effective solution. Furthermore, we propose PI-based stabilization for the stabilization controller and PID based pointing for the position controller. To verify the proposed algorithms, different disturbances are applied to the system.

In this paper, we propose a 4 axis antenna which is capable of tracking satellites, through the precise real-time control of azimuth, cross-level, elevation and polarization angles. Due to the well-known gimbal lock problem in 2-3 axis antenna systems, our design has a redundant degree of freedom, and this redundancy is utilised to solve the issue of inverse kinematics around singularities. In the paper, we derive the kinematics of the 4 DOF redundant orientation mechanism and propose a method to avoid the gimbal lock issue. Also, utilising satellite position data (from TLE Database) and GPS localisation, we derive the necessary kinematics equations to position the antenna towards any visible satellite from any position on earth. Furthermore, we propose a PI-based cascade controller for tracking the satellites. To verify our analytical results we present experimental results on a 4 axis antenna test platform that we have built and compare our results with the feedback from a satellite in the form of guidance RF signals. The proposed system can find and track satellites with a 1-degree error without satellite signal feedback.

In cancer treatments photodynamic therapy, because of being a minimally invasive method and having no significant harmful side effect to the patient. This method begins with the administration of photosensitizer material prior to treatment in the tissue, which reacts chemically to the specific wavelength light. Dosimeter calculations determine the time required for photosensitizer the body to remain only in the tumor. At the end of this period, the tumor tissue becomes eligible for targeted treatment. In order to be able to create the thermal effect on the tissue with the laser beam, a laser with a constant wavelength must be used. When photosensitizer provides sufficient energy in the target tissue exited with specific wavelength light energy in the cell, the cancer cells lead to necrosis. In order to complete the treatment successfully, Photodynamic therapy system's laser diode must have a constant temperature and constant output power so that it can generate constant wavelengths. Since the laser diode temperature is directly related to the wavelength and the output power, control of the temperature at a single point is an essential issue.. In the study, the selected laser diode need to be stabilized at a temperature of 15 °C so that it could operate steadily at a wavelength of 635 nm which can excite the ALA- 5. The thermoelectric controller (TEC), the thermistor directly connected to the laser diode, detects the temperature of the diode. The sensed temperature value is applied to the (Dspic30f4013) microcontroller's ADC circuit via an analog circuit that will produce 5V for the maximum value and 0V for the minimum value of the thermistor. This feedback from the thermistor with PID control on the microprocessor determines the PWM on and off time ratio. Cooling control is provided instantaneously by PWM signal applied on the thermoelectric material.

Data augmentation is a widely adopted technique for avoiding overfitting when training deep neural networks. However, this approach requires domain-specific knowledge and is often limited to a fixed set of hard-coded transformations. Recently, several works proposed to use generative models for generating semantically meaningful perturbations to train a classifier. However, because accurate encoding and decoding are critical, these methods, which use architectures that approximate the latent-variable inference, remained limited to pilot studies on small datasets. Exploiting the exactly reversible encoder-decoder structure of normalizing flows, we perform on-manifold perturbations in the latent space to define fully unsupervised data augmentations. We demonstrate that such perturbations match the performance of advanced data augmentation techniques -- reaching 96.6% test accuracy for CIFAR-10 using ResNet-18 and outperform existing methods, particularly in low data regimes -- yielding 10--25% relative improvement of test accuracy from classical training. We find that our latent adversarial perturbations adaptive to the classifier throughout its training are most effective, yielding the first test accuracy improvement results on real-world datasets -- CIFAR-10/100 -- via latent-space perturbations., In Proceedings of the IEEE International Conference on Computer Vision