Sharif Digital Repository

Data representation plays an important role in machine learning and the performance of machine learning algorithms for instance, in supervised learnings (e.g. classifcation), and unsupervised ones (e.g. image denoising), are heavily influenced by the input applied to them. Regarding the fact that data usually lacks the desirable quality, efforts are always made to make a more desirable representation of data to be used as input to machine learning algorithms. Among many different representation of data, sparse data representation preserves much more information about data while it is simpler than data. We proposed a new stacked sparse autoencoder by imposing power two of smooth L0 norm of... 

Markov Random Fields are suitable and applicable tools for modeling dependency of data dimensions; But since time complexity of parameter learning for these models is exponential with respect to data dimension size, density estimation using these models is restricted in action. In the other hands, with growing use of neural networks in many problems a class named autoregressive networks were applied for density estimation. Although the learning time of neural network parameters is not very low, there are many efforts for parameter learning acceleration. In this thesis Markov Random Fields structural properties are used in autoregressive networks. In our proposed method the hypothesis space... 

There are a variety of multivariate industrial processes in real world problems. It seems to be necessary to control them through strong tools such as control charts. One of the state-of-the-art methods to monitor processes is neural network. Neural networks are data processing systems inspired by human brain and they are capable of processing data with a variety of small processors working in parallel forming an integrated network to solve a problem. Chaotic models, one of the states of being out of control, are deterministic non-linear models which have extremely complex behavior under determined assumption. Researches have shown neural networks have excellent performance in such systems.... 

Deep learning methods have become more popular in the past years. These methods use complex network architectures to model rich, hierarchical datasets. Although most of the research has been centered around Discriminative models, however, recently a lot of research is focused on Deep Generative Models. Two of the pioneering models in this field are Generative Adversarial Networks and Variational Auto-Encoders. In addition, knowing the structure of data helps models to search in a narrower hypothesis space. Most of the structure in datasets are models using Probabilistic Graphical Models. Using this structural information, one can achieve better parameter estimations. In the case of... 

High dynamic range (HDR) images provide more realistic experience in displaying real-world scenes than conventional low dynamic range (LDR) images by providing much more detailed luminance information; However, most imaging content is still available in low dynamic range. Inverse tonemapping is known as the problem of inferring an HDR image from a single-exposure LDR image in which the lost data caused by saturation of bright parts and quantization must be reconstructed.To address this problem, in this thesis, two fully-automatic architectures based on convolutional neural networks, are proposed. Both these architectures utilize a number of convolutional auto-encoders as... 

One of the most important alternative technologies in the field of security monitoring is millimeter-wave imaging technology, which is a good alternative both for performance and cost-effectiveness. Conventional security monitoring techniques use optical images or metal detectors to control people in crowded, sensitive places, but with the help of electromagnetic waves, these technologies can be obtained. This way, metal and non-metallic objects hidden inside clothing, bags or shoes can also be detected that are not identifiable by conventional security-control methods. This thesis examines the implementation of an improved algorithm for automatic detection of objects in millimeter-wave... 

Exoskeletons, or wearable robots, are electromechanical devices that have become the focus of academic and industrial research in recent years, and their applications in power augmentation have increased. One of the most important challenges of these applications is controlling the robot and synchronizing it with the user. The close interaction of the robot with the user and the change of movement pattern between different users and different gait cycles show the importance of estimating the user's movement intention, but the need for online method of estimating the movement intention and the complexity of accurate dynamic modeling has caused researchers to use reinforcement learning (RL) in... 

Entanglement is one of the key resources of quantum information science which makes identification of entangled states essential to a wide range of quantum technologies and phenomena.This problem is however both computationally and experimentally challenging.Here we use autoencoder neural networks to find semi-optimal measurements for detection of entangled states. We show that it is possible to find high-performance entanglement detectors with as few as three measurements. Also, with the complete information of the state, we develop a neural network that can identify all two-qubits entangled states almost perfectly.This result paves the way for automatic development of efficient... 

Today, the use of deep learning algorithms in the design of communication systems has received much attention. One of these areas is the partial or total design of these systems using deep networks. The overall design of a communication system using deep networks allows for global optimization and can provide better performance in cases where classical methods have suboptimal performance without significantly increasing the computational load. In this research, a comprehensive architecture for designing communication systems based on Auto-encoder neural networks is presented. This architecture has the same functionality as classical systems, considering all parts of these systems including... 

Identifying anomalous events is one of the vital topics in research as it often leads to the detection of actionable and critical information such as intrusions, faults, and system failures. With its importance, there has been a substantial body of work for network anomaly detection using supervised and unsupervised machine learning techniques with their own strengths and weaknesses. In this work, we take advantage of both worlds of unsupervised and supervised learning methods. The basic process model we present in this paper includes (i) clustering the training data set to create referential labels, (ii) building a supervised learning model with the automatically produced labels, and (iii)... 

In the signal transmission cycle in a telecommunication network, there are two coding blocks, which can be realized in two ways. In the traditional way, source coding is first used to remove redundancies and compress information. Then, channel coding is used to transmit on the telecommunication channel and deal with noise and other destructive factors of the channel. In other words, source coding and channel coding are done separately. In contrast, signal transmission can be done by joint source¬channel coding. According to Shannon separation theorem and with the fulfillment of the conditions mentioned in the theorem, the method of separation of source and channel coding in a point to point... 

Autoencoder, as an essential part of many anomaly detection methods, is lacking flexibility on normal data in complex datasets. U-Net is proved to be effective for this purpose but overfits on the training data if trained by just using reconstruction error similar to other AE-based frameworks. Puzzle-solving, as a pretext task of self-supervised learning (SSL) methods, has earlier proved its ability in learning semantically meaningful features. We show that training U-Nets based on this task is an effective remedy that prevents overfitting and facilitates learning beyond pixel-level features. Shortcut solutions, however,are a big challenge in SSL tasks, including jigsaw puzzles. We propose... 

One of the most widely used protocols used on the Internet is the SSL protocol, which is also used in many applications to exchange information between the server and the user. Therefore, the analysis of this traffic can help decision makers in many analyses. In this thesis, we are going to present a mapping for feature vectors extracted from SSL traffic that will lead to improving the performance of machine learning algorithms.In this treatise, three methods for learning mapping are proposed, all of which are based on deep learning. The first method is to use a simple self-encoder for map learning that tries to learn a compact map from the input feature vector.The second method is the... 

Parkinson's disease is a progressive and severe neurodegenerative disorder in the central nervous system that is common in adults and, to some extent, in young people. One way to control and improve this disease is through rehabilitation, specifically in the form of rehabilitation activities. Due to the difficulty of travelling to and from the clinic, systems have been developed today that, using inertial sensors, provide the possibility of remotely monitoring this type of treatment up to 95%. In this regard, a system named SEPANTA has been developed at the Djavad Mowafaghian Research Center to diagnose and monitor the rehabilitation activities performed by patients with Parkinson's disease.... 

With the spread of threats against industrial control systems, preserving the security of these systems faces serious challenges. On the other hand, with the increase of communication between industrial control networks and external communication networks, the entry points of these networks have also increased and this exposes them to IP network threats. Beside that, traditional attacks on these systems, which generally occur by infiltrating the internal network, are also constantly changing and becoming more complex. These attacks mainly have a phase of hiding the attack from the monitoring systems, which eliminates the possibility of identifying the attacker's operations to a great extent... 

The rapid evolution of artificial intelligence (AI) technologies has led to the widespread adoption of AI systems in diverse research and industrial fields. Deep neural networks, at the forefront of AI's power, demonstrate high performance by leveraging large volumes of training data. However, acquiring such vast amounts of data requires collaboration among individual data owners, who may have concerns about privacy. To address these concerns, various privacy-preserving methodologies have been proposed. These methodologies share a common goal of striking a balance between preserving privacy and maintaining data utility. This study aims to explore and analyze these privacy protection... 

Today, the additive manufacturing process is one of the most important manufacturing processes. In the additive manufacturing process, the mass of the parts is directly related to the printing duration, in addition to the costs of consumables. On the other hand, the use of lighter parts in industries, such as the aviation, is one of the important requirements of those industries, and topology optimization is needed to reduce the consumables and the mass of the parts. Also, since in the optimal design of engineering parts, more than one design objective is usually considered, the multi-objective optimization of topology is of great importance. The Moving Morphable Components is one of the new...