Sharif Digital Repository

Speech recognition is one of the first goals of speech processing. Our goal in this thesis is to use deep learning for speech recognition. In recent years little improvement of speech recognition accuracies are reported. Deep learning is a new learning algorithm that results in improvement in many machine learning tasks. Following improvements reported in speech recognition in English language by deep learning, in this thesis we tried to improve accuracy over common and new recognition methods for Persian language.
First the overall structure of a typical speech recognition system is introduced. For this purpose, the modules of a speech recognition system are introduced. Deep multilayer... 

Fine-grained image classification is image classification where the considered classes are all sub-classes of a certain, more general class. In this setting of the problem, the classes are visually very similar to each other, such that an unskilled human cannot discriminate between them. In this case, proposed methods for the ordinary image classification problem do not obtain good classification accuracy. So proposing new methods for solving this problem is necessary. In this thesis two new methods, based on recent advances in deep learning are proposed for solving the fine-grained image classification problem. First by improving several parts of one of the recent proposed methods for this... 

With the advances in technology, Nowadays digital cameras are everywhere. As a result very large amount of images are on the web. Searching through these images intelligently and purposively is an essential need. Recently the possibility of retrieving images with some conceptual words along side of content based image retrieval has been studied in computer vision. For this purpose it’s required that for each image several words that describe its content be assigned automatically. One of the main problems for this task is semantic gap, meaning that the low level features such as color, texture,… don’t have the ability to describe the high level concepts in images which are comprehensible by... 

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... 

We construct and analyse a computational model that predicts the outcome of alternative splicing by recognizing features in RNA sequences. The computational model can be viewed as a “splicing simulator” for a range of healthy human tissues. It takes as input a pre-mRNA sequence surrounding a possibly alternatively spliced exon and estimates the inclusion level of that exon in mature RNA, after splicing occurs. The model is trained using a supervised machine learning framework where the training examples are the alternatively spliced exons, the feature vectors are derived RNA sequences near these exons, and the targets are their corresponding splicing outcomes in healthy individuals. The... 

Almost all of the network traffic classification systems use pre-defined extracted features by the experts in computer network. These features include regular expressions, port number, information in the header of different layers and statistical feature of the flow. The main problem of the traffic analysis and anomaly detection system lies in finding appropriate features. The feature extraction is a time consuming process which needs an expert to be done. It is notable that the classification of special kinds of traffic like encrypted traffic is impossible using some subset of mentioned features.The lack of integration in feature detection and classification is also another important issue... 

Cardiovascular diseases are the leading cause of death globally so early diagnosis of them is important. Many researchers focused on this field. First signs of cardiac diseases appear in the electrocardiogram signal. This signal represents the electrical activity of the heart so it’s primarily used for the detection and classification of cardiac arrhythmias. Permanent monitoring of this signal is not possible for specialists so we should do this by means of Artificial Intelligence. In this thesis, we use recurrent neural networks to classify electrocardiogram’s arrhythmias. This deep learning method, use two sources of data to learn from. The first part of data is global for everyone and the... 

Grammar induction is an important area of natural language processing. There are two general methods for recognizing the syntactic structure: constituency and dependency parsing. The unique nature of the dependency parsing, in which the word order does not affect the syntactic structure of the sentence, make it an appropriate option for parsing free-word-order languages such as the Persian. In this thesis, dependency-based methods are used to parse Persian sentences. Manual induction of the grammar is a time-consuming and tedious task. However, machine learning algorithms facilitated this task to a great deal. One of the most effective algorithms in this field is the deep neural networks... 

The extraordinary ability of the human brain to solve various problems has led scientists to simulate models of the human brain. One of these simulated models is artificial neural networks. Today, the power of artificial neural networks is not overlooked. The ability of artificial neural networks to solve various types of issues led us to use the thesis to solve some of the graph-based problems. Quite accurately, this graph-based problem is a matter of identifying the source of rumor in a network. In many graph networks, whether natural networks such as the network of neurons in the human brain or synthetic ones such as the types of social networks, it is possible that a rumor spreads across... 

News plays a very important role in stock market trading. Nowadays news from a different part of the world and about different fields can be accessed easily, and for a successful trade, it is necessary to analyze accurately and use this big data and information as soon as possible. For this reason, this thesis tries to present and study models based on Deep Learning networks and Natural Language Processing for financial news analysis and predicting stock indices movement. This research takes advantage of a language model for learning and representing news text, and beside this language model it uses deep learning networks at multiple levels to extract proper features from each news in a day... 

Lots of essential structures can be modeled as sequences and sequences can be utilized to model the structures like molecules, graphs and music notes. On the other hand, generating meaningful and new sequences is an important and practical problem in different applications. Natural language translation and drug discovery are examples of sequence generation problem. However, there are substantial challenges in sequence generation problem. Discrete spaces of the sequence and challenge of the proper objective function can be pointed out.On the other, the baseline methods suffer from issues like exposure bias between training and test time, and the ill-defined objective function. So, the... 

Nowadays, Human Action Recognition is one of the most widely used and active areas of research in computer vision. the purpose of Human Action Recognition is to label an action in a video. This field has numerous applications like human-computer interaction, video analysis, medical care, surveillance camerate, and etc. Like other subcategories of computer vision, today with the advent of deep learning networks and its development, considerable progress has been made in the accuracy and speed of the methods. The main purpose of this research is to improve human action recognition networks on RGB-D videos. In this study, three methods for action recognition using deep neural networks are... 

One of the subfields of natural language processing is sentiment analysis. Generally, sentiment analysis, analyzes the positivity, or negativity of an opinion expressed in a sentence or document. Because each person's opinions have a huge impact on the decisions of other people and businesses, automatic analysis of texts has a particular importance; on which, extensive researches have been conducted in recent years. One of the common problems in sentiment analysis of some languages, including Persian, is the lack of proper resources in them. Cross-lingual sentiment analysis is one solution to this problem. In these methods, the goal is, through using the rich resources available in a source... 

The importance of confidentiality and anonymity maintaining mechanisms are not hidden to anybody these days. With the worldwide web spreading rapidly, protecting the users' data flowing through it has become one of the most critical challenges to anonymity mechanisms. Nonetheless, machine learning algorithms have shown that they can reveal some explanatory information, even from encrypted traffic. Website fingerprinting attacks are a group of traffic analysis attacks that aim to detect the website which the monitored user has already visited. The current research takes a brief survey over website fingerprinting attacks presented in recent studies plus the defenses which took devised against... 

Deep neural networks have not progresses comparative until last decade due to computational complexity and principal challenges as gradient vanishing. Thanks to newly designed hardware architecture and great breakthroughs in 2000s leading to the solution of principal challenges, we currently face a tsunami of deep architecture utilization in various machine learning applications. Sparsity of a representation as a feature to make it more descriptive has been considered in different deep learning architectures leading to different formulations where sparsity is impose on specific representations. Due to the gradient based optimization methods for training deep architecture, smooth regularizers... 

Nowadays, due to advanced digital imaging technologies and internet accessibility to the public, the number of digital images taken has increased dramatically. Therefore, the need for automatic image enhancement techniques are felt more than ever. A successful approach in recent years is deep learning. In this thesis, besides introducing some of the works done in this field, an image enhancement system based on convolutional neural networks is presented. Our goal is to apply two existing approaches in the literature, convolutional neural networks and bilateral grid, which have been used in previous works to enhance the quality of images. The bilateral grid is a three-dimensional array which... 

In recent years, there have been so many studies in the field of machine learning, aiming to exploit the correlation among tasks. Among those, there are some types of tasks called primal and dual, which output of one is the input of the other. Dual learning is a method in which the dual and primal tasks are trained together. Many AI tasks emerge in dual form, e.g., English to Persian translation vs. Persian to English translation and image classification vs. image generation.Recently, several methods have been proposed to utilize the correlation between dual tasks. These methods can be divided into three groups of data-level, model-level, and inference-level dual learning. They have been... 

IoT security is an important step in the rapid development of its applications and ser-vices. Many security protocols have been defined to maintain security at higher communication layers, such as the transport layer and the application layer, but with the advent of quantum computers due to high-speed parallel computing, the previous methods are no longer secure and we have to establish security at lower layers such as the physical layer. The efficiency of IoT devices depends on the reliability of their message transmission. Cyber-attacks such as data injection, eavesdropping and man-in-the-middle can lead to security challenges. In this study, we propose a new deep learning based component... 

Today, estimating housing prices is very important in different societies. The reason for this growing importance can be considered the role of housing in the economic decisions and policies of any society. This estimate is made using the quantitative and qualitative characteristics of each housing and in different ways. In the past, housing prices were estimated using traditional models, but in the present century, due to easier and more access to the Internet and the development of organizations and businesses in this context, which produces a huge amount of data, the use of traditional models for this purpose is not possible. Therefore, today the use of other methods such as machine...