Sharif Digital Repository

Grammar induction is a basic step in natural language processing. Based on the volume of information that is used by different methods, we can distinguish three types of grammar induction method: supervised, unsupervised, and semi-supervised. Supervised and semisupervised methods require large tree banks, which may not currently exist for many languages. Accordingly, many researchers have focused on unsupervised methods. Unsupervised Data Oriented Parsing (UDOP) is currently the state of the art in unsupervised grammar induction. In this paper, we show that the performance of UDOP in free word order languages such as Persian is inferior to that of fixed order languages such as English. We... 

This paper presents a simple yet efficient method for an anomaly-based Intrusion Detection System (IDS). In reality, IDSs can be defined as a one-class classification system, where the normal traffic is the target class. The high diversity of network attacks in addition to the need for generalization, motivate us to propose a semi-supervised method. Inspired by the successes of Generative Adversarial Networks (GANs) for training deep models in semi-unsupervised setting, we have proposed an end-to-end deep architecture for IDS. The proposed architecture is composed of two deep networks, each of which trained by competing with each other to understand the underlying concept of the normal... 

Supervised speech emotion recognition requires a large number of labeled samples that limit its use in practice. Due to easy access to unlabeled samples, a new semi-supervised method based on auto-encoders is proposed in this paper for speech emotion recognition. The proposed method performed the classification operation by extracting the information contained in unlabeled samples and combining it with the information in labeled samples. In addition, it employed maximum mean discrepancy cost function to reduce the distribution difference when the labeled and unlabeled samples were gathered from different datasets. Experimental results obtained on different emotional speech datasets... 

In this paper, we introduce a novel method for human activity recognition that benefits from the structure and sequential properties of the test data as well as the training data. In the training phase, we obtain a fraction of data labels at constant time intervals and use them in a semi-supervised graph-based method for recognizing the user's activities. We use label propagation on a k-nearest neighbor graph to calculate the probability of association of the unlabeled data to each class in this phase. Then we use these probabilities to train an HMM in a way that each of its hidden states corresponds to one class of activity. These probabilities are used to learn the transition probabilities... 

Anatomical image segmentation is one of the foundations for medical planning. Recently, convolutional neural networks (CNN) have achieved much success in segmenting volumetric (3D) images when a large number of fully annotated 3D samples are available. However, rarely a volumetric medical image dataset containing a sufficient number of segmented 3D images is accessible since providing manual segmentation masks is monotonous and time-consuming. Thus, to alleviate the burden of manual annotation, we attempt to effectively train a 3D CNN using a sparse annotation where ground truth on just one 2D slice of the axial axis of each training 3D image is available. To tackle this problem, we propose... 

Multilayer perceptron (MLP) neural networks have been used extensively for estimation/regression of parameters. Moreover, recent studies have shown that learning algorithms of MLP which are based on Gaussian function are more accurate. In this paper, the mass attenuation coefficient (MAC) of gamma radiation for light-weight materials (e.g. O-8), mid-weight materials (e.g. Al-13), and heavy-weight materials (e.g. Pb-82) is modelled using Gaussian function based regularization of MLP (i.e. Bayesian regularization (BR)) and by a modular estimator. The results are compared with the Reference results. To show better performance of the utilized algorithm, the results of the different supervised... 

In this study, the estimation of the uranium price as one of the most important factors affecting the fuel cost of nuclear power plants (NPPs) is investigated. Supervised learning algorithms, especially, multilayer feed-forward neural network (FFNN) are used extensively for parameters estimation. Similar to other supervised methods, FFNN can suffer from overfitting (i.e. imbalance between memorization and generalization). In this study, different regularization techniques of FFNN are discussed and the most appropriate regularization technique (i.e. Bayesian regularization) is selected for estimation of the uranium price. The different methods including different learning algorithms of FFNN,... 

The motivation of this study is development of a technique to construct energy spectrum of higher price/high resolution detectors (e.g. NaI (Tl)) using spectrum of lower price/low resolution detectors (e.g. NE102). Since there is no explicit mathematical model between these type of detectors (i.e. organic and inorganic scintillator detectors), it is necessary to utilize model-free methods. Construction of mapping function to generate spectrum of inorganic scintillator using spectrum of organic scintillator can be done by supervised model-free methods. Different supervised learning methods including localized neural networks, statistical methods, feed-forward neural networks, and conditional... 

Anatomical image segmentation is one of the foundations for medical planning. Recently, convolutional neural networks (CNN) have achieved much success in segmenting volumetric (3D) images when a large number of fully annotated 3D samples are available. However, rarely a volumetric medical image dataset containing a sufficient number of segmented 3D images is accessible since providing manual segmentation masks is monotonous and time-consuming. Thus, to alleviate the burden of manual annotation, we attempt to effectively train a 3D CNN using a sparse annotation where ground truth on just one 2D slice of the axial axis of each training 3D image is available. To tackle this problem, we propose...