نشریه پژوهش های نظری و کاربردی هوش ماشینی
پیاپی 2 (پاییز و زمستان 1402)

With the rapid growth of video systems to monitor human behaviors, demands are increased on such systems which can detect violence events automatically. The violence detection is one of the active research area in machine learning and image processing to attract new researchers. The methods of violence detection are divided into two major categories which are traditional machine learning techniques and deep learning methods. In this article, deep learning methods have been reviewed and the variety of methods and structures of deep neural networks have been examined in this area. First, traditional and deep methods are compared with each other, and the superiority of deep methods over traditional methods is investigated from different aspects. Then, different structures of deep networks have been investigated regarding the detection of violence. Moreover, the available datasets for the analysis of violence in video are also introduced. Finally, it is discussed about the conducted research that can be useful for the development of future works.

Feature selection methods are known to be effective in improving the learning process. The purpose of a feature selection method is to identify relevant features and remove irrelevant features in order to obtain a suitable subset of features, so that the redundancy between the selected features is minimized. In multi-label data, if there is a correlation between features, it is possible that the amount of redundancy in the feature set is increased. The existence of redundancy between features along with the challenge of high dimensions of multi-label data can grow the computational calculations, decrease the accuracy and finally increase the probability of errors in the prediction and classification of multi-label data. In this article, with the aim of minimizing the redundancy of features, a multi-label feature selection algorithm is proposed considering the least squares regression model and sparse regularization. Finally, using a number of well-known multi-label data sets, the efficiency of the proposed method is verified and the results are compared with some common multi-label feature selection methods.

Despite the significant advancement of artificial intelligence methods in recent years, there is still a need for a lot of data to learn these methods. To meet this need, a new machine learning model called Few-shot learning has been proposed. One of the methods in this field is the prototypical networks approach, which is actually a combination of metric learning and meta-learning methods. In these networks, the classifier tries to generalize to these classes according to only a small number of samples of each new class. In this study, an attempt was made to propose a modified form of the prototype networks to solve the finite shot classification problem. Initially, in order to improve the performance of the prototype networks, instead of the Euclidean distance, the Mahalanobis distance was used to measure the distance between the samples. This improved the performance of these networks in classifying omniglot and miniImageNet images so that the proposed network was able to achieve 99.1% and 68.5% accuracy on these two datasets, respectively. The next section introduces a general approach that can automatically improve the architecture of convolutional neural networks using a genetic algorithm. In this research, this approach has been used specifically on omniglot datasets with the proposed primary architecture in prototype networks. Finally, by using this approach and replacing the proposed architecture with the main architecture of the prototype network, the network accuracy was improved and was able to achieve 99.5% accuracy.

The increasing popularity of the Internet has led to the dramatic growth of e-commerce. However, such activities have significant security challenges, mainly due to cyber fraud and identity theft. Therefore, checking the legitimacy of visited web pages is a very important task to secure the identity of customers and prevent phishing attacks. The use of machine learning methods, and deep learning is widely recognized as a promising solution. Research is full of studies that use machine learning and deep learning methods to detect website phishing. However, their findings depend on the data set and are far from generalizable. The two main reasons for the lack of generalization are impractical replication and lack of appropriate benchmark data sets for fair evaluation of systems. Furthermore, phishing methods are constantly evolving and the proposed models do not keep up with the rapid changes. In this article, we review the methods of identifying phishing sites from legal sites and finally reach the final conclusion.

Telegram Messenger is a suitable platform for users who are looking to buy a product or receive services online. In these messengers, it is not possible to have direct access to the providers of goods and services, and in order to request the product, one must first become a member of the related groups and channels telegram. The purpose of this study is to directly identify users of service providers using the classification of Persian messages published in Telegram. One of the problems with categorizing these messages is the large size of the feature space, which reduces accuracy and increases classification time. Feature selection methods were used to solve this problem. The proposed method of this research is based on a combination of feature selection methods based on local and global filters. In this regard, in the first step, using the most widely used methods for selecting local and global filter feature, related features are selected. In the second step, a combination of local and global filtering methods is used to identify better features and increase classification accuracy. The innovation of this research is in using the combined methods of feature selection for automatic classification of Telegram Persian messages, in order to identify the users of the service provider. The proposed method, while reducing the number of features and selecting related features, improves the performance of classification and Identification of service providers.

A person's emotional state changes due to the emotional stimulus he is experiencing. During a conversation, these changes are noticeably reflected in his voice and tone of speaking as well as in the linguistic components of his speech. On the receiving side, the listener can recognize the emotional content of the received message and also, predict the emotional behavior associated with it. The human nervous and cognitive system is able to perform this process efficiently and effortlessly despite all the complexities involved in computationally defining a model for this process. In other words, it recognizes the emotional content in the speech and predict a set of related behaviors for the speaker according to that recognized emotional state. This issue is crucial in human-machine interaction studies. Here, a model based on Bidirectional Associative Memory is presented, which can predict the corresponding emotional behavior based on the emotional state. The results of the simulation show that this model is able to perform the process of mapping bidirectionally between emotional states and the set of behaviors related to emotional states and is robust against some degrees of information incompleteness and the presence of some level of noise.

This paper outlines the optimization of the two first parts of the three major components of the  scene descriptions of the geometrical shapes, namely (1) fuzzy logic scheme, (2) an integrated algebraic and logical reasoning, and (3) the machine learning technique. After arguing the need for using fuzzy expressions in spatial reasoning, the integration of approximate references into spatial reasoning using absolute measurements is outlined. The integration here means that the satisfiability of a spatial expression including possibly fuzzy one is conducted by both logical and algebraic reasoning. Then, the implementation of spatial expression evaluation is briefly described. The paper ends by the conclusion and the problems to be studied. The reasoning algorithm of this system not only uses the one-piece system, algebraic method and first-order logical method, but by introducing fuzzy spatial propositions and how to integrate them, as well as classical propositions in this system, it has a high power compared to classical systems.

This research aims to enhance the accuracy of credit assignment for rules in a learning classifier system using Markov Reinforcement Learning for predicting the secondary structure of proteins. Markov Reinforcement Learning has replaced the Bucket Brigade algorithm in the learning classifier system. The Protein Data Bank dataset is utilized to train the system, specifically the protein 4L1W with 5741 samples, where 70% is used for training and 30% for testing purposes. Following the system's training, a set of valuable classifiers (rules) is generated, which will be employed in the testing phase to predict the protein's secondary structure. The experimental results demonstrate an improvement in the accuracy of the Markov Reinforcement Learning classifier system, both in the Learning classifier system and the eXtended classifier system. Through Markov Reinforcement Learning, the credit assignment to each rule is enhanced, resulting in an accuracy improvement of 82.5% for the Learning classifier system and 85% for the eXtended classifier system.

Analyzing individuals' emotions from the content of social media through text, speech, and images is necessary for various types of applications and purposes. Most recent research studies in the field of sentiment analysis have focused on textual data. However, social media users share more images and videos compared to text. In other words, images are the most effective way to convey emotions to others. Therefore, focusing on the development of a sentiment analysis model based on images in social media is important. In this article, we will use the DenseNet-121 transfer learning model to analyze emotions based on images. To implement this approach, we will utilize the images available in the Image Sentiment dataset. This dataset includes internet links to images along with their emotional polarities. Based on the obtained results, the accuracy of the proposed model in this article is 89%, which, compared to previous work in the field of visual sentiment analysis, shows a 5% to 10% improvement.

In the field of knowledge of philosophical schools, each person's thinking is different according to the type of attitude he can have regarding different schools. Recognizing the author's attitude and its similarity to each of the philosophical schools has always been one of the important issues in the field of humanities. In this article, a method based on deep learning is proposed to distinguish philosophical schools from text. In the proposed method, first the texts are normalized and redundant and meaningless words are removed. After the normalization stage, the text is broken into sentences and words, and then using the fasttext library, each word is converted into a numerical vector, after that, the features of the texts are extracted using the designed network, and finally, the system has learned and is ready to extract data, and by giving a new sentence, its similarity to each school is expressed. Based on the evaluation, the accuracy of the proposed method is 94%.

The morphology of the ventricles is used in studies of hydrocephalus, schizophrenia, tumors, trauma, Alzheimer's disease, Parkinson's disease, aging and atrophy to diagnose neurological diseases such as stroke, dementia and Huntington's disease. In this study, an automatic method for morphometric analysis of cerebral ventricles is proposed. Segmentation of cerebral ventricles is an important step to detect the landmarks of cerebral ventricles. Furthermore, the initial estimation of the cerebro-ventricular area can be effective in the proper segmentation of the ventricles. In this regard, Fuzzy C-Means Clustering (FCM) optimized with Harris-Hawk Optimization (HHO) and Aggregate Channel Features (ACF) are used. In order to measure the linear indices of cerebral ventricles, including Evans Index, Bicaudate Ratio, Bicaudate-Frontal Index, Bicaudate-Temporal Index, and Huckman Number, locating a number of landmarks on MRI images is required. This process is based on the geometric features of cerebral ventricles and the use of Hough transformation. The implementation results demonstrate that the proposed algorithm has the best performance with precision 90%, sensitivity 82%, specificity 99%, peak signal-to-noise ratio 77/11, dice similarity coefficient 86%, Jaccard index 75% and contour matching score 92% in the segmentation of cerebral ventricles among other compared methods. Additionally, the results show that the measurement accuracy of the proposed algorithm in the mentioned morphometric linear indices is 97%, 74%, 77%, 76% and 92% respectively.

Extractive summarization of text is an essential technique in natural language processing, which helps to produce compact versions of text by extracting the most important sentences. Since the task of shortening and summarizing a text document is time-consuming and exhausting, an automatic system for creating these short versions of the text seems necessary. In extractive summarization, sentences that contain useful and relevant information are usually selected for the final summary. In order to identify these sentences, there are different algorithms, the performance and summary created by each one is different based on the type and scope of the text and the size of the required summary. In this article, a method called Sa-TRB is presented, which is derived from two algorithms, TextRank and BERT, and in addition to using these two methods, it also uses the common sentences created by other algorithms to achieve high accuracy in selection. Have final summary sentences. The most important criterion for evaluating the performance of algorithms is the quality of their final summary, so the more the final summary created by these algorithms is similar to the summary created by humans, the better the quality of the created summary is. ROUGE criteria have been used to obtain the size of this similarity. Finally, by conducting experiments on the cnn-dailymail dataset with different sizes of summaries, it is shown that the proposed method, by increasing the size of the required summaries, despite the decrease in the recall criterion, has accuracy, score and, as a result, higher quality of the final summaries. So, in the last two tests, the score of the proposed method has reached 24.68 and 23.34%, which is almost one percent better than the best tested methods.