نشریه مهندسی برق و مهندسی کامپیوتر ایران
سال بیستم شماره 3 (پیاپی 73، پاییز 1401)

Today, with the increase in data production volume, attention to machine learning algorithms to extract knowledge from raw data has increased. Raw data usually has redundant or irrelevant features that affect the performance of learning algorithms. Feature selection algorithms are used to improve efficiency and reduce the computational cost of machine learning algorithms. A variety of methods for selecting features are provided. Among the feature selection methods are evolutionary algorithms that have been considered because of their global optimization power. Many evolutionary algorithms have been proposed to solve the feature selection problem, most of which have focused on the target space. The problem space can also provide vital information for solving the feature selection problem. Since evolutionary algorithms suffer from the pain of not leaving the local optimal point, it is necessary to provide an effective mechanism for leaving the local optimal point. This paper uses the PSO evolutionary algorithm with a multi-objective function. In the proposed algorithm, a new mutation method that uses the particle feature score is proposed along with elitism to exit the local optimal points. The proposed algorithm is tested on different datasets and examined with existing algorithms. The simulation results show that the proposed method has an error reduction of 20%, 11%, 85%, and 7% in the Isolet, Musk, Madelon, and Arrhythmia datasets, respectively, compared to the new RFPSOFS method.

Due to the increasing speed of information production and the need to convert information into knowledge, old machine learning methods are no longer responsive. When using classifications with the old machine learning methods, especially the use of inherently lazy classifications such as the k-nearest neighbor (KNN) method, the operation of classifying large data sets is very slow. Nearest Neighborhood is a popular method of data classification due to its simplicity and practical accuracy. The proposed method is based on sorting the training data feature vectors in a binary search tree to expedite the classification of big data using the nearest neighbor method. This is done by finding the approximate two farthest local data in each tree node. These two data are used as a criterion for dividing the data in the current node into two groups. The data set in each node is assigned to the left and right child of the current node based on their similarity to the two data. The results of several experiments performed on different data sets from the UCI repository show a good degree of accuracy due to the low execution time of the proposed method.

Quantum computing is a new method of information processing that is based on the concepts of quantum mechanics and leads to strange and powerful events in the quantum field. The logic synthesis of quantum circuits refers to the process of converting a given quantum gate into a set of gates that can be implemented in quantum technologies. The most famous logic synthesis methods are CSD and QSD. The main goal of this study is to present a multi-objective logical synthesis method combining the above two methods in the quantum circuit model with the aim of optimizing the evaluation criteria. In this proposed method, the solution space is created from different combinations of CSD and QSD decomposition methods. The created solution space is a space with a very large exponential size. Then, using a bottom-up approach of multi-objective dynamic programming, a method is presented to search only a part of the entire solution space to find circuits with the optimal Pareto costs. The obtained results show that this method creates a balance between the evaluation criteria and produces many optimal Pareto solutions that can be selected according to different quantum technologies.

Using the unlabeled data in the semi-supervised learning can significantly improve the accuracy of supervised classification. But in some cases, it may dramatically reduce the accuracy of the classification. The reason of such degradation is incorrect labeling of unlabeled data. In this article, we propose the method for high confidence labeling of unlabeled data. The base classifier in the proposed algorithm is the support vector machine. In this method, the labeling is performed only on the set of the unlabeled data that is closer to the decision boundary from the threshold. This data is called informative data. the adding informative data to the training set has a great effect to achieve the optimal decision boundary if the predicted label is correctly. The Epsilon- neighborhood Algorithm (DBSCAN) is used to discover the labeling structure in the data space. The comparative experiments on the UCI dataset show that the proposed method outperforms than some of the previous work to achieve greater accuracy of the self-training semi-supervised classification.

In recent years, optical neural networks have received a lot of attention due to their high speed and low power consumption. However, these networks still have many limitations. One of these limitations is implementing their nonlinear layer. In this paper, the implementation of nonlinear unit for an optical convolutional neural network is investigated, so that using this nonlinear unit, we can realize an all-optical convolutional neural network with the same accuracy as the electrical networks, while providing higher speed and lower power consumption. In this regard, first of all, different methods of implementing optical nonlinear unit are reviewed. Then, the impact of utilizing saturable absorber, as the nonlinear unit in different layers of CNN, on the network’s accuracy is investigated, and finally, a new and simple method is proposed to preserve the accuracy of the optical neural networks utilizing saturable absorber as the nonlinear activating function.

Session-based recommender systems predict the next behavior or interest of the user based on user behavior and interactions in a session, and suggest appropriate items to the user accordingly. Recent studies to make recommendations have focused mainly on the information of the current session and ignore the information of the user's previous sessions. In this paper, a personalized session-based recommender model with self-attention networks is proposed, which uses the user's previous recent sessions in addition to the current session. The proposed model uses self-attention networks (SANs) to learn the global dependencies among all session items. First, SAN is trained based on anonymous sessions. Then for each user, the sequences of the current session and previous sessions are given to the network separately, and by weighted combining the ranking results from each session, the final recommended items are obtained. The proposed model is tested and evaluated on real-world Reddit dataset in two criteria of accuracy and mean reciprocal rank. Comparing the results of the proposed model with previous approaches indicates the ability and effectiveness of the proposed model in providing more accurate recommendations.

Content-based image retrieval, which is also known as query based on image content, is one of the sub-branches of machine vision, which is used to organize and recognize the content of digital images using visual features. This technology automatically searches the images similar to the query image from huge image database and it provides the most similar images to the users by directly extracting visual features from image data; not keywords and textual annotations. Therefore, in this paper, a method is proposed that utilizes wavelet transformation and combining features with color histogram to reduce the semantic gap between low-level visual features and high-level meanings of images. In this regard, the final output will be presented using the feature extraction method from the input images. In the next step, when the query images are given to the system by the target user, the most similar images are retrieved by using semi-supervised learning that results from the combination of clustering and classification based on frequent patterns mining. The experimental results show that the proposed system has provided the highest level of effectiveness compared to other methods.