Analysis Of Different Architectures Research Articles

Analysis of methods of classification of electronic messages based on neural network models

In the article, the creation of a mechanism for detecting and classifying messages is considered, with an assessment of how effectively different neural networks work and can recognize and classify different types of electronic messages, including phishing attacks, spam, and legitimate messages. A preliminary analysis of incoming messages has been performed, encompassing their headers, text, and other relevant attributes. For instance, in the case of emails, these attributes could be the 'subject' and 'sender' of the message. Methods for data preparation and processing have been reviewed, including text vectorization, noise removal, and normalization, to be utilized in training neural networks. Message tokenization has been performed by transforming them into a numerical format while considering the selection of features. For text messages, it is crucial to execute both tokenization and text vectorization. The model training was performed on the test data with prior splitting into two parts: 80% for training and 20% for testing. The training set is utilized for training the model, while the test set is used to evaluate its effectiveness. The peculiarity of the class structure of the data, namely the uniformity of the distribution of classes, is considered. In this case, spam occurs less frequently than legitimate messages, so class balancing techniques such as random deletion of redundant examples, upsampling, and subsampling were applied to ensure adequate model training. Optimization of network parameters was performed, by researching the optimal parameters of neural networks, such as the number and size of layers, activation functions, and optimization of hyperparameters to achieve the best performance. Hyperparameter optimization includes determining optimal settings for neural networks, such as layer size, activation functions, learning rate, and other parameters. The effectiveness was assessed by comparing the results and performance of various classification methods based on neural networks using metrics such as precision and F1-score. It was determined how well the methods can avoid misclassifications where legitimate messages are mistakenly identified as spam, and vice versa. A comparison of the methods' effectiveness in processing a large volume of messages in real time was conducted. An analysis of different architectures of neural network models was performed. Based on the analysis, it was revealed how effectively different neural network models can recognize and classify messages as spam.

Comparison of the Effectiveness of Yolov5 and yolov8 Algorithms for Detecting Human Personal Protective Equipment

The article deals with the comparative analysis of different architectures of convolutional neural networks of YOLO class: YOLOv5 and YOLOv8. Neural network was trained on a dataset consisting of more than 2,200 marked digital images. Detection and recognition of 8 classes of personal protective equipment such as medical masks, gloves, helmets, goggles, uniforms and others was performed. To evaluate the efficiency of the object detection algorithms such characteristics as accuracy metrics, learning time and memory size of the neural networks were used. The research confirms the high efficiency and accuracy of the YOLO algorithms in the detection and recognition of objects in digital still and video images. The research shows that the average accuracy of mAP object detection in YOLOv8 is 3% higher than in YOLOv5, while the performance in the new algorithm has increased by more than 50% compared with the earlier version of the neural network. The results obtained can be used to improve the object detection system in various fields, such as the automotive industry, medical and scientific research, the security field, and etc. Based on the results of the experiment, conclusions were made regarding the selection of an algorithm for the detection of human personal protective equipment. Further research in this area may be based on expanding the volume of training datasets to improve the accuracy of object recognition and evaluate the performance of algorithms on large amounts of data. Further research is planned in the area of optimizing the architecture of convolutional neural networks to improve the efficiency, speed and accuracy of object detection in digital images.

Retinal disease classification based on optical coherence tomography images using convolutional neural networks

The challenges in today’s medicine are progressively more related to the application of artificial intelligence and supervised learning techniques. Optical coherence tomography (OCT) is a noninvasive imaging technology used to obtain high-resolution cross-sectional images of the retina. The layers within the retina can be differentiated and retinal thickness can be measured to facilitate early detection and diagnosis of retinal diseases and conditions. This research paper is aimed at exploring different possibilities of applying deep learning, specifically convolutional neural networks in retinal diseases. During the research, several different architectures—AlexNet, VGG, Inception, and residual network, were evaluated, and the convolutional neural network that proved to be the most successful in the classification was based on the Inception architecture. Hyperparameter tuning was applied as the main method to find the most optimal solution. The key contributions of this research refer to the analysis of different architectures that can be applied in the classification of retinal diseases based on OCT images, as well as the evaluation of the test set obtained by comparing different models with different hyperparameters. This research yielded the best results obtained with Inception1, when training by means of the root mean square propagation optimizer with a batch size of 32, learning rate of 1e − 6, momentum of 0.99, and L2 regularization rate of 0.001. This model achieved an accuracy of 0.95528. In the conclusion of the paper, the advantages of the proposed and implemented solutions were discussed, and a proposal for further improvements was proposed.

Semantic segmentation of satellite images of airports using convolutional neural networks

The paper is devoted to the development of an effective semantic segmentation algorithm for automation of airport infrastructure labelling in RGB satellite images. This task is addressed using algorithms based on deep convolutional artificial neural networks, as they have proven themselves in a wide range of tasks, including the terrestrial imagery segmentation, where they show consistently high results. A new dataset was labelled for this particular task and a comparative analysis of different architectures and backbones was carried out. A conditional random field model (CRF) was used for postprocessing and accounting of contextual information and neighborhood of objects of different classes in order to eliminate outliers. Features of the solutions applied at all preparatory stages of the algorithm were described, including data preparation, neural network training and post-processing of the training results.

Optimizing Clifford gate generation for measurement-only topological quantum computation with Majorana zero modes

One of the main challenges for quantum computation is that while the number of gates required to perform a non-trivial quantum computation may be very large, decoherence and errors in realistic quantum architectures limit the number of physical gate operations that can be performed coherently. Therefore, an optimal mapping of the quantum algorithm into the physically available set of operations is of crucial importance. We examine this problem for a measurement-only topological quantum computer based on Majorana zero modes, where gates are performed through sequences of measurements. Such a scheme has been proposed as a practical, scalable approach to process quantum information in an array of topological qubits built using Majorana zero modes. Building on previous work that has shown that multi-qubit Clifford gates can be enacted in a topologically protected fashion in such qubit networks, we discuss methods to obtain the optimal measurement sequence for a given Clifford gate under the constraints imposed by the physical architecture, such as layout and the relative difficulty of implementing different types of measurements. Our methods also provide tools for comparative analysis of different architectures and strategies, given experimental characterizations of particular aspects of the systems under consideration. As a further non-trivial demonstration, we discuss an implementation of the surface code in Majorana-based topological qubits. We use the techniques developed here to obtain an optimized measurement sequence that implements the stabilizer measurements using only fermionic parity measurements on nearest-neighbor topological qubit islands.

Optimizing Access to Memory Pages in Software-Implemented Global Page Cache Systems

This paper is based on a dissertation “Techniques for organizing shared access to distributed memory pages in cloud computing systems” defended at the Igor Sikorsky Kyiv Polytechnic Institute in 2017. The paper describes distributed page processing in Oracle Real Application Clusters (Oracle RAC) and compares it with other well-known processing methods. The comparison includes analysis of different architectures (including shared nothing, shared disk, and replication-based architectures) in terms of SQL query processing and asserts the soundness of the distributed page approach (also known as global cache fusion) to cloud database management systems (DBMSs). As a result of analyzing the global cache fusion approach, the main drawback of Oracle RAC systems—increasing queue problem—is revealed; it causes the impossibility to process queries once their rate exceeds a certain threshold inversely proportional to the packet delivery time between nodes. To eliminate the increasing queue problem when accessing distributed pages, a new access method is proposed that introduces an additional page state—unloading state—which improves the efficiency of distributed page processing by reducing the number of transfers between nodes during hot page processing. In addition to cloud DBMSs, the proposed method can also be used in other cloud systems with page-organized distributed memory architecture.

Оптимизация доступа к страницам памяти в системах, использующих программную реализацию глобального страничного кеша

This article is based on a thesis “Techniques of organizations shared access to distributed memory pages in cloud computing systems”, defensed in Igor Sikorsky Kyiv Polytechnic Institute in 2017. The paper describes distributed pages processing in Oracle Real Application Clusters (Oracle RAC) and compares it with other known processing methods. The comparison comprises analysis of different architectures (including shared nothing, shared disk and “based on a replication” architecture) in the context of SQL query processing and asserts reasonableness of distributed pages approach (also known as Global Cache Fusion) choice for cloud DBMS. As a result researching the Global Cache Fusion approach there was revealed main drawback of Oracle RAC systems - “increasing queue problem”: impossibility process requests after intensity of requests exceeds threshold intensity, which is inversely proportional packet sent time between nodes. To eliminate “increasing queue problem” during distributed page access the new access method is proposed, which is based on the initiation of one more page state - the unloading state, which increases the efficiency of processing distributed pages by reducing the number of transfers between nodes during hot page processing. The considered method can be used not only in cloud DBMS but also in other cloud systems in a case if they use page-organized distributed memory architecture.

Performance analysis of different architectures and TCP congestion-avoidance algorithms using WMN-GA simulation system

In this paper, we evaluate the performance of two Wireless Mesh Networks (WMNs) architectures considering throughput, delay, jitter and fairness index metrics. For simulations, we used ns-3, Distributed Coordination Function (DCF) and Optimized Link State Routing (OLSR). We compare the performance of WMN for different Transmission Control Protocol (TCP): Tahoe, Reno and NewReno considering normal and uniform distributions of mesh clients by sending multiple Constant Bit Rate (CBR) flows in the network. The simulation results show that for normal and uniform distributions and both WMN architectures, the PDR values are almost the same. For Hybrid WMN, the throughput of TCP NewReno is good, but for I/B WMN, the throughput of TCP Tahoe is higher than other algorithms. For normal distribution, the delay and jitter of I/B WMN are lower compared with Hybrid WMN, while for uniform distribution, the delay and jitter of TCP NewReno are a little bit lower compared with other algorithms. The fairness index of normal distribution is higher than uniform distribution.

Program optimizations: The interplay between power, performance, and energy

We provide an analysis of the power and energy effects of program optimizations.The analysis relies on per application phase and per system component studies.We provide guidance on tradeoffs when tuning for performance, power, and energy.We identify energy and runtime correlations for optimizations on three architectures.Multi-objective optimizations require per component and application phase analysis. Practical considerations for future supercomputer designs will impose limits on both instantaneous power consumption and total energy consumption. Working within these constraints while providing the maximum possible performance, application developers will need to optimize their code for speed alongside power and energy concerns. This paper analyzes the effectiveness of several code optimizations including loop fusion, data structure transformations, and global allocations. A per component measurement and analysis of different architectures is performed, enabling the examination of code optimizations on different compute subsystems. Using an explicit hydrodynamics proxy application from the U.S. Department of Energy, LULESH, we show how code optimizations impact different computational phases of the simulation. This provides insight for simulation developers into the best optimizations to use during particular simulation compute phases when optimizing code for future supercomputing platforms. We examine and contrast both x86 and Blue Gene architectures with respect to these optimizations.

Escherichiacoli autoinducer-2 uptake network does not display hysteretic behavior but AI-2 synthesis rate controls transient bifurcation

Analysis of different architectures of quorum sensing networks has been the center of attention in recent times. The approach employs mathematical models to uncover the factors behind the dynamics. Quorum sensing networks mostly display autoregulation such as Pseudomonas aeruginosa and Vibrio cholerae. However, Escherichia coli autoinducer-2 (AI-2) synthesis does not display autoinduction (i.e. autoregulation). This and other features have raised questions about the actual function of AI-2 inside the cell. In this paper we propose a model for lsr operon regulation which explains or at least is consistent with AI-2 uptake in E. coli. The model was employed to determine the main factors that control the concentration of the signal and the uptake activation. We investigated deterministic and stochastic variants of the network model and we found no states that could lead to the typical bistability in quorum sensing systems. However, stochastic simulations predict a transient bifurcation (positively regulated by AI-2 synthesis) that could provide some advantage in adapting to new environments. LsrR inactivation was found to play a crucial role in the uptake activation compared to AI-2 synthesis, lsr transcription and AI-2 excretion. Our hypothesis is that positive regulation of the level of expression is the main factor in understanding the function of the lsr operon. This is in contrast to the conventionally held belief that the main factor is the onset of activation.