Computer Science Journal (AGH University of Science and Technology, Krakow)
Not a member yet
476 research outputs found
Sort by
USING SPLITTER ORDERING HEURISTICS TO IMPROVE BISIMULATION IN PROBABILISTIC MODEL CHECKING
Full text linkModel checking is used to verify computer-based and cyber-physical systems, but faces challenges due to state space explosion. Bisimulation minimization reduces states in transition systems, easing this issue. Probabilistic bisimulation further simplifies models with stochastic behaviors. Recent techniques aim to improve time complexity of iterative methods in computing probabilistic bisimulation for stochastic systems with nondeterministic behaviors. In this paper, we propose several techniques to accelerate iterative processes to partition the state space of a given probabilistic model to its bisimulation classes. The first technique applies two ordering heuristics for choosing splitter blocks. The second technique uses hash tables to reduce the running time and the average time complexity of the standard iterative method. The proposed approaches are implemented and run on several conventional case studies and reduce the running time by one order of magnitude on average
Microservices, a Definition Analyzed by ßMACH
Full text linkMicroservices are an answer to various scalability challenges. They enable building large and complex systems by scaling the number of services, and development teams. Microservices allow self-management and agile processes such as Scrum. Nevertheless, a minimal management process should also be defined and documented. Documentation must be easily understandable and applicable to the developing teams to foster efficiency. We propose to use the ßMACH method, a software management guidance. The result is a minimal and systematic description. Based on the definition obtained using ßMACH, we can state that the service and team isolation is essential for a scalable microservice system. In addition, we introduce ßMACH and show how to document a software management process. The documentation is easy to understand by software developers and interesting for software engineers
BIELIK 7B V0.1: POLISH LANGUAGE MODEL - DEVELOPMENT, INSIGHTS, AND EVALUATION
Full text linkWe introduce Bielik 7B v0.1, a 7-billion-parameter generative text model for Polish language processing. Trained on curated Polish corpora, this model addresses key challenges in language model development through innovative techniques. These include Weighted Instruction Cross-Entropy Loss, which balances the learning of different instruction types, and Adaptive Learning Rate, which dynamically adjusts the learning rate based on training progress. To evaluate performance, we created the Open PL LLM Leaderboard and Polish MT-Bench, novel frameworks assessing various NLP tasks and conversational abilities. Bielik 7B v0.1 demonstrates significant improvements, achieving a 9 percentage point increase in average score compared to Mistral-7B-v0.1 on the RAG Reader task. It also excels in the Polish MT-Bench, particularly in Reasoning (6.15/10) and Role-playing (7.83/10) categories. This model represents a substantial advancement in Polish language AI, offering a powerful tool for diverse linguistic applications and setting new benchmarks in the field
The Power of Intelligence Emerging from Swarms
Full text linkSwarm intelligence (SI) is a field of study that seeks to understand and model collective behaviors observed in natural social systems, such as ant colonies, bee hives, termite mounds, flocks of birds or schools of fish. The central principle of SI is that complex intelligent behaviors can emerge from the interactions of large numbers of simple individual entities, without any centralized control or monitoring. SI researchers aim to uncover the underlying principles and mechanisms behind this SI, with the aim of applying these concepts to solve complex problems in areas such as optimization, robotics, transport, IT, etc. As the field continues to evolve, SI is expected to have an increasingly significant impact on our understanding of biological systems and our ability to design intelligent systems capable of adapting and thriving in complex environments and dynamic. This article aims to introduce the reader to the field of SI, presenting its fundamental concepts, key principles, existing applications, and prospective future developments
Eye Disease Segmentation using Hybrid Neural Encoder Decoder based Unet Hybrid Inception
Full text linkDiabetic retinopathy (DR) is one of the major causes of vision problems worldwide. With proper treatment, early diagnosis of DR can prevent the progression of the disease. In this paper, we present a combinative method using U-Net with a modified Inception architecture for the diagnosis of both the diseases. The proposed method is based on deep neural architecture formalising encoder decoder modelling with convolutional architectures namely Inception and Residual Connection. The performance of the proposed model was validated on the IDRid 2019 contest dataset. Experiments demonstrate that the modified Inception deep feature extractor improves DR classification with a classification accuracy of 99.34% in IDRid across classes with comparison to Resnet. The paper Benchmark tests the dataset with proposed model of Hybrid Dense-ED-UHI: Encoder Decoder based U-Net Hybrid Inception model with 15 fold cross validation. The paper in details discusses the various metrics of the proposed model with various visualisation and multifield validations
IMPROVING PET SCANNER TIME-OF-FLIGHTRESOLUTION USING ADDITIONAL PROMPT PHOTON
Full text linkPositronium Imaging requires two classes of events: double-coincidences originated from pair ofback-to-back annihilation photons and triple-coincidences comprised with two annihilation photonsand one additional prompt photon. The standard reconstruction of the emission position along theline-of-response of triple-coincidence event is the same as in the case of double-coincidence event;an information introduced by the high-energetic prompt photon is ignored. In this study, we proposeto extend the reconstruction of position of triple-coincidence event by taking into account the timeand position of prompt photon. We incorporate the knowledge about the positronium lifetime distributionand discuss the limitations of the method based on the simulation data. We highlight that theuncertainty of the estimate provided by prompt photon alone is much higher than the standard deviationestimated based on two annihilation photons. We finally demonstrate the extent of resolutionimprovement that can be obtained when estimated using three photons
CHARACTERISTIC SKY BACKGROUND FEATURES AROUND GALAXY MERGERS
Full text linkIn the context of finding galaxy merger in large-scale surveys, we applied MachineLearning algorithms that, instead of using the images as it is the currentstandard, made used of flux measurements. Training multiple NNs using aclass-balanced dataset of mergers and non-mergers Sloan Digital Sky Survey,we found that the sky background error parameters could provide a validation92.64 ± 0.15 % accuracy of and a training accuracy of 92.36 ± 0.21 %.Moreover, analysing the NN identifications led us to find that a simple decisiondiagram using the sky error for two flux filters is enough to get a 91.59 % accuracy.By understanding how the galaxies vary along the diagram, and trying toparametrize the methodology in the deeper images of the Hyper Suprime-Cam,we are currently trying to define and generalize this sky error-based methodology
FL-MEC: FEDERATED LEARNING FOR NETWORK TRAFFIC CLASSIFICATION ON THE NETWORK EDGE
Full text linkNowadays, two technological trends, Federated Learning (FL) and Edge Computing (EC), are becoming more and more important and influential. FL is a distributed machine learning strategy that allows learning on distributed data. It primarily allows performing learning operations close to the user, where we can gather data. This approach lies in the EC domain, where the main goal is to move computation closer to the end user (e.g., from the centralized cloud). In our work, we apply the FL and EC in the context of network flow classification. We achieved an accuracy of 0.957 with the FL model, compared to 0.924 for the best local model. We achieved these results thanks to the federated averaging performed on neural network layers. To verify our approach, we executed all our experiments on a virtualized environment that emulates existing mid-scale EC network infrastructure, including limitations related to resource constraints on edge nodes
DEVELOPING ARTIFICIAL INTELLIGENCE IN THE CLOUD: THE AI INFN PLATFORM
Full text linkThe INFN CSN5-funded project AI INFN (“Artificial Intelligence at INFN”) aims to promote ML and AI adoption within INFN by providing comprehensive support, including state of-the-art hardware and cloud-native solutions within INFN Cloud. This facilitates efficient sharing of hardware accelerators without hindering the institute’s diverse research activities. AI INFN advances from a Virtual-Machine-based model to a flexible Kubernetes-based platform, offering features such as JWT-based authentication, JupyterHub multitenant interface, distributed file system, customizable conda environments, and specialized monitoring and accounting systems. It also enables virtual nodes in the cluster, offloading computing payloads to remote resources through the Virtual Kubelet technology, with InterLink as provider. This setup can manage workflows across various providers and hardware types, which is crucial for scientific use cases that require dedicated infrastructures for different parts of the workload. Results of initial tests to validate its production applicability, emerging case studies and integration scenarios are presented
PRELIMINARY STUDY ON ARTIFICIAL INTELLIGENCE METHODS FOR CYBERSECURITY THREAT DETECTION INCOMPUTER NETWORKS BASED ON RAWDATA PACKETS
Full text linkMost of the intrusion detection methods in computer networks are based ontraffic flow characteristics. However, this approach may not fully exploit thepotential of deep learning algorithms to directly extract features and patternsfrom raw packets. Moreover, it impedes real-time monitoring due to the neces-sity of waiting for the processing pipeline to complete and introduces depen-dencies on additional software components.In this paper, we investigate deep learning methodologies capable of de-tecting attacks in real-time directly from raw packet data within network traffic.Our investigation utilizes the CIC IDS-2017 dataset, which includes both benigntraffic and prevalent real-world attacks, providing a comprehensive foundationfor our research