Search academic texts

Information × Registration Number 0212U008145, 0111U005982 , R & D reports Title Parallel Grid-aware Library for Neural Networks Training popup.stage_title Head Turchenko Volodymyr, Registration Date 20-11-2012 Organization Ternopol National Economic University popup.description2 Object of research - accelerating the execution of computationally-intensive algorithms training artificial neural networks (NN). Purpose - to develop improved methods of training artificial neural network for heterogeneous parallel computing systems in the computational grid systems that provide high performance parallelization learning processes and the development of grid-based library programs for parallel study of artificial neural network. Methods - the methods of the theory of neural systems and networks, computer systems design methods, methods for evaluating the computational complexity of algorithms, optimization techniques, methods of computer simulation methods of experiment planning. The study - new methods of parallel training artificial neural network, including multilayer perceptron, recirculation neural network and the neural network with radial-basis activation function on parallel computer systems with different architectures, including the parallel computers with shared memory and computing clusters distributed architecture, broker resources for heterogeneous grid systems that will keep the high efficiency of parallelization of the developed algorithm learning by identifying Pareto-optimal computing resources (according to the criteria of minimum execution time, maximum efficiency and minimum prices parallelization computing resource), which is appropriate to perform a specific parallel algorithm library routines in the programming language C, which implements the above methods and broker resources and results of application of the developed library for solving practical problems. Scientific novelty lies in the creation of new methods for parallelization of artificial neural networks based on the theory of group learning, which, unlike existing ones, can significantly improve acceleration and parallelization efficiency on parallel computers with shared memory computing on clusters with distributed architecture and on GRID computing systems. Product Description popup.authors Турченко Володимир Олександрович popup.nrat_date 2020-04-02 Close