Parallel Programming Using CUDA Technology for Neural Networks

<p>This graduation thesis is dedicated to artificial neural networks and it&rsquo;s work acceleration by means of parallel programming.</p><p>Neural networks being a very powerful tool, allow to analyze huge volumes of data and to draw on their basis conclusions which can help experts from different areas. But one of the disadvantages of neural networks is their operating time with large volumes of data. This shortcoming can considerably be approved, using parallel algorithms at realization of neural networks in practice. Not all of the algorithms can be effectively parallelized, but neural networks algorithms possess a lot of elements which allow parallel performance.</p><p>In this work besides general idea about neural networks, the interval neural networks, Kokhonen&#39;s self-organizing maps and algorithms of their parallelization, is consideredthe CUDA technology and its application practice.</p><p>Thusfirst chapter of this diploma is devoted to neural networks. There are consideredSpheres of their application, pluses and minuses. As in detail there are considered two examples of neural networks, namely interval neural networks and Kokhonen&#39;s self-organizing maps.</p><p>The second chapter contains information on a parallelization of neural networks, is consideredthe parallelization of our examples. Also thereare considered the main methods of parallelization and is sorted the CUDA technology.</p><p>The third chapter is completely devoted to practice of application of the CUDA technology for writing of the parallel algorithms of neural networks. There are considered data of the realization of our examples and are represented the results of their work.</p>