An Online Neural Learning-Based Method to Enhance Scientific Lossy Compression

View the PDF file from the paper entitled Neulz: A nervous educational method on the Internet to enhance scientific pressure, by Wenqi Jia and 12 other authors

PDF HTML (experimental) view

a summary:Scientific simulation operations are widely generated by huge data collections, which represents storage challenges and I/O. The traditional pressure of pressure is struggled for more progress in the pressure rate, data quality, and the ability to adapt to various scientific data features. While deep learning solutions have been explored, their common practice of relying on large models and training limits to non -connection to the Internet with dynamic data and mathematical efficiency features. To face these challenges, we suggest Neurolz, a nervous method designed to enhance the pressure lost by integrating online learning, learning across the field, and organizing strong errors. The main innovations of Neurolz include the following: (1) INS online learning at the time of pressure with lightweight DNN models, adapting to the remaining errors without contact with non -connection mode, (2) The ability to reduce errors, and restore the fine details from the pressure errors on $ that is ignored by traditional compressors, ($ 1) $ 1 / $ 2 strict limits or relaxation $ 2 $ Times $ to get a better comprehensive quality, (4) Learning across the crossing field. Take advantage of the links between the field in scientific data to improve traditional methods. Comprehensive reviews on HPC data sets show, for example, NYX, Miranda, Hurricane, against modern compressors, neurolz. During the first five learning period, Neulz achieves a 89 % reduction in the bit, with more improvement that reaches about 94 % of the decrease in equivalent distortion, greatly outperforms the current roads, indicating the outstanding Neurolz performance in enhancing lost scientific pressure as a developed and effective solution.