All of this is supposed to be a reliable way to control a computer or its software. But that will hardly work, who n They try to classify human intelligence algorithms with classical methods. They need January 2019 Calendar South Africa to take advantage of data-driven approaches such as deep learning. Impressive successes have been achieved here, especially when compared with classical logic-based systems. “Does not the hardware have its limits?” Keuper: Machines learn a wealth of information-driven data. The data mass required for this leads to the hitherto customary use of highendshared systems.
January 2019 Calendar South Africa
memory Multi-GPU systems for the calculation of a midsize PARADIGIN CHANGE: DEEP LEARNING SUPERCOMPUTERNDespite a multitude of trend-setting developments, machine learning has to undergo a paradigm shift – because the current need for data and computing power to calculate learning models is growing faster than the quality of the calculated solutions. Janis Keuper from the Fraunhofer Institute for Industrial and Industrial Mathematics ITWM.
FRAUNHOFER ITWM23Models needs several days. The maximum achievable computing power in a local system can hardly keep up with the demands of machine learning. That’s why machine learning algorithms are facing a fundamental change: we need to switch to a distributed computation on heterogeneous high-performance computers, HPC for short.This requirement focuses on learning per se? Keuper: Exactly. The computational effort always relates to the actual learning process at HPC.
If something is learned, the software could run behind on your phone. But learning, for example, requires the automotive industry, which requires a variety of algorithms for its autonomous vehicles: for the driver himself, to recognize traffic signs or pedestrians, and so on. These models must be pre-trained, increasingly on mainframes. At the FraunhoferITWM we are working on methods for better and simpler use of HPC for deep learning. A central problem here is the question of how the vast amount of data can actually “pass through” the actual billing process, as it were. We are working in the range of 10 to 15 or even 10 to 20 arithmetic operations.
Therefore, we try to get the necessary data transport under control on almost all levels – from mathematical models that calculate the amount of data down to the communication protocols down to the hardware with which we transport This is also the subject of projects such as the »High Performance Deep Learning Framework«, which we carry out on behalf of the Federal Ministry of Education and Research (BMBF).
www.itwm.fraunhofer.de / mlDr. Janis Keuper24In industrial practice, things are much more difficult than in the classroom: Whether in the chemical industry, the pharmaceutical industry, or the food industry: necessary separations of different substances are often considerably more complicated than the scholastic separation of solids and liquids with maximum adjustable fluency – besides, tens of thousands of liters have to be processed per hour.
“As a rule, there are several phases of separation here. When processing raw milk, for example, milk, cream and various solid substances have to be separated cleanly, “explains Sebastianvon Enzberg of the Fraunhofer Institute for Design Technology Mechatronics.” If separators are not set exactly and adjusted continuously, January 2019 Calendar South Africa quality losses and product damage can result. “The correct setting of the separators was geared to a long time Flair and hearing of the respective centrifuge experts – usually the manufacturer or machine operator.