The new algorithm proposed by the Joint International Research Institute of the Institute of Neuroscience and Medicine at the German Institute of Neuroscience and Medicine in Lich, solved the problem of limiting the simulation of brain neural networks on an E-class supercomputer—that is, the original network creation algorithm makes each processor The amount of computer memory required increases with the increase of neural networks. At the same time, after testing, it has been found that the new algorithm also improves the simulation speed of the supercomputer to some extent. The human brain is an incredibly complex organ composed of 10 billion interconnected nerve cells. Even with the help of the most powerful supercomputer, it is currently impossible to simulate the exchange of neuronal signals in a network of this size. However, recently, an international research team has taken a decisive step in the simulation of brain networks on an E-class supercomputer. This study allows the use of the same amount of computer memory to represent the larger part of the human brain. The new algorithm significantly accelerates The brain simulation effect on the existing supercomputer. The study was published in Frontiers in Neuroinformatics. "Since 2014, our software has been able to simulate the connection between about one percent of neurons in the human brain," said Markus Diesmann, director of the German Institute of Neuroscience and Medicine (INM-6). In order to achieve this feat, the software needs Gigabit supercomputers, such as Kobe's K computer and Yurich's supercomputer JUQUEEN. Diesmann has studied simulation software NEST for more than 20 years. NEST is a free, open source simulation code that is widely used by the neuroscience community and is also the core simulator of the European Brain Project. Diesmann himself led projects in the field of theoretical neuroscience and high-performance analysis and computational platforms in the European Brain-like Project. Using NEST, the behavior of each neuron in the network is represented by some mathematical equations. Future E-class computers, such as the Post-K computers planned to be built in Kobe and the JUWELS computers built in Ulrich, will be 10 to 100 times more powerful than today's high-end supercomputers. This will be the first time that researchers have computer capabilities that simulate a large-scale neural network such as the human brain. Looks like a dead end: Simulate human brain size, the processor's memory is 100 times larger than the supercomputer Although the current simulation technology enables researchers to begin research on large-scale neural networks, it also represents the end of E-class technology development. The current supercomputer consists of about 100,000 small computers called nodes. Each computer is equipped with multiple processors that perform actual calculations. "Before conducting neural network simulations, we need to create neurons and their connections virtually, which means that they need to be instantiated in the node's memory. During the simulation, neurons do not know where their target neurons are. A node. Therefore, its short electrical pulse needs to be sent to all nodes, and then each node examines which of these electrical pulses is related to the virtual neurons present on the node.†Susanne Kunkel, KTH Royal Institute of Technology, Stockholm, explains Say For the current stage, this kind of network-creation algorithm is effective because all nodes construct a specific part of their network at the same time. However, sending all electrical pulses to all nodes is not suitable for simulation on an E-class system. "In order to effectively check the correlation of each electrical pulse, we need an information bit for each processor of each neuron in the entire network. For a network with 1 billion neurons, each node is large. Part of the memory will be consumed by this information bit of the neuron,†adds Diesmann. Using petascale supercomputers (bottom left), previous simulations can simulate about 1% of neurons (neurons) in the human brain (the dark red area of ​​the brain in the left side of the picture). Although the performance of the next generation of supercomputers is 10 to 100 times greater than that of today's supercomputers, the application of previous simulations has only marginal progress in brain simulation (the dark red area of ​​the brain in the middle of the picture). Using the same amount of computer memory (bottom right corner), the new technology proposed in this study can be used to simulate more of the human brain. About 10% of our brain is equal to the size of the entire cerebral cortex (the deep red area of ​​the brain on the right side of the picture) and up to 14 billion nerve cells, which is essential for more advanced processing. Another part of the neurons is located in the cerebellum (blue part). Picture from Forschungszentrum Jülich This is the main problem encountered when simulating larger networks: The amount of computer memory required by each processor increases as the number of neuron networks increases. If you want to simulate the size of the human brain, this requires that each processor has 100 times more memory available than current supercomputers. However, this is unlikely to happen in next-generation supercomputers. The number of processors per compute node in next-generation computers may increase, but the number of memory and compute nodes per processor will remain unchanged. Breakthrough in the new algorithm: Data exchanges between neuron activity data between nodes are organized and there is no need to add bits to neurons The breakthrough in the frontiers of neuroinformatics is the creation of a new algorithm for supercomputer neuron networks. Due to this algorithm, the memory required on each node no longer increases as the network increases. At the beginning of the simulation, the new technology allows nodes to exchange information about who needs to send neuron activity data to whom. Once this knowledge is acquired, the exchange of neuron activity data between nodes is organized, so that each node only receives the information it needs, without having to add a bit for each neuron in the network. The beneficial side: Making existing supercomputers simulate faster While testing the new algorithm, the scientists presented another useful discovery, Susanne Kunkel said: "When analyzing the new algorithm, we realized that this new technology can not only complete the simulation of the E-class system, but also make the existing The supercomputer simulates faster." In fact, as memory consumption is controlled, the speed of simulation becomes the main focus of the further development of technology. For example, a large artificial neural network connected by a 5.8 trillion synapse that runs on Lich's supercomputer JUQUEEN takes 28.5 minutes to calculate a second of biological time. As the simulated data structure improves, time is reduced to 5.2 minutes. With this new technology, we can make better use of the parallelism of modern microprocessors than ever before, which will become even more important in E-class computers. "The main author of the study, Jakob Jordan, commented. The combination of E-class hardware and corresponding software promotes the study of basic brain functions such as plasticity and how to learn quickly. Said Markus Diesmann. In the next version of simulation software Nest, researchers will provide their results as free resources to the community. "We have been using NEST on the K-Computer to simulate the complex dynamics of the basal nucleus circuit of the brains of healthy people and Parkinson's patients. We are pleased to hear about the new generation of NEST, which will enable us to post on the Post-K computer. Running a whole-brain simulation to elucidate the neural mechanisms of motor control and psychological function, said Kenji Doya of the Okinawa Institute of Science and Technology (OIST). "This research will be a good example of building an international cooperation for E-class computers. It is important that we have prepared our applications and will be able to use them on the first day of the construction of these supercomputers." Mitsuhisa, Institute of Chemical Chemistry, Kobe Sato concluded in the end. switch and socket, Wall switch and socket, push switch and socket Guangdong Shunde Langzhi Trading CO., Ltd , https://www.langzhielectrical.com
March 15, 2023