The new algorithm will allow to simulate neural connections of the whole human brain on future exaflop supercomputers
In 2013, to simulate 1 second of 1% of the human brain, it took 40 minutes on a cluster of 82,944 processors of a 10-petaflops K computer . Scientists have tried to repeat the work of 1.73 billion nerve cells and 10.4 trillion synapses connecting them, each of which took 24 bytes.
The power of supercomputers of the new generation will be calculated in exaflops, but with existing software solutions, only 10% of brain activity will be enough to simulate it. An international team of scientists has created an algorithm that can change this and present up to 100% of the activity. For the first time, researchers will have the power of existing computers to simulate a neural network across the entire human brain.
The new simulation algorithm is designed to help simulate 100 billion interconnected neurons on exaflops supercomputers, that is, “digitize” neurons on the scale of a whole brain. It is based on the NEST neuro- stimulation tool with which the Human Brain Project works. Researchers from several countries took part in creating the algorithm: the Juelich Research Center (Germany), the Norwegian University of Natural Sciences and Technologies NMBU, the Rhine-Westphalian Technical University of Aachen (Germany), the Institute of Physical and Chemical Research RIKEN (Japan), the Royal Institute of Technology ( Sweden).
')
“Before it is possible to simulate a neural network, neurons and their connections must be created in a virtual environment,” says study author Susanne Kunkel. In the process of simulating 100 thousand neurons represented by the same number of network nodes, the excitation wave from the neuron should be sent to all 100 thousand nodes. Each node is equipped with processors for performing calculations. When receiving a signal, they are used, among other things, to check its relevance - whether this impulse refers to it.
Each signal corresponds to one bit of information per processor for each neuron network. For a network of billion neurons, a large amount of node memory will be spent on checking relevancy. As the network grows, this number grows, so to increase the activity share in the simulation from 1% to 100%, it will require an increase in computer memory by a factor of 100. Since 2014, neural networks have been simulated using NEST on K computer and JUGENE with a capacity of 10 petaflops and 222 teraflops, respectively. Future supercomputers will become more powerful, the number of processors for computing will grow, but the amount of memory per processor will remain the same.
This is where the new NEST algorithm comes in handy. At the beginning of the simulation, it will allow the nodes to determine which data on neural activity should be sent and where. Once this information is clear, it will be possible to send data in an address order. This eliminates the need to process data bits for each neuron of the network.
Software today is able to represent about 1% of the activity of neurons in the cerebral cortex on supercomputers whose power is calculated in petaflops. The next generation of exam-scale supercomputers will increase this figure to 10%. The new algorithm will give the opportunity to simulate up to 100% of brain activity using the same amount of memory at the same capacity.
The researchers are confident that after optimizing the use of memory, the main task will be to increase the speed of the simulation. For example, today a simulation of 0.52 billion neurons combined by 5.8 trillion synapses on the JUQUEEN supercomputer takes 28.5 minutes per 1 second of biological time. The new algorithm for the calculations of scientists will reduce this time to 5.2 minutes. “The combination of the test-grade hardware and the new software will allow you to explore the fundamental functions of the brain, such as plasticity and learning, which take place within minutes of biological time,” says Markus Diesman. In 2013, the scientist argued that the computational power necessary to imitate the work of the brain will become available after 2020 — in two years we will be able to find out if he is right.
In one of the next software releases from the Neural Simulation Technology Initiative, the code will be freely available to the entire community. The algorithm will speed up simulations on existing petaflops supercomputers, the developers say.
Kenji Doya, an engineer and neuroscientist at the Okinawa Institute of Science and Technology, could be one of the first users of the new algorithm: “We worked with NEST to simulate the complex dynamics of the basal gangs of a healthy person and with Parkinson’s disease on the K computer. We are pleased to hear about the new version of NEST, which will enable the launch of a simulation of a whole brain on next-generation computers, which will make it possible to clarify the mechanisms of motor and mental functions. ”
The power of supercomputers of the new generation will be calculated in exaflops, but with existing software solutions, only 10% of brain activity will be enough to simulate it. An international team of scientists has created an algorithm that can change this and present up to 100% of the activity. For the first time, researchers will have the power of existing computers to simulate a neural network across the entire human brain.
The new simulation algorithm is designed to help simulate 100 billion interconnected neurons on exaflops supercomputers, that is, “digitize” neurons on the scale of a whole brain. It is based on the NEST neuro- stimulation tool with which the Human Brain Project works. Researchers from several countries took part in creating the algorithm: the Juelich Research Center (Germany), the Norwegian University of Natural Sciences and Technologies NMBU, the Rhine-Westphalian Technical University of Aachen (Germany), the Institute of Physical and Chemical Research RIKEN (Japan), the Royal Institute of Technology ( Sweden).
')
“Before it is possible to simulate a neural network, neurons and their connections must be created in a virtual environment,” says study author Susanne Kunkel. In the process of simulating 100 thousand neurons represented by the same number of network nodes, the excitation wave from the neuron should be sent to all 100 thousand nodes. Each node is equipped with processors for performing calculations. When receiving a signal, they are used, among other things, to check its relevance - whether this impulse refers to it.
Each signal corresponds to one bit of information per processor for each neuron network. For a network of billion neurons, a large amount of node memory will be spent on checking relevancy. As the network grows, this number grows, so to increase the activity share in the simulation from 1% to 100%, it will require an increase in computer memory by a factor of 100. Since 2014, neural networks have been simulated using NEST on K computer and JUGENE with a capacity of 10 petaflops and 222 teraflops, respectively. Future supercomputers will become more powerful, the number of processors for computing will grow, but the amount of memory per processor will remain the same.
This is where the new NEST algorithm comes in handy. At the beginning of the simulation, it will allow the nodes to determine which data on neural activity should be sent and where. Once this information is clear, it will be possible to send data in an address order. This eliminates the need to process data bits for each neuron of the network.
Software today is able to represent about 1% of the activity of neurons in the cerebral cortex on supercomputers whose power is calculated in petaflops. The next generation of exam-scale supercomputers will increase this figure to 10%. The new algorithm will give the opportunity to simulate up to 100% of brain activity using the same amount of memory at the same capacity.
The researchers are confident that after optimizing the use of memory, the main task will be to increase the speed of the simulation. For example, today a simulation of 0.52 billion neurons combined by 5.8 trillion synapses on the JUQUEEN supercomputer takes 28.5 minutes per 1 second of biological time. The new algorithm for the calculations of scientists will reduce this time to 5.2 minutes. “The combination of the test-grade hardware and the new software will allow you to explore the fundamental functions of the brain, such as plasticity and learning, which take place within minutes of biological time,” says Markus Diesman. In 2013, the scientist argued that the computational power necessary to imitate the work of the brain will become available after 2020 — in two years we will be able to find out if he is right.
In one of the next software releases from the Neural Simulation Technology Initiative, the code will be freely available to the entire community. The algorithm will speed up simulations on existing petaflops supercomputers, the developers say.
Kenji Doya, an engineer and neuroscientist at the Okinawa Institute of Science and Technology, could be one of the first users of the new algorithm: “We worked with NEST to simulate the complex dynamics of the basal gangs of a healthy person and with Parkinson’s disease on the K computer. We are pleased to hear about the new version of NEST, which will enable the launch of a simulation of a whole brain on next-generation computers, which will make it possible to clarify the mechanisms of motor and mental functions. ”
Source: https://habr.com/ru/post/374367/
All Articles