The oncoming data tsunami could overwhelm information-rich research projects from small neutrino to explosive supernova as well as puzzles in the brain.
When LIGO raises a gravitational wave from a distant black hole and neutron stars, it starts time to capture the first possible light with them. The information gathered from the sensors that control the activity of the brain exceeds the capacity of the computer. Fractured beam data from the larger Hadron Collide (LHC) will soon exceed 1 pt.
To address this imminent data block, a team of researchers from nine universities, including MIT, led by MIT, received $ 15 million in funding to establish the Institute for Accelerated AI Algorithms for Data-Driven Discovery (A3D3). From MIT, the research team includes Assistant Professor of Physics, Philip Harris, who serves as Deputy Director of the A3D3 Institute. Song Han, Assistant Professor of Electrical Engineering and Computer Science, Man A3D3’s Associate PI; And Eric Katsavonidis, senior research scientist at the MIT Kavley Institute of Astrophysics and Space Research.
A3D3, supported by Big Five Ideas and advanced cybersecurity bureaus in this five-year data revolution, focuses on three rich environments: multi-messenger astrophysics, high-energy particle physics and brain imaging neuroscience. By enriching AI algorithms with a new processor, A3D3 seeks to accelerate AI algorithms to solve complex problems in conflict physics, neutron physics, astronomy, gravitational physics, computer science, and neuroscience.
“I am delighted with the new institute’s opportunity to conduct research on nuclear and particle physics,” said Bolelaw Wisles, director of the Laboratory of Nuclear Science. “Modern particle searchers generate a huge amount of data, and we are looking for unusual rare signatures.
A3D3 seeds were planted in 2017, when Harris and his colleagues Fermilab and CERN decided to integrate real-time AI algorithms. By LHC To process unreliable data rates. In an email exchange with Han, Harris team developed an HLS4ML compiler that could run in nanoseconds.
“The fastest process we knew before the development of HLS4ML was approximately one millisecond in AI, probably a little faster,” Harris said. “We recognize that all AI algorithms are designed to solve very slow problems, such as image and audio recognition. .
A few months later, Harris presented his research at a meeting of the Faculty of Physics, which caught the attention of Katsavonidis. In Building 7 Coffee, they discussed the waves of gravity by combining the Harris FPGA with the use of Katsavounisis machine learning. FPGAs and other types of processors, such as graphics processing units (GPUs), accelerate AI algorithms to quickly analyze large amounts of data.
“I worked with the first FPGAs on the market in the early ’90s and saw in the past how high-tech physics and data acquisition was a revolution. Reminds Katsavonidis. “Ever since I joined LIGO 20 years ago, the ability to break gravity data has been on my mind.”
They received their first support two years ago and joined the University of Washington’s Shi-Chih Hussein. The team published about 40 articles on this subject, inspired the fast machine, and built the team for about 50 researchers and said, “General work has begun. Industry on how to explore previously undiscovered AI territory, ”says Harris. “Basically, we started this without any financial support. We have been receiving small subsidies for various projects over the years. A3D3 represents our first major donation to support this effort.
“What makes A3D3 unique is the technical boundary exploration for MIT. Rob Simco, director of the MIT Caveley Institute of Astrophysics and Space Research and director of physics at Francis Friedman, said, “We are living in a time when experiments create a flood of data. Time analysis.
Huge information from a large Hadron Collage
With a data rate of more than 500 terabytes per second, LHC produces more information than any other scientific device on earth. Future aggregate data will soon exceed 1 petit per second, the largest data in the world.
“Using AI, the A3D3 aims to perform advanced analyzes such as unconventional detection and particle reconstruction at collisions up to 40 million times per second,” he said.
The goal is to find a way to identify a few conflicts to complete the picture, which can show new forces out of 3.2 billion collisions per second, explain how dark matter is formed, and how basic forces interact with matter. To process all this information, you need a custom computer system that can translate conflict information into extremely low latency.
“The challenge of running this in real time with all 100 therapies per second is very difficult and requires a complete overhaul of how we formulate and implement AI algorithms,” Harris said. “The greater the demand for information, the greater the increase in data rates.
The brain and the universe
Thanks to advances in techniques such as medical imaging and electromagnetic implants, neuroscience is also gathering vast amounts of information on how the brain’s neural networks respond to stimuli and how motor information works. A3D3 plans to develop and implement high-speed and low-latency AI algorithms to process, organize, and analyze real-time neural data to test brain function to enable new experiments and treatments.
With Multi-Messenger Astrophysics (MMA), A3D3 aims to quickly identify astronomical phenomena by efficiently processing information from gravitational waves, gamma rays, and neutrinos taken by telescopes and finders.
A3D3 researchers include project leaders from the University of Washington, Calcutta, Duke University, Purdue University, UC San Diego, the University of Illinois at Urbana-Champaign, the University of Minnesota and the University of Wisconsin-Madison. Includes Icecube and DUNE neutronos research and astronomy on the Zwiki Transit Facility and prepares in-depth workshops and boot camps on how students and researchers can contribute to the framework and expand the use of faster AI strategies.
“We have reached a stage where the development of the research network is making a difference, in terms of event size and astronomical discoveries and ultimately discoveries,” says Katsavonidis. “Fast and ‘efficient’ is the only way to fight the ‘weak’ and ‘stupid’ in the universe and use the best of our data. A3D3 on the one hand is going to bring product-scale AI to gravity-wave physics and multi-messenger astronomy; But on the other hand, we want to move beyond our immediate domains to accelerate AI across the country into information-based education.
.