The human brain contains a vast expanse of unmapped territory. An adult brain measures only about 1,300 cubic centimeters — less than 80 cubic inches — but it holds 86 billion neurons. Mapping the largely uncharted 100 trillion connections among them is called connectomics, and computer scientists at Argonne National Laboratory are processing huge datasets of human brain imagery to map interneuron links.
Argonne’s Nicola Ferrier, who recently co-chaired an SC24 session on extreme-scale analysis and visualization, and ANL colleague Thomas Uram are collaborating with Harvard neuroscientist Jeff Lichtman. The project leans on artificial intelligence and machine learning analyses of 3D electron microscopy images to better understand the shapes of neurons and how they connect.
The researchers have received a Department of Energy (DOE) Innovative and Novel Computational Impact on Theory and Experiment (INCITE) award of 250,000 hours on Argonne’s supercomputer Aurora, the world’s second-fastest high-performance machine and the second to achieve exascale. Aurora can perform more than a million trillion operations per second. To prepare for the INCITE award, the team is using ALCF’s Polaris supercomputer and pre-Aurora hardware, under an Aurora Early Science Project (ESP) award.
Uram compares connectomics today to having no road map while trying to understand how people travel from place to place. “If we have a map, then we can think about things like how neurons communicate with each other,” says Uram, the data sciences and workflows team lead in the Argonne Leadership Computing Facility. “How does connectivity make us who we are, or produce our ability to learn? Or how is it affected by learning, diseases or aging? Having the map should enable us to study these sorts of questions.”
Completing the map will require electron microscopes and supercomputers that are more advanced than any currently available, as well as computational tools to complement those future technologies. To develop those tools as they map, Ferrier and Uram will process datasets from nanoscale-resolution electron microscope images generated by Lichtman’s lab. Each dataset will comprise millions of gigabytes, an undertaking that will require Aurora’s more than 60,000 Intel GPUs.
“Our ability to pursue these problems depends on availability of the world’s largest supercomputers; with these machines, we are able to chase questions that would not be answerable otherwise,” Uram says.
Connectomics to date has mostly involved organisms with small brains, such as worms, flies or mice. A mouse brain contains about 70 million neurons. To map their connections, researchers take a one-centimeter-sided cube-shaped sample, slice it into thin sections for electron microscopy, then computationally assemble the tile-like images to create a 3D image of the sample.
‘You can train machine learning to do amazing things for many tasks involving images if you have enough data and the right model.’
“There are computer vision steps in some of the preliminary alignment and stitching,” says Ferrier, a senior computer scientist at Argonne. But machine learning completely outclasses computer vision methods for later processing steps.
“There’s no comparison,” Ferrier says. “You can train machine learning to do amazing things for many tasks involving images if you have enough data and the right model. The earlier work, trying to use computer vision, met with some success, but it just can’t compete with the various machine learning approaches.”
The INCITE project will process data from humans rather than mice. Neuroscientist Lichtman, a pioneer in high-speed electron microscopy, will provide the images
Large connectomics projects typically involve thousands or tens of thousands of human hours spent correcting AI-generated data errors.
“That’s a significant challenge,” Uram says. “While AI is capable of identifying individual neurons in a sample of tissue with great accuracy, it is still imperfect, and a small error rate can produce many mistakes in a large body of data.”
The INCITE project will train AI models to trace the shapes of millions of neurons and their connections. The results need to be corrected by humans because, like ChatGPT, the models generate results that range from highly valuable to ridiculous. “That’s just the nature of AI,” Uram says, though “the connectomics community continues to pursue more accurate AI models and post-processing to ameliorate the task” of human fact-checkers.
The trained model then will be applied to huge volumes of data on Aurora. Preparations for that moment began with the selection of the related Aurora ESP. During the ESP, Ferrier and Uram demonstrated computational readiness for the INCITE award by adapting their codes to operate efficiently and at large scales on Aurora.
As with their connectomics research, the Argonne pair often collaborate with scientists in other fields. Ferrier, for example, has applied her computation skills to medical systems and problems in manufacturing and plant-phenotyping. Uram has done likewise in high-energy physics and cosmology.
“One thing that’s exciting about being a computer scientist is you get exposure to different fields,” Ferrier says. “I always view that as a bonus.”
Collaborating with people in other fields often involves learning unfamiliar terminology. “It can be challenging to have to get to the level where you’re conversing appropriately. But I embrace that challenge,” she says.
Ferrier and Uram also work with Argonne’s Advanced Photon Source (APS), a powerful DOE X-ray generator that enables researchers to probe and improve the materials and chemical processes that affect nearly every aspect of life. Connectomics methods can easily transfer to brain imagery produced at the APS, and possibly also to data generated in experiments at the APS for many different scientific fields.
A better understanding of the brain may also generate benefits outside of biology. Assuming that neural networks are biologically inspired — and some experts question that notion — detailed brain maps might also help improve AI models.
In the far future, insights gleaned from connectomics may also benefit neuromorphic computing endeavors, Ferrier notes. Today’s supercomputers require tens of megawatts. The human brain, which contains a similar number of computing elements, operates on 20 watts. That’s a million times less power.
“If we could harvest at least some of that fraction to make computers more efficient, that would be a huge game changer for us in power consumption,” Uram says.