Building a better brain

Can you build a smarter computer by imitating the human brain? Catherine Schuman of ORNL thinks so.

Guest post by: Alisa Alering, Science Node
First published on Science Node on 3rd January 2018

Brain glue. Neuromorphic computing systems are inspired by human neurobiology like this astrocytic glial cell (glial = glue). Courtesy Gerry Shaw. (CC BY-SA 3.0)

The human brain weighs three pounds and is made up of more than 100 billion nerve cells that allow us to remember birthdaysrecognize and evade dangercompose symphoniesbuild bridges, and design super-smart machines to take over the tasks we find too difficult, too dirty, or too boring.

But even though scientists admit that there’s still a lot they don’t know about how the brain works, Catherine Schuman of Oak Ridge National Laboratory (ORNL)thinks that machine brains might work better if they were designed to be more like human brains.

Rather than looking to the seventy year history of neural networks constructed on Von Neumann architecture, researchers like Schuman instead take inspiration from current discoveries in neuroscience to build a very different kind of artificial intelligence.

These neuromorphic computers are designed to be massively parallel, constructed from many simple computational elements connected to mimic the neurons and synapses in the human brain.

Most recently, Schuman has been intrigued by the astrocytic glial cells that insulate neural pathways and allow signals to travel fasters along certain routes.

“We’re looking deeper into the way the biological brain functions and trying to take more inspiration from biology,” says Schuman. “For example, the fascinating thing about astrocytes is that they actually change and develop as you learn, so they could be important in building learning systems.”

Survival of the fittest

Aptitude test. Scientists are experimenting to discover what applications neuromorphic systems excel at. Courtesy Catherine Schuman, et al.

Neuromorphic systems excel at data analysis tasks. For example, high energy physicists work with tremendous amounts of data, and a neuromorphic system can help them to analyze that data in a quick and energy-efficient way.

Brain-based neural networks are also very good at real-time anomaly detections—useful in cybersecurity situations to flag intrusions and other attacks. Neuromorphics also shine in robotic control tasks, such as operating a vehicle.

But making a brain that can teach itself is a big project. As a software developer, Schuman writes algorithms and builds models and simulations. But she collaborates with hardware developers in a co-design process.

“I go to hardware designers with a request, and they come back and say, ‘We can’t do exactly what you want but what if we tried this?’ and we just iterate on that process, looking at the restrictions of the hardware and trying to optimize energy efficiency and the size of the system,” says Schuman.

The neuromorphic field is extremely collaborative. Neuroscientists, computer scientists, electrical engineers, and material scientists—we’re all working together to build these new types of systems. ~ Catherine Schuman

Schuman can run a neuromorphic simulation on a traditional computer, but she turns to ORNL’s Titan supercomputer when she wants to simulate many different systems at once. The genetic algorithm she used to build the neuromorphic systems benefits from massive parallelization.

“We have something like 18,600 compute nodes at Oak Ridge, and I have used 18,000 of them at once to build a program to control a robot that we then deploy into a neuromorphic device,” says Schuman.

Her genetic algorithms are just what they sound like: survival of the fittest.

Schuman begins with a random set of potential solutions for a particular problem. At the beginning, all of the solutions perform poorly. But after training, Schuman evaluates the whole population, and the ones that do the best get to reproduce.

“We take two pieces of two networks and put them together to produce the child population. Then we iterate that process to get better and better performing networks until eventually you get a network that works,” says Schuman. “Nature and evolution have come up with some really smart things, so why not emulate it?”

Evolution of a career

Schuman came late to the computing field. Her own education didn’t expose her to programming in middle or even high school. It was actually a rivalry with her older brother, who is a software developer, that finally spurred her into the computational arena.

“My attitude was, if he can do it, then I can do it, too,” says Schuman.

But now that she’s found the field, she’s clearly discovered the ideal outlet for her talents.

“I want to think big thoughts and push the field forward,” says Schuman. “I get bored if I’m not developing new systems and coming up with new ideas.”

From user to creator. Children are avid users of computers but Schuman wants to encourage them to think about what–and who–is behind the screen. Courtesy US Air National Guard.

Schuman is passionate about getting more women involved in computing. She encourages everyone in the field to make an effort to get out to local schools and make sure that kids know that computing is an option.

Despite some negative experiences being a woman in a male-dominated career, Schuman believes the trials—for her, at least—have been worth it, and she wants to pass her excitement on to the next generation.

“Kids are interacting with computers every day, but they don’t necessarily think about what’s underneath—they don’t think about who’s building the apps, video games, and software they interact with every day,” says Schuman. “I like to tell them, ‘This is what I do for a job, and it’s awesome.’”

About Catherine Schuman, ORNL

  • Catherine D. Schuman (Katie) is a Liane Russell Early Career Fellow in the Computational Data Analytics group at Oak Ridge National Laboratory. Katie received her doctorate in computer science in 2015 from the University of Tennessee, where she completed her dissertation on the use of evolutionary algorithms to train spiking neural networks for neuromorphic systems. She is continuing her study of models and algorithms for neuromorphic computing as part of her fellowship at ORNL. Katie has co-authored over 20 publications in neuromorphic computing, presented her work at several conferences and workshops, and holds two patents. Katie is also a joint faculty member at the University of Tennessee (UT), where she, along with four professors at UT, leads the TENNLab neuromorphic research team.

About the author: Alisa Alering, Managing Editor, Science Node

  • Originally trained as a librarian, Alisa loves tracking down the science behind her stories and learning something new about technology every day. With previous experience as a freelance writer and photo editor, she has held positions at Indiana University Press, PBS, and Google and earned degrees from Penn State and Indiana University. She particularly enjoys writing about women and diversity in technology, digital humanities, and the intersection of science and the natural world.

About Science Node

  • Science Node is an online magazine that connects the global research community, exploring how tech works and showing why it matters to our everyday lives. sciencenode.org