Theoretical physics and social networks work may seem as different as apples and giraffes, but Julian Kates-Harbeck says there are deep connections.
His social network modeling research at Harvard University involves complex problems, with multiple interacting components. Kates-Harbeck savors the challenge of distilling these systems to “the key things that are happening that cause the behavior that we’re observing.” He hopes to define the system’s fundamental properties in order to make predictive insights.
“This is why theoretical physics has always been one of my favorite subjects, because that’s exactly what you’re taught,” says Kates-Harbeck, a Department of Energy Computational Science Graduate Fellowship recipient. “You’re taught to remove all the unnecessary complexities and stick with the necessary complexities required to explain what you’re observing.” Reducing a social network to its interacting components “allows you to tackle many problems with one unified paradigm.”
Although the rapid spread of an idea or concept is called “going viral,” major differences between that phenomenon and a disease spreading make previous theories ineffective, Kates-Harbeck says. The probability of infection via an epidemic depends on how many times someone is exposed to previously infected individuals. With social contagion, in contrast, people don’t independently observe what neighbors on the network do but instead “form an aggregate opinion and then, based on that aggregate opinion of what everyone around (them) is doing,” decide whether to adopt the idea.
“This introduces a key nonlinearity that changes everything,” he says. Researchers can’t reduce social phenomena to “interactions between two individuals. You have to consider the entire neighborhood of each individual.”
Kates-Harbeck says his research with Michael Desai sticks largely to the theoretical and avoids direct applications. In light of recent controversies over social network misuse, “we want to separate ourselves from that and just build a theory of how this works.” It’s possible, however, that what they learn could help predict which ideas will spread quickly and provide information to ensure good ideas spread and bad ones (such as fiction disguised as news) don’t.
The Harvard researchers don’t yet have that predictive capability, which would depend on a host of factors. “But we do have the kind of theoretical insight that allows us to understand which features of a network make an idea easier to spread, which features of an idea make it more viral.”
Kates-Harbeck has done his network research on Odyssey, a Harvard high-performance computing (HPC) cluster. He’s used much bigger machines for a project started during his 2016 practicum at DOE’s Princeton Plasma Physics Laboratory (PPPL). That includes the United States’ most powerful supercomputer, the Titan system at Oak Ridge National Laboratory.
Kates-Harbeck worked with PPPL’s William Tang, who researches fusion energy reactors called tokamaks – donut-shaped chambers in which powerful magnets contain plasmas hotter than the center of the sun. Under these conditions, hydrogen nuclei in the plasma fuse, releasing tremendous energy.
But the plasma can escape its magnetic container and damage the reactor walls. Researchers want to analyze sensor data on the fly to predict when these so-called disruptions are likely to occur and take steps to stop or limit them.
As detailed in this DEIXIS Online article, Kates-Harbeck is lead architect of the Fusion Recurrent Neural Network, an artificial intelligence model to forecast disruptions. The team trained the algorithm on data from JET, the world’s largest tokamak, and from DIII-D, the only operating U.S. machine, and then showed that its predictive power surpasses previous methods. Tests using thousands of processors on Titan and other machines demonstrated that the algorithm runs faster in direct proportion to the number of processors employed.
Kates-Harbeck still collaborates with the PPPL team while working on his doctoral research. “The fact that I have that freedom, to take the time and do this, is something I wouldn’t have without this fellowship. I’m really grateful for that.”
Kates-Harbeck aims to graduate in 2019 and will most likely pursue an academic post after serving a postdoctoral fellowship. So far, however, he admits his knowledge of social networks probably has little influence on how he uses forums like Twitter.
“I’m an average user of social networks for my generation,” he says, and he tries to be aware of the information pitfalls they harbor. “But I’m just a human like everyone else, so I’m probably just as prone to biases” as anyone. “Just like with insights into psychology or behavioral economics, knowing about it doesn’t necessarily make you immune to it.”
Image caption: Visualization of the community structure in a section of a popular online social media network. Each color represents a separate community. Complex social contagions, such as new ideas, depend on these densely connected communities to incubate and spread wide across the network. Data from the Stanford Network Analysis Project. Image courtesy of Julian Kates-Harbeck.