Your browser is unsupported

We recommend using the latest version of IE11, Edge, Chrome, Firefox or Safari.

Mesrob Ohannessian awarded National Science Foundation CAREER grant

Mesrob Ohannessian

Mesrob I. Ohannessian, an assistant professor of electrical and computer engineering and, by courtesy, of computer science, wants to understand both the limits of what is possible with data, and how to achieve this potential fairly.

“Data is a resource, we can do something with it,” Ohannessian said. “But how much? Is it something simple, or can we afford more sophistication?”

With his new National Science Foundation CAREER grant, he plans to do this by streamlining the design of “competitive” algorithms.

Just as humans learn, machine learning algorithms used for artificial intelligence also form patterns and connections from the data they receive, improving themselves with more experience.

But if humans try to see too many patterns, they can hallucinate and arrive at unlikely explanations. The credo of Occam’s razor, or law of parsimony, addresses this. It states: the simplest answer is the best. But “simple” can mean a great many things.

“We don’t follow this credo often. When we adopt data-driven algorithms, we tend to be optimistic and throw the algorithm at everything, without caution to simplicity,” Ohannessian said. “We expect them to match the best current scientific and engineering solutions.”

Ohannessian says, in contrast, even the most sophisticated machine learning models are restrictive. In theory, they must start with the right assumptions about what is simple and must have enough parameters to make accurate predictions, but not so many that they stumble and over-explain.

“Then what is the root of this optimism?” asked Ohannessian. “Nature is well-behaved. With enough observation, we expect to discover it. I interpret this as believing that there are algorithms that, no matter what we face in nature, we could learn it, as though nature itself whispered in our ear in what way it is ‘simple.’”

Such algorithms can be called “competitive”, because they would work equally as well as creating a group of specifically designed algorithms.

With his $545,000 support, Ohannessian and his students in the DICE (Data, Information, and Computation, Equitably) research group will explore the principles that enable such competitive performance. He believes the key to this is methods that help cope with rare events, which are not represented well in data . “The challenge in all learning is understanding the unobserved. That’s where we need to rely on simplicity to simultaneously find the truth of nature and adapt to it,” he said.

Ohannessian’s project will also weave research with data science education efforts at several levels, with active outreach and engagement of traditionally underserved students.