Francesco Restuccia has been awarded a Presidential Early Career Award to improve the reliability of wireless systems and AI models.
In the AI age, industries from biotechnology to defense are relying on autonomous systems with increasing frequency.
One clear example is the use of drones that do not rely on direct human guidance to deliver packages, explained Francesco Restuccia, a professor of electrical and computer engineering at Northeastern University.
These aerial machines regularly navigate areas that could look different from the data their artificial intelligence systems were trained on, he said. That’s dangerous because the drones may face obstacles that would be difficult for them to foresee, like hidden walls.
At Northeastern, Restuccia will develop new mathematical frameworks that could help these machines operate more smoothly in complex environments.
“Over the years, I’ve discovered that a lot of these systems are very uncertain,” he said. “They are surrounded by uncertainty in their data, in their communication systems, everywhere. But there is no principled way of characterizing that uncertainty. This is what my project fills out.”
To support this research effort, Restuccia has been awarded the Presidential Early Career Award for Scientists and Engineers, one of the highest honors awarded by the federal government given to young researchers in the U.S. He was nominated for the award by the U.S. Army Research Office.
As part of the award, Restuccia, 38, has been awarded $1 million to pursue the five-year project.
Restuccia will work on building on Classical information theory, a framework developed in 1948 useful for ensuring reliable and consistent information sharing between tele-communication systems.
Restuccia’s work deals with the difference between “randomness” and “ignorance” when it comes to wireless technology and AI systems. Today’s systems are excellent at understanding randomness in their environment, but they are less good at understanding ignorance, he said.
For example, if you were to quiz an AI model on the likely outcome of a roll of a pair of dice, it could easily understand the concept. Since all sides are equal, any outcome is likely to occur. It’s completely random, Restuccia said. However, if one of those dice were weighted, and the AI model did not know that fact, it could easily make a mistake when asked about the likely outcome of a roll. An AI model can, however, reduce its ignorance by rolling the pair of dice, over and over again until it recognizes a pattern.
The Classical information theory, Restuccia explained, provides a great framework for accounting for randomness. It fails to account for ignorance, a major issue that needs to be addressed, he said.
“This is becoming increasingly important because we’re deploying AI and communication systems in environments that are complex and dynamic, and so there is going to be ignorance,” he said, stressing the importance for researchers to be able to establish the mathematical framework that will allow a system to understand the “two sources of uncertainty” demonstrated in the dice experiment.
“What we want is a framework that allows a system to say, ‘Here’s my best estimate, but here’s what the best evidence allows me to conclude,’” he said.
That is the framework Restuccia is planning to develop with the federal government’s support. If successful, his algorithm could be helpful in high risk scenarios.
“The goal here isn’t to eliminate uncertainty,” Restuccia said. “It’s to know what we know. Know what we don’t know. And design the system accordingly.”