Seed grant award winner researches ways to tailor AI applications to Maine climate
In July 2020, Salimeh Yasaei Sekeh was announced as one of four Seed Grant Award winners through the UMaine AI Initiative. Her project titled “Improved Adversarial Attack Detection Toward Robustness of Deep Neural Networks,” is currently in the first two phases of research and data analysis.
Sekeh is an assistant professor of Computer Science in the School of Computing and Information Science. She is currently teaching a graduate-level course in machine learning while conducting research projects out of the Sekeh Lab.
Sekeh and her team are looking at deep neural networks, which is an artificial intelligence technique that simulates the human brain. They also study the shortcomings of these networks when faced with adversarial attacks.
In neural networks, there are layers that information passes through in order to perform or learn tasks. They are trained through data and computer vision learning architectures. Deep neural networks are used in several machine learning applications, which show them how to interpret the world and perform different tasks.
In extreme weather or other natural phenomena, these machines can have difficulty in object detection tasks such as reading signs because the details of the environment are altered or physical adverse weather, such as road signs covered in snow.
“One of the problems that sometimes people in AI and the computer community realize is that [these networks] are not robust enough, and what we mean by they are not robust is that sometimes we have a series of videos or images, or we want to drive a drone and the weather is brutal…or we had a storm and the sign is rotated. Some neural networks are not performing well. The autonomous vehicles or robots are not accurate enough to be trusted,” Sekeh explains.
In the first phase of their research, Sekeh and her team are zooming into deep neural network architectures to look at connectivity between layers and between neurons to analyze the deep features’ robustness. Eventually, they intend to be able to control networks robustness to better interpret adversarial examples, like snow or covered signs.
The success of AI-driven technology heavily depends on the source data. Next, Sekeh hopes her team is able to develop a unique data set for the state of Maine.
“We get a lot of storms. The big question here is how much autonomous vehicles can perform accurately and well here in this state. The data that have been developed [in other states] can’t actually be applied here because our weather, our climate, is not the same as other states,” Sekeh said. Developing their own data set will help the team test the robustness of neural networks in a Maine-specific setting.
Sekeh and her team would like to apply their new techniques to Autonomous vehicles possibly via Autonomous Visualization Systems or AVS and sensor fusion. AVS is a new open-source standard for the industry to share data, as a way to encourage industry-wide improvements and new development.
“We are hoping when we progress in the project toward the end of the year, we will be able to generate and develop some interesting techniques. And hopefully, in the near future, we will be able to apply them to our collected data in Maine,” Sekeh said.
Written by Ali Tobey