Published by Penn Engineering Today
Authored by Melissa Pappas
Image by Melissa Pappas
Nikolai Matni, Assistant Professor in Electrical and Systems Engineering, is one of Penn Engineering’s most recent recipients of the National Science Foundation’s CAREER Award. The CAREER Award is given to early-career faculty researchers who demonstrate the potential to be role models for research and education and are committed to outreach and public engagement.
Matni’s work sits at the intersection of two branches of applied mathematics, control theory and machine learning, and focuses on the theoretical side of artificial intelligence.
“My work attempts to combine the strengths of control theory and machine learning,” says Matni. “An example of a traditional control system is cruise control in a car. This simple algorithm operates based on a feedback principle, measuring the difference between the current and desired speed of the car and accelerating or slowing down based on this difference. This feedback allows cruise control to operate well across a broad range of road conditions, accounting for some amount of uncertainty in the environment.”
Making computer-based systems “smarter” means granting them more autonomy to deal with a wider range of situations. However, this requires the systems to take in more information, some types of which may be harder to quickly assess and compare.
“As cars become more autonomous, for example, their cruise control systems use cameras to detect the environment around them” Matni says. “Machine learning is incorporated into these control systems to make decisions based on the images coming from the cameras. However, how to properly interpret the information contained in that stream of images is also not well understood, which introduces its own unique uncertainties in the control system.”
A core concept to his work is “robustness,” which can be simply defined as the ability of a machine to correctly respond to the varying amounts of uncertainty in the real world. Matni’s research focuses on the link between the robustness of feedback control and the uncertainty in machine learning to help inform safe and reliable systems we use in our everyday life.
“When designing systems that are applied to real-world machines, it is very important to be able to guarantee a certain level of safety. This is what robustness refers to,” says Matni. “We need to create mathematical models that span a large range of uncertainty so that we do not need to create a specific model for every unique system or circumstance that the machine could run into. For example, autonomous cars need a control system that can account for uncertainties such as people crossing the road, car collisions, stop lights and potholes. A robust system will be able to account for a wide range of worst-case scenarios.”
Matni’s goal is to push the underlying mathematics that allows the integration of control theory and machine learning. A better understanding of how generalizable these theories are will result in better machine learning approaches, which in turn will result in autonomous machines that are able to reliably adjust to more environments.
“The way we answer this question is by chipping away at a main problem,” says Matni. “Whether it’s self-driving cars facing diverse and potentially dangerous weather patterns, medical devices implanted into patients with different conditions, or bipedal robots having to run across different types of terrain, we would work to identify the core issues at play, then use these to pose a theoretical problem. However, reality is messy, and so simplifications and abstractions are almost always needed in order to make progress. Nevertheless, the insights we gain from such abstractions can tell us a lot about the fundamental properties of machine-learning enabled control systems. Once we feel we have a solid handle on the current version of the problem, then we can start to reintroduce more complications from the real-world environment to see how much more we can close the gap between our theoretical guarantees and reality.”
Matni plans to use funds from the CAREER Award to develop specific courses for undergraduate and graduate students at Penn. These courses will take the many disciplines of this emerging field and translate them into a common language and perspective that will aid in the development of a new and interdisciplinary research field.
“I imagine these courses will be a combination of both theory and practical lab work. We will focus heavily on the theoretical side of machine learning, but I would add in a simulator experience that allows students to test their theories,” says Matni. “I am very excited to be able to develop these courses and receive training to create the curriculum through the Center for Teaching and Learning at Penn.”
In addition to increased student engagement, Matni plans to bring his research to the public, a mission that sets the CAREER Award apart from other research grants.
“Another goal I have with funding from the CAREER Award is outreach through the Franklin Institute’s ‘Portals of the Public,’ where my graduate students and I will perform table-top demos of our research to showcase our research to the general public. The aspect of collaboration and focus on outreach is very unique to these awards and has been the most rewarding part of writing and receiving the grant,” says Matni.