Eric Nalisnick joins the Johns Hopkins University as an assistant professor of computer science.
Nalisnick received his PhD from the University of California, Irvine. Before joining Hopkins, he was an assistant professor at the University of Amsterdam and a postdoctoral researcher at the University of Cambridge.
Tell us a little bit about your research.
A broad theme in my research is uncertainty quantification for AI systems. Consider an AI agent tasked with diagnosing diseases based on patient data such as blood pressure, heart rate, and blood tests. A system that just says “this patient has heart disease” is not useful in that clinicians will ask for more information. A better answer is, “With a 75% chance, this patient has heart disease. Collecting more information about their family history will increase the system’s confidence in this diagnosis.” My research aims to equip AI systems with the self-awareness to quantify what information they know and don’t know. In turn, this will result in AI systems that are more reliable and safer to deploy in high-stakes settings.
Tell us about a project you are excited about.
In collaboration with the Bosch group, I have been investigating uncertainty quantification for computer vision. A fundamental task in computer vision is generating bounding boxes around objects in an image. Bounding boxes allow, for example, a self-driving car to locate the cars and pedestrians that are also on the road. My group’s work develops methods that provide error bands around bounding boxes. These error bands have theoretical guarantees about the chance that they will include the true—but unknown—bounding box.
So far we have released a pre-print; see below for a visualization of the results. Notice that for “typical” pedestrians, such as the second person from the left, the error bars (shown in green) are relatively tight around the bounding box (shown in red). The presence of “non-typical” objects will then increase the width of the error bands. For instance, the pedestrian on the far left is wearing a backpack, which inflates the right-side error bands. On the far right, a stroller is entering the field of view, and the bounding box contains only the child and not the stroller itself. Fortunately, our method recognizes that this box may be incorrect and increases the size of the error band to include the stroller. This would then result in a self-driving car navigating around the stroller and not just the child.
Examples of conformal bounding box intervals. True bounding boxes are in red; two-sided prediction interval regions are shaded in green.
Why this? What drives your passion for your field?
Ever since my first class on AI, I have been interested in the topic of uncertainty quantification. This interest firstly stems from curiosity about how the human brain works: Humans are notoriously bad at thinking in terms of chances and probability; creating machines that do it well may mean that we are making progress in representing and engineering possibly superhuman intelligence. I am also interested in the topic from the standpoint of practice: If we are to use AI to improve health care, develop self-driving cars, etc., then these AI systems must be aware of what they know—and, in turn, which tasks they are good at—and what they don’t know.
What classes are you teaching?
This fall, I am teaching a graduate course titled Human-in-the-Loop Machine Learning. Enabling humans and AI systems to collaborate effectively is another topic of interest for me. This course will teach students the mathematical principles behind this emerging research field. We will study topics such as how an agent can learn from (imperfect) human demonstrations, how to switch between human and machine decision-makers, and how humans can teach safe behaviors, even when AI goes beyond human abilities.
Why are you excited to be joining the Johns Hopkins Department of Computer Science?
I find the department so exciting because of its multidisciplinary, impactful research—which spans applications ranging from large language models to surgical robotics to assured autonomy. I have encountered few other departments that have such close and consistent contact with today’s cutting-edge technologies. I also think that Baltimore is a fantastic place to live—for the city itself and for its proximity to the other great cities on the East Coast.
Besides your work, what are some of your other hobbies and passions?
In my free time, I like to enjoy the outdoors, practice American Sign Language, attend a baseball game, or read a book. Some people might be surprised to know that horror is my favorite genre of books and films.