Jeff Druce, Ph.D., is the Principal Investigator for a variety of artificial intelligence programs at Charles River Analytics, including two DARPA-sponsored efforts. For ALPACA (a DARPA CAML program), Dr. Druce leads research on competency awareness in autonomous systems. On the CAMEL project (DARPA XAI), he is developing explanation methods and interfaces for black-box machine learning algorithms. Other research interests include robust classification strategies in memory and computationally limited environments.

---

The Trajectory of Jeff Druce

As a Senior Scientist at Charles River Analytics, Jeff specializes in applied artificial intelligence (AI), with an emphasis on testing, evaluating, and engendering trust in AI systems. But he began his scientific career studying mathematics and physics at the University of Michigan, a fact that’s still apparent in some of his outside interests.

“Every time my brother and I get together,” Jeff says, “We build something. And we usually blow it up. One time, we decided to jump an RC car over the garage and try to have it pass super close to a quad copter.”

Jeff’s interests prepared him for some unexpected jobs during the years between college and graduate school. He taught high school science for a year in a Florida public school, where he particularly enjoyed the lab component of his classes. “We built a giant sling shot that we used to study projectile motion. Our school mascot was the Pirates, so I painted the sling shot matte black, with a huge Jolly Roger on the side.” Jeff had the students calculate the angle needed to take out an “enemy ship”—the mascot of a rival school. For a lab on integrating data, students built rockets and placed accelerometers inside them.

Jeff applied some of these same skills to the project with his brother. After they built a launch ramp, Jeff put a GPS module on the RC car, figured out its speed, and worked out the kinematic equations for the trajectory, so they’d know exactly where the quad copter should hover. Jeff also ran a simulation to figure out exactly when to take the best photographs.

How Jeff wound up teaching high school science is a story in itself. “I actually didn’t apply for the job,” he said. “The principal called me directly.” At the time he received this unexpected job offer, Jeff was engaged in an entirely different career: playing professional golf.

Jeff had been a competitive junior golfer as a high school student, but didn’t play competitively during college due to an injury. After college, he knew he couldn’t attend graduate school and play professional golf at the same time. He chose golf. “My advisor said, ‘Socks before shoes. Golf before grad school,’” Jeff explains.

The highlight of Jeff’s golf career was the professional win in a 4-hole, sudden-death playoff against a former PGA tour player. The low point may have been when he hit the same house three times in one tournament.

The RC car launch had a low point as well, on the last day of Jeff’s visit. It had snowed the night before, and they had to shovel, sweep, and salt the driveway before they could get started. When they finally launched the car, it bounced over the side of the garage—and into the passenger side mirror of a car. It was, Jeff laments, “smashed to bits.”

But Jeff is no novice at making course corrections. Though his undergraduate research was in astrophysics, he wanted his graduate work to be more applied in nature. Choosing his degree program based on the opportunity to work with an up-and-coming faculty member, Jeff joined the Department of Civil, Environmental, and Geo-Engineering at the University of Minnesota, where he studied optimization, machine learning, and mathematics.

For his dissertation, “Agnostic Anomaly Detection Methodologies with Robust Inference Capability in Solid Media,” Jeff used techniques from optimization and computer vision. Though the term machine learning wasn’t in widespread use at the time, Jeff’s research employed a dictionary learning technique that eventually led him to an internship at the Air Force Research Laboratory (AFRL). At AFRL, he worked on structural diagnostics using machine learning techniques, and received a grant from the Center for Surveillance Research.

After completing his Ph.D. in 2016, Jeff came to work at Charles River Analytics. He chose Charles River not only for the opportunity to work in artificial intelligence and machine learning, but also for our atmosphere of “liveliness and curiosity,” which distinguished us from the other places he interviewed. He was deeply impressed by the people who would be his colleagues. “I couldn’t pass up that opportunity,” he said.

Jeff’s first project at Charles River, on insider threat detection and modeling, brought the excitement of a whole new technical area, probabilistic modeling. Other projects have involved computer vision, deep learning, reinforcement learning, homomorphic encryption, and medical autonomy.

His first opportunity to lead a project as the Principal Investigator came with CAMEL (Causal Models to Explain Learning), a four-year contract with the Defense Advanced Research Project Agency (DARPA) valued at close to $8 million. Jeff is also the Principal Investigator on ALPACA (Advancing Learning via Probabilistic Causal Analysis for Competency Awareness), which is developing an agent for robot navigation in outdoor environments and using the agent as a platform for research on competency-aware machine learning.

As a leader in the field of XAI, Jeff has co-authored a number of papers on the emerging discipline, including “Brittle AI, Causal Confusion, and Bad Mental Models: Challenges and Successes in the XAI Program,” which describes lessons learned from the DARPA XAI program and possible directions for future research in XAI. “With complex AIs, the challenge is to ‘open the black box’ of the AI’s decision-making process, without sacrificing performance.”

Jeff cites his teaching experience as a major influence on his current work: “I went from working on astrophysics to teaching first-year high school students. Now, in the XAI program, we have to present explanations that a non-expert can understand and use to make better decisions. Figuring out how to help students understand a concept was valuable practice for that.”

Thus far, CAMEL has been Jeff’s favorite project at Charles River; it was so rewarding, he says, “to see that explanation user interface work and see our hypotheses validated.” And the most rewarding moment of that extracurricular project? “The second time we launched it… the car flew so gracefully. It landed perfectly, in a snowbank on the opposite side of the garage. Exactly where the equations predicted.”