As a Ph.D. student at Arizona State University, Rushil Anirudh wanted to make the most of the time between academic years. He spent one summer at a surgical robotics company in the Bay Area. The next year, he worked for a startup company that was eventually bought by Google. In 2015, he interned at LLNL under the guidance of computer scientists Peer-Timo Bremer and Jayaraman Thiagarajan.

“A research laboratory is very different from commercial companies,” says Anirudh. “Here we are focused on research, not on putting out fires. For me, it was a seamless transition from academics.” His postdoctoral appointment with LLNL’s Center for Applied Scientific Computing (CASC) began in 2016. In 2018, he converted to full-time research staff.

As part of CASC’s Data Analysis Group, Anirudh contributes to a range of machine learning (ML), computer vision, and high-dimensional data analysis projects, often collaborating with scientists outside the Computing directorate—including National Ignition Facility experts, computational biologists, and engineers. He describes his Ph.D. in computer vision and ML as “quite theoretical,” so he is excited to expand his skills and apply that knowledge to solving real-world problems.

Part of Anirudh’s work involves bridging the gap between scientific research and machine learning. For one project, he reconstructs computed tomography (CT) images using convolutional neural networks from partially scanned objects. This approach could potentially cut the amount of time a CT scan takes in half, such as during a medical procedure or at an airport security checkpoint.

“We have to be careful when applying ML techniques to critical problems,” cautions Anirudh. “We can’t just throw data at a model.” Fortunately, the Laboratory is full of subject matter experts in specialized fields, from biomedicine to nuclear physics. Because of this proximity, he notes, “The Lab provides computer scientists with a unique opportunity to address these challenges.”

Anirudh presented a paper on CT reconstruction at the 2018 Computer Vision and Pattern Recognition conference, and his related work with generative adversarial networks was featured in NVIDIA’s developer blog. He also makes time to mentor summer students who, like he once did, seek LLNL internships during graduate school. LLNL’s Data Science Institute recently honored Anirudh for his many accomplishments.

Every day, Anirudh is motivated by the rapid pace of research. “I’m part of an amazing group in CASC,” he says, explaining that his projects have in common the application of ML to scientific simulations. He states, “CASC provides me with the freedom to pursue research all over the Lab. Eventually, most scientists want to explore ML for their projects, so there’s always something new happening.”

Figures: In a recent paper, Rushil Anirudh and colleagues trained a generative adversarial network on a data set of more than 200,000 faces. The team applied an unsupervised machine learning technique to image deblurring (image a, left) and edgemap recovery (image b, right). PGD stands for projected gradient descent. Click each image to enlarge.