Rising Stars in Machine Learning Talk Features Gintare Karolina Dziugaite
![](https://www.umiacs.umd.edu/sites/default/files/styles/medium/public/Karolina.jpg?itok=k8DUT3F2)
The University of Maryland Center for Machine Learning concludes its Rising Stars in Machine Learning program on Monday, Oct. 21 with an hour-long talk on “PAC-Bayesian approaches to understanding generalization in deep learning.”
The talk will be given by Gintare Karolina Dziugaite, a fundamental research scientist at the Canadian start-up, Element AI. Her research combines theoretical and empirical approaches to understanding deep learning, with a focus on generalization, network compression and adversarial training.
Dziugaite is one of three speakers chosen to present their work this fall at the University of Maryland as part of the Rising Stars in Machine Learning program, which supports up-and-coming female researchers as they pursue new scientific and academic opportunities.
“The winners have demonstrated strong track records in successfully tackling fundamental and practical research problems in the areas of machine learning and deep learning,” says Soheil Feizi, an assistant professor of computer science and core member of the Center for Machine Learning.
Dziugaite was a fellow in the Simons Institute for the Theory of Computing at the University of California, Berkeley in the Foundations of Machine Learning program.
“It was at Simons where I started studying generalization in deep learning, using tools from statistical learning theory, but combining them with a strong emphasis on empirical evaluation,” Dziugaite says. This research became the focus of her doctoral dissertation, which she defended last March at the University of Cambridge.
“The question of whether or not known generalization bounds explain the impressive performance we see in practice is subtle,” says Dziugaite about her research. “In my recent work, I challenge the ‘asymptotic viewpoint’ and seek out numerically non-vacuous bounds, and argue that the empirical evaluation of generalization bounds is essential to uncover the actual explanatory power of a generalization bound.”
Dziugaite “has also done some great recent work generalizing some past frameworks into the differential privacy space,” says John Dickerson, an assistant professor of computer science and core member of the Center for Machine Learning.
Dickerson also praised Dziugaite’s previous work that looked at Generative Adversarial Networks.
“She had the influential idea of using a two-sample test statistic based on the maximum mean discrepancy instead of an adversarial network in GANs, which may end up lending itself to helping us understand GAN theory and practice better,” says Dickerson.
Previous Rising Stars in Machine Learning speakers include Adji Bousso Dieng, a doctoral student in statistics at Columbia University, and Surbhi Goel, a doctoral student in computer science from the University of Texas at Austin.
The UMD Center for Machine Learning, supported in part by financial and technology leader Capital One, is part of the University of Maryland Institute for Advanced Computer Studies.
–Story by Colleen Curran