The Computing Research Association (CRA) has named fourth-year undergraduates Jacob Kelly and Winnie Xu as finalists in the 2022 Outstanding Undergraduate Researcher Award program.
The program recognizes undergraduate students in North American colleges and universities who show outstanding potential in an area of computing research.
Assistant Professor David Duvenaud
Kelly and Xu have each led neural network experiments under the guidance of Assistant Professor David Duvenaud.
“Jacob and Winnie both produced world-class research projects as third-year undergrads, which is extremely impressive,” remarked Duvenaud. “The key ingredients were determination, being able to say ‘I don't know,’ and to be willing to use unfinished, bleeding-edge tools.”
Winnie Xu
Xu led the experiments on a project to develop a new kind of deep neural network architecture that has infinitely many parameters and infinitely many layers.
“Of course, learning and prediction in such a network can't be handled exactly, but different approximations correspond to different existing architectures,” explained Duvenaud. “The advantage of this approach is that the user can adjust the quality of these approximations at any point, trading off compute cost against solution quality, without having to restart from scratch.”
Earlier this month, this paper was accepted to one of the top 5 international machine learning conferences, the conference on Artificial Intelligence and Statistics (AISTATS).
Infinitely Deep Bayesian Neural Networks with Stochastic Differential Equations
Winnie Xu, Ricky T.Q. Chen, Xuechen Li, David Duvenaud
Jacob Kelly
Kelly led the experiments on a project to make training infinitely-deep neural networks, and differential equations more generally, easier to solve.
“Specifically, when parameters of ordinary differential equations are being fit to data, the resulting model sometimes ends up being very expensive to simulate. Jacob investigated ways to encourage the resulting models to be as fast to solve as possible, while still fitting the data well,” said Duvenaud.
The resulting paper was accepted to the top international machine learning conference, Neural Information Processing Systems (NeurIPS).
Learning Differential Equations that are Easy to Solve
Jacob Kelly, Jesse Bettencourt, Matthew James Johnson, David Duvenaud