Projects

Provably robust conformal prediction with Improved efficiency

To appear on The Twelfth International Conference on Learning Representations (ICLR24)

Conformal prediction is a method that provides prediction sets with guaranteed coverage. However, this guarantee fails at adversarial attacks. In this project, we propose a novel framework called RSCP+ which provides guaranteed robustness against adversarial examples. Besides, we design two methods to boost the efficiency of conformal prediction (reduce the size of prediction sets). Our methods achieve 4.36x, 5.46x and 16.9x efficiency boost on CIFAR10, CIFAR100 and ImageNet, respectively.

Failure probability estimation via Bayesian neural networks

Neural networks could be used as a PDE solver with much lower cost comparing to traditional finite elements methods, which makes them to be a possible kind of model when cheap approximation of PDE solution is needed. In this project I applied Bayesian neural networks as a PDE solution approximator in failure probability estimation. Details could be found in my B.S. thesis.