Advances in Nearest Neighbor Methods: Non-Plug-in f-Divergence Estimation and Applications
김수현
27동 220호
0
2061
2025.08.21 14:52
| 구분 | ACM |
|---|---|
| 일정 | 2025-10-31(금) 10:30~12:00 |
| 세미나실 | 27동 220호 |
| 강연자 | 노영균 (한양대) |
| 담당교수 | 홍영준 |
| 기타 | ACM Seminar |
Nearest neighbor methods have long provided a simple yet flexible nonparametric framework, enabling density estimation followed by its plug-in use. Such algorithms include vanilla nearest neighbor classification, which is equivalent to making decisions based on comparisons of the plug-in density estimates. However, regardless of how easily these approaches can be applied to practical real-world applications, a fundamental issue with this straightforward methodology lies in the misconception that the plug-in principle is universally applicable with asymptotic guarantees and fully represents all nearest neighbor methods. In this talk, I will present both theoretical advances and practical applications of plug-in and non-plug-in nearest neighbor approaches for utilizing density functionals. This research explores developments specifically tailored for f-divergence estimation and their applications in adjusting trustworthiness and AI safety. I will show how an appropriate k-nearest neighbor methods can be derived by leveraging inverse Laplace transforms, offering a contrast to previous plug-in methodologies, which exhibit theoretical limitations when using a fixed k. Applications of these methods will be briefly discussed to address various challenges confronted in artificial intelligence, such as handling imperfect information, ensuring fairness, and eliminating artifacts in simulated data.
Bio:
Yung-Kyun Noh is a Professor in the Department of Computer Science at Hanyang University. He is currently the Director of the Generative AI Research Center and serves as the Chair of the Department of Artificial Intelligence at Hanyang University. His research interests range from understanding the fundamental principles of machine learning to developing practical applications that advance medicine and scientific discovery. These principles concern generalization and learning as mechanisms for maintaining life through information processing, decision-making, and intelligent communication. He is an official research collaborator with the Mayo Clinic and a Visiting Fellow at Weill Cornell Medicine in the United States. He has been a Visiting Scientist at RIKEN-AIP in Japan since 2018 and an Affiliate Professor at the Korea Institute for Advanced Study since 2021. He also served as a NeurIPS organizer in 2015.
Yung-Kyun Noh is a Professor in the Department of Computer Science at Hanyang University. He is currently the Director of the Generative AI Research Center and serves as the Chair of the Department of Artificial Intelligence at Hanyang University. His research interests range from understanding the fundamental principles of machine learning to developing practical applications that advance medicine and scientific discovery. These principles concern generalization and learning as mechanisms for maintaining life through information processing, decision-making, and intelligent communication. He is an official research collaborator with the Mayo Clinic and a Visiting Fellow at Weill Cornell Medicine in the United States. He has been a Visiting Scientist at RIKEN-AIP in Japan since 2018 and an Affiliate Professor at the Korea Institute for Advanced Study since 2021. He also served as a NeurIPS organizer in 2015.
