연사: 황경하 교수(영남대학교 수학과)

일시: 2026.09.01(화) 15:00-16:00 자연과학대학 B06호

초록:  Deep learning has become a fundamental computational tool for image

processing, computer vision, and inverse problems. This lecture introduces

the basic ideas of deep learning from the viewpoint of image restoration.

We begin with supervised learning and discuss single-layer perceptrons,

multi-layer perceptrons, nonlinear activation functions, backpropagation,

the universal approximation theorem, and convolutional neural networks.

 We then consider situations in which clean target images are unavailable.

Deep Image Prior is introduced as an example of unsupervised image

restoration, followed by a self-supervised denoising strategy,

BS2SD, for removing salt-and-pepper noise. BS2SD exploits spatial

correlation between neighboring pixels and combines self-supervised

prediction with iterative refinement and data consistency.

 Numerical examples are presented to illustrate the reconstruction

performance, together with possible extensions to imaging inverse problems

such as CT and MRI.

연사: 지정민 교수(University of Louisville)

일시: 2026.07.07(화) 16:00-17:00 자연과학대학 219호

초록:  Singular perturbation problems in fluid mechanics, characterized by a small (viscosity) parameter multiplying the highest-order derivatives, give rise to multiscale behavior with sharp transitions such as boundary and interior layers. These features pose significant challenges for both analysis and computation. In this talk, I will present recent developments in semi-analytic Physics-Informed Neural Network (PINN) methods designed to address these difficulties in slightly viscous fluid flows. By incorporating viscous layer correctors into the learning framework, the proposed approach significantly improves both accuracy and computational efficiency. This work highlights the synergy between asymptotic analysis and machine learning in advancing the simulation of singularly perturbed fluid equations.