- Speaker: Prof. Seunghoon Hong (KAIST)

- Title: Modeling complex conditional distributions with hierarchical and diversity-sensitive conditional GANs

- Time: 16:00, Nov 11, 2019

- Place: Rm. 1501, E3-1 (SoC building)

- Language: English

Here is the speaker's short biography and the talk's abstract:

Bio

Seunghoon Hong is an assistant professor at the School of Computing, KAIST. Before joining KAIST, he had been a postdoctoral fellow at the University of Michigan and visiting research faculty at Google Brain team. His research interests lie in the intersection of machine learning and computer vision, with a specific focus on learning with least supervision and deep generative models. He received the B.S. and Ph.D. degree from the Department of Computer Science and Engineering at POSTECH, Pohang, Korea in 2011 and 2017, respectively.

홍승훈 교수는 2019년부터 KAIST 전산학부 조교수로 재직 중이다. 주 연구분야는 기계 학습 및 이를 활용한 자율적 시각 인지 시스템이다. 홍승훈 교수는 POSTECH에서 학사 및 박사 학위를 각각 2011년과 2017년에 받았으며, 2017-2019년에는 University of Michigan에서 박사 후 연구원으로, 2019년에는 Google Brain team에서 방문 교수로 재직하였다.

Abstract

Modeling a generative process of data conditioned on input signals is one of the important research problems in machine learning and derives a wide array of practical applications. Despite great advances in deep generative models, however, modeling complex conditional distributions between high-dimensional input and output data still remains a challenging open problem due to the complexity and ambiguity of the mapping. In this talk, I will discuss our recent attempts to resolve such challenges based on conditional Generative Adversarial Networks (cGAN). In the first half of the talk, I will introduce hierarchical approaches that decompose a complex mapping into a sequence of more feasible sub-generation tasks. Specifically, I will discuss approaches that construct semantic layout as an intermediate representation between input and output, which is turned out to be very useful in certain problems such as text-to-image synthesis and semantic image manipulation. In the second part of the talk, I will introduce a simple way to regularize the generator to address a mode-collapse problem in conditional GANs. Despite the simplicity, I will show that the proposed regularization is widely applicable to various cGAN architectures and tasks, and outperforms more complicated methods proposed to achieve multi-modal conditional generation for specific tasks, such as image-to-image translation, image inpainting, and video prediction.

If you want to get more information about the colloquium, please refer to colloquium page: https://cs.kaist.ac.kr/colloquium/