821. Learning adaptive universal denoisers with deep neural networks

¾연사: 문태섭 (성균관대 전자전기공학부 조교수)

¾일시: 2017년 09월 21일(목) 오후 5:00~6:00

¾장소: 서울대학교 제1공학관(301동) 118호

Abstract:

In this talk, I will first briefly reflect on my paths that I took after graduating from SNU EE, namely, Stanford, Yahoo!, Berkeley, Samsung, DGIST and SKKU, and summarize how my research topics have varied among various principles, such as information theory, applied statistics, and deep learning. After arguing that the topics I considered share a coherent theme of estimating an unknown source from its observation, I will introduce a recent work on applying deep neural networks to devising a novel universal denoiser, which attempts to combine the principles I learned along my paths. More specifically, I will show how the neural networks can be adaptively trained as a denoiser solely with statistical “loss estimators” obtained from noisy data (and without any ground-truth labels). I will then show that the performance of such adaptively trained denoiser can be boosted when combined with supervised learning, by presenting strong empirical results on two very different data sources, namely, image and DNA sequence. Our scheme significantly outperforms the previous state-of-the-art developed from the information theoretic principle. Finally, I will conclude the talk with some on-going and future research directions.

Biography:

• 학력

  2008   Stanford University 전자공학 박사

  2004   Stanford University 전자공학 석사

  2002   서울대학교 전기공학부 학사

   

• 주요 경력

  2017.3~          성균관대학교 전자전기공학부 조교수

  2015.9~2017.2   대구경북과학기술원 (DGIST) 정보통신융합공학전공 조교수

  2013.9~2015.8   삼성종합기술원 전문연구원

  2012.2~2013.8   UC Berkeley 통계학과 박사후 연구원