한국정보과학회 인공지능소사이어티 머신러닝 연구회
두 번째 딥러닝 워크샵


프로그램

Deep Learning: An Overview
강사

최승진 교수, POSTECH

http://mlg.postech.ac.kr/~seungjin/ 

약력 Professor, Department of Computer Science and Engineering, Pohang University of Science and Technology
Ph.D. in Electrical Engineering, 1996 University of Notre Dame, Indiana, USA
MS. in Electrical Engineering, 1989 Seoul National University
BS. in Electrical Engineering, 1987 Seoul National University
초록 In this talk I give a brief overview of deep learning, explaining restricted Boltzmann machines, deep belief networks, auto encoders, deep Boltzmann machines, and convolutional neural networks. In earlier workshop (on April 24), I was supposed to deliver this overview, but the limited time allowed me to cover only less than the first half. This time, one hour is assigned, so I will be able to finish this overview. I begin with restricted Boltzmann machine (a.k.a. harmonium) which is a basic building block for deep belief network. I emphasize two extensions of harmonium: (1) to exponential family in order to accommodate various probability distributions; (2) to multi-wings in order to accommodate multi-modal data. Auto-encoders, stacked auto-encoders are also introduced. Recent advances, such as variational auto-encoders are also introduced. I describe a few different deep networks for multi-view learning and pinpoint a distinction between deep belief network and deep Boltzmann machine. Finally, convolutional neural networks are described, with its application to human activity recognition.
Toward Scalable Deep Learning
강사

윤성로 교수, 서울대학교

http://best.snu.ac.kr/index.html

약력 Associate Professor, Electrical and Computer Engineering, Seoul National University
Assistant Professor, Electrical Engineering, Korea University (2007-2012)
Senior Engineer, Intel Corporation, Santa Clara, California, USA (2006-2007)
Postdoctoral Scholar, Stanford University, California, USA (2006)
Ph.D. in Electrical Engineering, Stanford University, California, USA (2006)
M.S. in Electrical Engineering, Stanford University, California, USA (2002)
B.S. in Electrical Engineering, Seoul National University (1996)
초록 In this presentation, I will first overview fundamental concepts and techniques for implementing machine learning algorithms in a scalable fashion. Both scale-up approaches (such as the use of GPUs and multi-core CPUs) and scale-out techniques (such as distributed parameter servers) will be covered. I will also review notable approaches to scalable deep learning and present DeepSpark, a memory-based distributed deep learning platform that is currently under development in collaboration with the POSTECH Machine Learning Center.
Structured Prediction using CNNs
강사 한보형 교수, POSTECH

http://cvlab.postech.ac.kr/~bhhan/

약력 Associate Professor in Dept. of Computer Science and Engineering, POSTECH, 2014-present
Assistant Professor in Dept. of Computer Science and Engineering, POSTECH, 2010-2013
Ph.D. in Computer Science, University of Maryland at College Park, 2005
M.S. in Computer Engineering, Seoul National University, 2000
B.S. in Computer Science, Seoul National University, 1997
초록 Abstract: Convolutional neural network is one of the most powerful representation learning frameworks and has been used for various classification problems in computer vision. Recent advances in convolutional neural network shows its potential extension to more complex and structured problems, and typical examples include semantic segmentation, visual tracking, image generation, etc. In this talk, I will introduce several structured prediction problems based on convolutional neural networks, and discuss how such problems can be formulated using convolutional neural networks straightforwardly. Also, I will present how convolutional neural networks can be combined with other deep learning frameworks such as recurrent neural networks.
Compressing CNN for Mobile Device
강사 김용덕 박사, Senior Engineer, 삼성전자 DS부문 소프트웨어연구소
약력 2015~current: Senior Engineer, 삼성전자 DS부문 소프트웨어연구소
2005~2014: PhD, POSTECH
2001~2005: BS, POSTECH
초록 Convolutional Neural Network (CNN)은 imagle classification/object detection에서 뛰어난 성능을 보이나, 계산 자원을 많이 요구하기 때문에 (예: Oxford's VGG19 ==> 548Mbytes,20Gops), 모바일/임베디드 환경에서 구동하기에는 무리가 있다. 그러나 최근 들어 CNN 모델 파라미터 내에 redundancy가 존재함이 밝혀지고, 이를 바탕으로 모델 압축 기술에 대한 연구과 동시 다발적으로 진행되고 있다. 본 강연에서는 다양한 모델 압축 기법들을 소개하고, 압축 기법을 적용했을 때의 성능(메모리,속도,소모전력)과 정확도 사이의 tradeoff에 대해 살펴본다.  
Multiple Object Class Detection & Localization with Deep Learning (CNN)
강사 이한성 수석, 삼성전자 SW R&D Center
약력 Principal Engineer, SW R&D Center, Samsung Electronics. (2014 – Current)
Senior Researcher, SW·Content Research Lab., ETRI. (2009 – 2014)
Ph.D. in Computer Science, Korea University (2008)
M.S. in Computer Science, Korea University (2002)
Engineer, Daewoo Engineering (POSCO Eng.) (1996 – 1999)
B.S. in Computer Science, Korea University (1996)
초록 This talk will share the brief overview of multiple object class detection & localization problems with deep learning. Some issues for applying deep learning, in particular convolutional neural network, to multiple object class detection will be presented. The recent methodologies and applications about deep learning based multiple object class detection & localization will be introduced.
Deep Learning and Medical Image Analysis
강사

이예하 박사(yeha.lee@vuno.co), VUNO

http://vuno.co

약력 CEO of VUNO Inc.
Reserch Staff Member, Samsung Advanced Institute of Technology, 2014
Ph.D. Computer Science & Engineering, POSTECH, 2012
M.S. Computer Science & Engineering, POSTECH, 2005
초록 딥러닝은 영상,음성,텍스트 등의 다양한 분야에서 기존의 방법론을 뛰어넘는 성능을 보여주고 있으며, 최근에는 의료분야에서도 다양하게 활용되고 있다. 특히 의료영상은 전체 의료데이터의 90%이상을 차지하고 있으며 환자의 진단 및 처방, 치료를 위한 중요한 수단중 하나로 사용되고 있다. 본 강연에서는 딥러닝에 대한 소개와 의료분야에서 딥러닝이 어떻게 활용되고 있는지를 살펴본다.
Sequence to Sequence Learning for NLP
강사

정상근 박사, SK Telecom 미래기술원

약력 한국전자통신연구원(ETRI) 선임연구원 2012-2014
삼성전자 책임연구원 2010.3-2012
Ph.D. POSTECH, Computer Science and Engineering, 2010
MS. POSTECH, Computer Science and Engineering, 2006
BS. POSTECH, Computer Science and Engineering, 2004
초록 최근 여러 분야에서 탁월한 연구 결과를 보여주고 있는 딥러닝을 이용한 Sequence to Sequence(S2S) learning 방법 및 그 응용분야를 소개한다. 특히 자연어 처리 분야에서 S2S 방법론이 적용되는 여러 사례와 응용방법을 보임으로써, 패턴인식을 넘어서는 딥러닝의 활용 가능성을 소개한다.