한국정보과학회 인공지능소사이어티 머신러닝 연구회
딥러닝 워크샵: 딥러닝의 현재와 미래


프로그램

Deep Learning: A Quick Overview [발표자료]
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, beginning with a basic building block such as harmonium (aka restricted Boltzmann machine). 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-encoder and stacked auto-encoder are also introduced. I also describe two different deep networks for multi-view learning. Finally, I introduce my earlier work on multiplicative up-propagation learning which is used to train a deep network with nonnegative weights.
Recent Advances in Deep Learning: Learning Structured, Robust, and Multimodal Models (skype 원격 영상 강의) [발표자료]
강사

Dr. Ruslan Salakhutdinov, University of Torronto

http://www.cs.toronto.edu/~rsalakhu/

약력 Assistant Professor, University of Toronto, Department of Computer Science and Department of Statistical Sciences
Postdoctoral Research Associate, Brain and Cognitive Sciences (BCS) and Computer Science and Artificial Intelligence Lab (CSAIL), MIT.
PhD, Department of Computer Science, University of Toronto. Thesis: Learning Deep Generative Models, advised by Geoffrey Hinton.
MS, Department of Computer Science, University of Toronto. Thesis: Optimization Algorithms for Learning, advised by Sam Roweis.
BS, High Point University, NC, USA. Double major in Computer Science and Mathematics, Honors Degree.
초록 Building intelligent systems that are capable of extracting meaningful representations from high-dimensional data lies at the core of solving many Artificial Intelligence tasks, including visual object recognition, information retrieval, speech perception, and language understanding.
In this talk I will first introduce a broad class of hierarchical deep learning models and show that they can learn useful hierarchical representations from large volumes of high-dimensional data with applications in information retrieval, object recognition, and speech perception. I will then describe a new class of more complex models that combine deep learning models with structured hierarchical Bayesian models and show how these models can learn a deep hierarchical structure for sharing knowledge across hundreds of visual categories. Finally, I will introduce deep models that are capable of extracting a unified representation that fuses together multiple data modalities as well as discuss models that can generate natural language descriptions of images. I will show that on several tasks, including modelling images and text, video and sound, these models significantly improve upon many of the existing techniques.
Artificial intelligence and Intelligence Business [발표자료]
강사 정상근 박사, SKT 미래기술원
약력 한국전자통신연구원(ETRI) 선임연구원 2012.4-2014.4
삼성전자 책임연구원 2010.3-2012.4
Ph.D. POSTECH, 컴퓨터공학 박사 2010
MS. POSTECH, 컴퓨터공학 석사 2006
BS. POSTECH, 컴퓨터공학 학사 2004
초록 이번 강연에서는 최근에 부상하고 있는 인공지능과 이를 이용한 산업의 변화에 대해 소개하고, ICT기업 특히 통신사 입장에서 인공지능 시대를 준비하는 관점을 소개한다. Intelligence Business생태계를 만들어 내려는 여러 노력에 대해 소개하고, 특히 딥러닝으로 대표되는 머신러닝 기법이 Intelligence Business 와 어떻게 접목되어 가치를 창출할 수 있을지에 대해 소개하도록 한다.
Recent Advances in Recurrent Neural Networks [발표자료]
강사 최영상 박사, Project Leader, 삼성종합기술원
약력 2009~ Research Staff Member and Project Leader, Samsung Advanced Institute of Technology
2004~2009: PhD, Georgia Institute of Technology
1999~2004: Software Engineer, Samsung Electronics
1997~1999: MS, Seoul National University
1993~1999: BS, Seoul National University
초록 최근 딥러닝의 연구방향이 이미지등의 정적인 데이타 분석에서, 음성이나 텍스트 혹은 비디오 등의 동적인 데이타 분석으로 급속히 확산되는 추세이다. 본 강연에서는 이러한 연구에 중점적으로 사용중인 recurrent neural networks (RNNs) 의 알고리즘과 관련 기법들을 살펴본다. 또한 음성인식 등 관련 분야에서의 최근 성과에 대해 소개한다.  
Accelerate Deep Learning app with GPU [발표자료]
강사 유현곤 차장, NVIDIA Korea
약력 2009. 1 ~ : NVIDIA Korea, PSG Solution Architect for CUDA
2008.3~ :  PhD. cource,  Dept. of Mathematics  , Yonsei Univ.
2008.2 : MS. Dept. of Mathematics, Yonsei Univ.
초록 In this talk, I give brief overview of GPU architecture and how to accelerate app with cuBLAS library which include SGEMV and SGEMM routines. I’ll show how to represent image processing algorithm to matrix-vector operation such as image batch to 4D Tensor operation, 2d image convolution to SGEMV and 2d image multi convolution to SGEMM. this routines are main hot-spot of convolutional neural network.  after that, I’ll show how to integrate cuDNN library to open-source Deep Learning frameworks(caffe, torch7 and Theano).
Deep Learning for Natural Language Processing [발표자료]
강사

이창기 교수, 강원대학교

http://cs.kangwon.ac.kr/~leeck/

약력 Assistant Professor, Dept. of Computer Science, Colleage of Information Technology, Kangwon National University, Chuncheon, Korea (강원대학교 IT대학 컴퓨터과학과 조교수 이창기)
PhD in Computer Scinece and Engineering, POSTECH, Korea 2004
MS in Computer Science and Engineering, POSTECH, Korea 2001
BS in Computer Science, KAIST, Korea, 1999
초록 본 강좌에서는 자연어처리를 위한 딥 러닝 기술을 소개한다. 먼저 딥 러닝 기술을 자연언어처리 분야에 적용할 때 필요한 word embedding에 대해 소개하고, 이를 한국어 자연어처리(구문분석, 개체명 인식 등)에 적용한 예를 살펴본다.
Deep Learning, Long Learning, and Human-Level AI [발표자료]
강사

장병탁 교수, 서울대학교

http://bi.snu.ac.kr/~btzhang/

약력 Ph.D. Compuer Science, University of Bonn, Germany
BS MS. Computer Science and Engineering Seoul National University
초록 Recent successes of AI have been predominantly due to the advancement of machine learning technology, such as deep learning. However, AI still is far away from the human-level or human-competitive intelligence and, probably, we first have to achieve human-level machine learning. In this talk, I will contrast the properties of machine learning to those of human learning to identify the essential capabilities for achieving human-level ML. As an important finding, I will argue that “long” learning (in addition to “deep” learning, of course), i.e. learning continuously over an extended period of time or “lifelong” in a changing environment, is fundamental for building the future generation of AI machines. Time permitting, I will also discuss the principles with which the brain deals with this challenge.
Visualization and Discriminative Learning of Convolutional Neural Networks [발표자료]
강사

김인중 교수, 한동대학교 전산전자공학부

ijkim@handong.edu

약력 Prof. In-Jung Kim, School of CSEE, Handong Global University
Ph.D in Computer Science, KAIST, Korea, 2001
M.S. in Computer Science, KAIST, Korea, 1995
B.S. in Computer Science, KAIST, Korea, 1994
초록 Convolutional neural networks (CNN) have shown outstanding performance in detection and recognition of visual targets. Comprising hierarchical structure, CNN is effective in learning mid- and high-level image representation. For further improvement, it is necessary to understand the content CNN learnt from training samples. Unlike encoder-decoder models, it is not straightforward to visualize and analysis the training result of CNN. In this talk, I will introduce a few recent works on CNN visualization and analysis. Then, I'll describe a method to improve the discrimination ability of CNN which was inspired by the visualization techniques.
Rotating Your Face Using Multi-task Deep Neural Network [발표자료] [추가자료]
강사

김준모 교수, 카이스트

약력 Assistant Professor, KAIST (2009 - current)
Research Staff Member, Samsung Advanced Institute of Technology (2005 - 2009)
Massachusetts Institute of Technology, Ph. D. (2005)
Massachusetts Institute of Technology, MS (2000)
Seoul National University, BS (1998)
초록 Face recognition under viewpoint and illumination changes is a difficult problem, so many researchers have tried to solve this problem by producing the pose- and illumination- invariant feature. Zhu et al. changed all arbitrary pose and illumination images to the frontal view image to use for the invariant feature. In this scheme, preserving identity while rotating pose image is a crucial issue. In this talk, we propose a new deep architecture based on a novel type of multitask learning, which can achieve superior performance in rotating to a target-pose face image from an arbitrary pose and illumination image while preserving identity. The target pose can be controlled by the user’s intention. This novel type of multi-task model significantly improves identity preservation over the single task model. By using all the synthesized controlled pose images, called Controlled Pose Image (CPI), for the pose-illumination- invariant feature and voting among the multiple face recognition results, we clearly outperform the state-of-the-art algorithms by more than 4~6% on the MultiPIE dataset.
Exploiting the Power of Deep Learning with Constrained Data [발표자료]
강사

백승욱 대표, 클디

http://www.cldi.io

약력 Co-founder and CEO of Cldi Inc.
PhD, MVLSI Lab., KAIST, 2014
MS. MVLSI Lab., KAIST, 2010
초록 In this talk, I will share the experiences and the practical challenges in applying deep learning technology to vertical applications. The first half of this talk will cover the case with a limited number of training data, which is common in real-world application. I will present our novel solution: multi-scale pyramid pooling(MPP) combining the discriminative power of CNN activation and Fisher kernel. In the second half, I will briefly introduce our recent achievements in fashion and diagnostic healthcare industries.
Deep Learning at NAVER [발표자료]
강사 김정희 수석연구원, 네이버
약력 석사, 서울대학교 전기공학부 컴퓨터비젼 computer vision, 1999
학사, 서울대학교 전기공학부, 1996
초록 Deep learning은 최근 기계학습분야에서 좋은 성능으로 주목받고 있다. NAVER 에서도 또한 deep learning 을 이용한 다양한 시도가 이루어지고 있다. 하지만 이를 활용한 서비스가 성공적으로 이루어지기 위해서는 풀어야 할 문제 역시 존재하며, deep learning 이 pattern recognition의 모든 어려운 점을 해결해 주는 절대반지는 아니다. 본 세션에서는 NAVER 에서는 성공적인 deep learning의 활용을 위해 어떤 노력을 기울이고 있으며, 앞으로 어떤 시도들이 이루어져야 하는지 이야기 하고자 한다.