패턴인식 및 기계학습 여름학교
2015년 8월 20일(목) - 21일(금), 고려대학교
미래창조과학부 기계학습 연구센터
정보기술의 발전과 더불어 도처에서 쏟아져 나오는 데이터를 효과적으로 처리하고 분석하기 위하여 다양한 분야에서 패턴인식 및 기계학습 이론의 필요성이 대두되고 있습니다. 이에 한국정보과학회 인공지능소사이어티에서는 2015년 8월 20일-21일 양일에 걸쳐 "2015 패턴인식 및 기계학습 여름학교"를 개최합니다.
국내외적으로 이 분야에서 활발하게 활동하시는 여섯 분의 연사를 모시고 알찬 강의를 준비해 보았습니다. 여름학교를 통해서 보다 더 깊은 지식을 습득하여 각자의 분야에서 유용하게 활용할 수 있는 기회의 장이 되었으면 합니다. 많은 분들이 여름학교에 참가하여 의미 있는 시간이 될 수 있기를 기대합니다.
한국정보과학회 인공지능소사이어티 회장 이성환
황성주(UNIST)
오일석(전북대학교), 이근배(POSTECH), 장병탁(서울대학교)
▲ 맨 위로
▶ Convex Optimization
신현정 교수, 아주대학교
내용
Convex optimization is an important enough topic that everyone who learns machine learning should know at least little bit about it. Many problems in machine learning are based on finding parameters that minimize some objective function. Very often, it is a weighted sum of two components: a cost term and a regularization term. If both of these components are convex, then the problem is a convex optimization problem. There are great advantages to recognizing or formulating a problem as a convex optimization problem. Most importantly, if a function is strictly convex, it is guaranteed to have a unique global minimum, and it can be solved, very reliably and efficiently, using standard methods. There are also theoretical or conceptual advantages that the associated dual problem, for example, often has an interesting interpretation in terms of the original problem, and sometimes leads to an efficient method for solving it. Typical examples of convex optimization problems in machine learning include support vector machines, semi-supervised learning, ridge regression with Tikhonov regularization, whereas neural networks, maximum likelihood mixtures of Gaussians are non-convex problems. In this talk, we give an overview of mathematical optimization, focusing on the special role of convex optimization, and then describe the convex programming formulations of widely known machine learning algorithms.
약력
2006 - 현재 : 아주대학교 공과대학 산업공학과 부교수
2005 - 2006 : Friedrich-Mierscher-Lab, Max-Planck-Institute(독일) 수석연구원
2000 - 2005 : 서울대학교 공과대학 (산업공학/데이터마이닝) 공학박사
▶ Recommender Systems
유환조 교수, POSTECH
내용
Recommender systems have high demands and various applications in industry. Developing an accurate and efficient recommender system is technically challenging thus actively researched in research communities. This talk first introduces fundamental concepts and techniques in recommender systems, and presents several practical issues in recommender systems such as novel recommendation, timing when to recommend, and fast parallel matrix factorization for big data.
약력
2008 - 현재 : POSTECH 컴퓨터공학과 교수
2004 - 2007 : 조교수, University of Iowa, USA
1998 - 2004 : PhD, Computer Science, University of Illinois at Urbana-Champaign, USA
▶ Bayesian Recommendation
최승진 교수, POSTECH
내용
Collaborative prediction is a popular technique used in recommender systems, the task of which is to fill in missing entries of a user-item matrix given a small set of observed entries, to predict preferences of users based on their observed preferences. In this tutorial, I present Bayesian models for recommender systems, what I called, “Bayesian recommendation”. To this end, I introduce Bayesian treatment of matrix factorization, especially emphasizing variational Bayesian matrix factorization as well as co-factorization. In addition, I also introduce a Bayesian binomial mixture model for collaborative prediction in the case of non-random missing world.
약력
2001 - 현재 : POSTECH 컴퓨터공학과 교수
1990 - 1996 : Ph.D. in EE, University of Notre Dame
▶ Structural SVMs
김건희 교수, 서울대학교
내용
Support Vector Machine (SVM) is one of most popular supervised learning algorithms for classification and regression, which aims to find a hyperplane that separates different classes of data with the maximal margins. Structural SVMs generalize the SVMs to handle arbitrary output spaces with non-trivial structure. Structural SVMs have been successfully applied to a large range of structured output prediction tasks, including diversified ranking, multi-class classification, parse tree generation, part-of-speech tagging, protein sequence alignment, image segmentation, and human pose estimation. This tutorial will provide a comprehensive introduction to the principles and algorithms for Structural SVMs, and their practical applications in computer vision and visual data mining.
약력
2015 - 현재 : 서울대학교 공과대학 컴퓨터공학과 조교수
2013 - 2015 : 박사후 연구원, Disney Research
2009 - 2013 : Ph.D. in Carnegie Mellon University, Computer Science Department
▶ Structured Sparsity
황성주 교수, UNIST
내용
Sparsity-inducing regularizations such as l1-regularizations allow to select small number of variables that are useful for a given learning task, by zeroing out most of the variables. They have been extensively explored in the last decades in both machine learning and computer vision field, as it enables to efficiently compute and compactly store the model, as well as analyze important factors in high-dimensional learning. In this lecture, we will go over the basics of norm-based regularization, including l1, l2, and l1/l2-norms, as well as the optimization techniques for them, while learning about statistical models that leverages sparsity such as lasso, elastic net, sparse group lasso, and tree-guided group lasso. Further, as a case study, we will focus on the problem of sparse coding and dictionary learning, and number of applications for it, such as multitask/transfer learning, and image denoising. At the end of the talk, we will also go over few sparsity-inducing regularizations for deep learning models.
약력
2014 - 현재 : UNIST 컴퓨터공학과 조교수
2013 - 2014 : 박사후 연구원, Disney Research
2013 : Ph.D. in Computer Science, University of Texas, Austin, USA
강의자료 다운로드
▶ Deconvolutions in Convolutional Neural Networks
한보형 교수, POSTECH
내용
Convolutional neural networks have shown remarkable performance in various problems in computer vision, and some researchers have beein interested in understanding the learned features using deconvolutional networks. Recently, deconvolution starts to be employed in several practical problems since deep networks with deconvolution are suitable to predict structural outputs in principle. There are several successful results based on deconvolutions in spite of challenges in training deconvolution layers such as a large number of parameters and a limited amount of resources. In this lecture, I introduce recent breakthroughs related to deconvolution networks, which include feature visualization, semantic segmentation, object generation, and image enhancement.
약력
2014 - 현재 : POSTECH 컴퓨터공학과 부교수
2010 - 2014 : POSTECH 컴퓨터공학과 조교수
2008 - 2009 : 연구원, Mobileye Vision Technologies, NJ
강의자료 다운로드
고려대학교 자연계캠퍼스 우정정보통신관 601호 (자세히 보기)
