일시 | 2021년 2월 16일(화) - 19일(금)
장소 | 온라인
주관 | 정보과학회 인공지능 소사이어티 , 고등과학원 거대수치계산연구센터
조직위원 | 최승진, 신현정, 최재식, 김세훈, 노영균, 손영우(고등과학원)
등록인원 | 제한없음
참가비 | 일반 : 200,000원, 학생(대학원생) : 100,000원
등록페이지 | 정보과학회 홈페이지 (고등과학원 소속은 개별 접수 (firstname.lastname@example.org))
#1. Recent advances in Autoregressive models and VAE (Part 1) [2/16, 10:00-12:00] #2. Recent advances in Autoregressive models and VAE (Part 2 & hands-on practice) [2/16, 14:00-16:00]
2020-현재, 카카오브레인 연구원 2017-2019, AITRICS 연구팀장 2018, POSTECH 컴퓨터공학 박사 2012, Microsoft Research Asia, Microsoft Research, 연구 인턴 2009, POSTECH 컴퓨터공학 학사
Abstract: 본 튜토리얼에서는 likelihood를 직접적으로 모델링할 수 있는 자기 회귀 (autoregressive) 모델과 변분 오토인코더 (variational autoencoder)의 최신 연구 결과를 포괄적으로 소개하며 간단한 자기 회귀 모델 실습을 진행할 계획입니다. 세부 튜토리얼 내용은 아래와 같습니다. - CNN기반 autoregressive 모델 (PixelCNN++, PixelSNAIL 등, 40분) - Transformer 기반 autoregressive 모델 (Image Transformer, Axial Transformer 등, 40분) - Hierarchical VAE 최신 연구 결과 (VDVAE 등, 40분) - Autoregressive 응용 (Colorization Transformer 및 Image GPT, 1시간) -간단한 autoregressive model 실습 (1시간) (실습 관련해서는 등록자에게 별도 안내 드릴 예정입니다.)
#3. Reinforcement Learning [2/17, 10:00-12:00]
Yoonsuck Choe is a Professor and Director of the Brain Networks Laboratory at Texas A&M University (2001-present). He received his Ph.D. degree in computer science from the University of Texas at Austin in 2001, and his B.S. degree in computer science from Yonsei University in 1993. His research interests are in neural networks and computational neuroscience, and he published over 100 papers on these topics, including a research monograph on computations in the visual cortex. He served as program chair and general chair for IJCNN (2015 and 2017, respectively), and served on the editorial board of the INNS journal Neural Networks. From 2017 to 2019, he served as head of the Machine Learning Lab and later as the head of the AI Core Team at Samsung Research AI Center in Seoul.
Abstract: This tutorial will go over the basic concepts of reinforcement learning and provide detailed coverage on foundational algorithms in the field such as Q learning. The basic introduction will be followed by a brief discussion of deep reinforcement learning algorithms (e.g. DQN), and the applications of reinforcement learning in sensorimotor control.
#4. Neural Processes [2/17, 14:00-16:00]
Dr. Juho Lee is an assistant professor in the graduate school of AI at KAIST. He received his Ph.D. degree in computer science from POSTECH and did his postdoc at the University of Oxford. After finishing his postdoc, he worked as a research scientist at AITRICS, an AI-based health-care startup. His research is mainly focused on Bayesian learning, ranging from Bayesian nonparametric models to Bayesian deep learning.
Abstract: Stochastic processes are widely used in machine learning as prior distributions over functions, applications including regression, classification, Bayesian optimization, bandit, and so on. Typically we posit a class of stochastic processes derived from mathematical objects (e.g., Gaussian processes), and hope it well describes the data at hand. On the other hand, neural processes implicitly define stochastic processes with a minimal assumption on the characteristics of functions to be generated, and let the model learn from examples. This tutorial describes how such neural processes are defined, trained, and employed for real-world problem-solving. The tutorial will cover several variants of neural processes, ranging from vanilla neural processes and recently developed convolutional neural processes.
#5. Adversarial Attacks: Why Are Machine Learning Algorithms Vulnerable to Attacks? [2/18, 10:00-12:00]
Dr. Hamm is an Associate Professor of Computer Science at Tulane University since 2019. He received his PhD degree from the University of Pennsylvania in 2008 supervised by Drs. Daniel Lee and Lawrence Saul. Dr. Hamm's research interest is in machine learning, from theory and to applications. He has worked on efficient algorithms for adversarial machine learning, deep learning, privacy and security, optimization, and nonlinear dimensionality reduction. Dr. Hamm also has a background in biomedical engineering and has worked on medical data analysis, computational anatomy and on modeling human behaviors. His approach can be summarized as using machine learning to find novel solutions for challenging problems in the applied fields. His work in machine learning has been published in top venues such as ICML, NIPS, CVPR, JMLR, and IEEE-TPAMI. His work has also been published in medical research venues such as MICCAI, MedIA, and IEEE-TMI. The academic community has recognized his contributions; among other awards, he has earned the Best Paper Award from MedIA (2010) and Google Faculty Research Award (2015).
Abstract: The current generation of machine learning algorithms exhibit impressive performance on various learning tasks often exceeding the human performance. However, state-of-the-art models are often trained with datasets which are relatively small compared to the millions of tunable parameters in a model. Consequently, an imperceptibly small deviation of the test example from the training distribution results in abject failure of the state-of-the-art models. Adversaries can exploit such a vulnerability to prevent learning-based models such as autonomous vehicles from being deployed in the real world. This tutorial aims to provide insights on the cause of the vulnerability and present an up-to-date review of papers on adversarial attacks and defenses.
#6. Pre-trained Language Models for NLP [2/18, 14:00 - 16:00]
2020-present: EXO-BRAIN National Strategy Project Leader 2005-present: Principal Researcher at ETRI 2016: Ph.D in CSE at Chungnam National Univ. 2002, 2005: B.S. and M.S. in CS at Korea Univ.
Abstract: 본 튜토리얼은 2018년 이후 언어처리의 주요 흐름으로 자리 잡은 딥러닝 사전학습 언어모델(BERT, GPT 등)에 대해 소개하며, 세부 내용은 다음과 같다. - Representation Learning for NLP - Transformer 모델 (CNN 및 RNN과 비교하여 특징 및 장점) - BERT 및 GPT 사전학습 언어모델 - BERT 및 GPT 이후 언어모델 연구 - Text-To-Text Transfer Transformer(T5) 언어이해생성 모델 - GPT-3와 퓨샷학습 본 튜토리얼은 딥러닝 또는 자연어처리 입문자부터를 대상으로 한다.
#7. 딥러닝과 HPC [2/19, 10:00-12:00]
2020.03. ~ 현재, 서울대학교 데이터사이언스대학원 교수, 학생부원장 2002.09. ~ 현재, 서울대학교 공과대학 컴퓨터공학부 조교수, 부교수, 교수 2019.01. ~ 현재, IEEE Fellow 2009.04. ~ 2018. 02., 매니코어 프로그래밍 연구단 단장(과학기술정보통신부 창의연구사업) 2000.01 ~ 2002.08, Michigan State University, Dept. of Computer Science and Engineering, Assistant Professor
Abstract: 최근 좋은 성능을 내고 있는 AI 또는 딥 러닝 응용은 학습을 수행할 때 수백, 수천 대의 GPU 컴퓨터 시스템을 빠른 네트워크로 엮은 슈퍼컴퓨터 클러스터와 같은 컴퓨팅 자원을 요구하고 있는 것이 보통이다. 본 강연은 이와 관련하여 고전적 High Performance Computing(HPC)에서 주로 사용하는 이종 컴퓨팅에 대하여 살펴본다. 또, 현재 딥 러닝 응용을 실행하는 딥 러닝 플랫폼의 발전 동향을 살펴보고 이와 고전적인 High Performance Computing의 관계를 알아본다. 이를 바탕으로 주어진 딥 러닝 응용을 자동으로 병렬처리하는 딥 러닝 프레임워크의 연구 방향에 대하여 살펴본다.
2021-present: Executive Advisor of BARO AI 2019-2020: CTO of BARO AI 2001-2019: Professor of Computer Science, POSTECH 1997-2001: Assistant Professor, Chungbuk National University 1997: Frontier Researcher in RIKEN, Japan 1996: Ph.D. in EE at University of Notre Dame, Indiana, USA
Abstract: Gaussian process (GP) is a stochastic process, which defines a distribution over functions that can be used for Bayesian regression. GP regression has been widely used in machine learning, since it provides a principled way to infer prediction uncertainty. It is also a core ingredient in Bayesian optimization, serving as a probabilistic surrogate model to infer posterior means and variances of the black-box function at particular points. In this tutorial, I begin with the introduction of stochastic processes, in particular, in the perspective of Bayesian nonparametric models. Then I will explain basic concepts of Gaussian processes and its application to Bayesian regression. A drawback of the vanilla GP is in its poor scalability since it requires the inverse of a kernel matrix whose dimension is determined by the number of training examples. Sparse GP regression, which is also known as low-rank GP regression, improves the scalability, which will be covered in this tutorial.