Abstract:

Humans convey their thoughts, emotions, and intentions through a concert of social displays: voice, facial expressions, hand gestures, and body posture. Despite advances in machine perception technology, machines are unable to discern the subtle and momentarynuances that carry so much of the information and context of human communication. The encoding of conveyed information by human body movements is still poorly understood, and a major obstacle to scientific progress in understanding human behavior is the inabilityto measure the full spectrum of social signals in groups of interacting individuals.

In this talk, I will describe my early exploration in building a sensor system, the Panoptic Studio equipped with more than 500 synchronized cameras, that can capture the wide spectrum of human social signaling---from voice, to facial expressions, to handgestures, to body posture---among groups of multiple people. Then, I will discuss my ongoing effort to build a system to perceive and understand human movements in 3D from the monocular videos captured in the wild.


Bio:

Hanbyul Joo is a Research Scientist at Facebook AI Research (FAIR), Menlo Park. His research is at the intersection of computer vision, graphics, and machine machine learning, focusing on building a system to perceive and understand humans in 3D. Hanbyul receiveda PhD in the Robotics Institute, Carnegie Mellon University. Hanbyul's research has been covered in various media outlets including Discovery, Reuters, NBC News, The Verge, and WIRED. He is a recipient of the Samsung Scholarship and the Best Student PaperAward at CVPR 2018.

Abstract:

Recent advances in both software and hardware opened a new horizon of robotics: artificial intelligence discovered dynamic motions in simulation and hardware became powerful enough to execute human-level stunts. However, the current state-of-the-art robots are yet far from operating in the real world due to lack of agility, robustness, efficiency, and safety. Therefore, we need to improve these features by building more intelligent control software and effective hardware mechanisms. However, this is a challenging problem that involves the optimization of control parameters, software architectures, and mechanical designs, where all the decisions jointly affect the motor capability of the robot. My research tackles these challenges by inventing novel optimization algorithms combined with prior knowledge and developing mathematical models for predicting the final performance of new robots. For instance, my on-going work includes learning methods for agile motions, interactive robot design software, and sim-to-real algorithms. In the long term, I aim to develop robotic companions in our home, search-and-rescue robots in disaster recovery scenes, and custom medical surgery robots that are tailored to individual patients.


Bio:

Sehoon Ha is currently an assistant professor at Georgia Institute of Technology. Before joining Georgia Tech, he was a research scientist at Google and Disney Research Pittsburgh. He received his Ph.D. degree in Computer Science from the Georgia Institute of Technology. His research interests lie at the intersection between computer graphics and robotics, including physics-based animation, deep reinforcement learning, and computational robot design. His work has been published at top-tier venues including ACM Transactions on Graphics, IEEE Transactions on Robotics, and International Journal of Robotics Research, nominated as the best conference paper (Top 3) in Robotics: Science and Systems, and featured in the popular media press such as IEEE Spectrum, MIT Technology Review, PBS News Hours, and Wired.

Abstract:

AI for the Real World
While deep learning has made remarkable progress over the recent years, deep-learning based AI systems are not yet fully ready to be deployed to any real-world applcations, due to lack of data, lack of resources, and safety/privacy concerns. In this talk, I will talk about some of my recent works that tackle these three practical challenges that arise when building practical AI systems.

In this talk, we will deliver recent developments in AI-based drug discovery. Since drug discovery is a very wide and complicated area of research, we will begin by explaining basic concepts and database resources on small-molecule drug or compound; target of drug; molecular signature before and after drug treatment; and phenotype such as drug sensitivity, toxicity, side effect, LADME (liberation, absorption, distribution, metabolism, and excretion). Then, in Part 2 of this talk, we will spent time to explain why AI-based drug discovery has emerged. Traditional drug discovery focused on predicting targets and phenotypes directly. Related research topics have been extensively investigated in the context of valid compound design, pharmacodynamics and pharmacokinetics. However, gap between compounds and phenotypes are big and wide. As molecular profiling techniques from genome and epigenome sequencing have been developed rapidly over the years, a relatively new concept called pharmacogenomics has emerged and has been extensively studied. In fact, information at the molecular level can be a bridge between compounds and phenotypes, which can be an innovative technology for drug discovery. However, computational analysis of data for drug discovery has become much more challenging since traditional concepts, such as valid compound design, pharmacodynamics and pharmacokinetics, already difficult computational problems and adding genomics dimension increases search space dramatically on top of already extremely large search space of the drug discovery problem. Fortunately, recent development of AI, deep learning, and graph mining technologies has begun to shed light on this daunting computational problem. In Part 3, we will introduce some of the representative examples of AI-based drug discovery technologies. A list of examples are: reinforcement learning for de novo molecule design, GAN and autoencoder for compound design, deep learning models for drug activity prediction, junction tree variational auto encoder for generating valid molecules, deep learning and symbolic AI for planning chemical syntheses, mixture representation learning for toxicity prediction, deep learning models for drug target interaction, GAN model for generating compounds from molecular biology data, and deep learning model for pharmacogenomcs study.

Short version:
AI technologies has begun to impact on the process of drug discovery. In this talk, I am going to introduce recent developments in AI-based drug discovery

빠르고 메모리 효율적인 딥러닝 모델은 모바일 디바이스에서 AI를 구현하기 위한 필수 구성 요소 중 하나다.
본 강연에서는 기존 딥러닝 모델의 정확도를 거의 그대로 유지하되 모델 크기를 대폭 줄이는 최신 경량화 연구를 소개한다.

Abstract:

본 강연은 딥 러닝 모델을 여러가지 가속기가 혼재하는 이종 클러스터에서 자동 병렬 실행하는 시스템에 대한 연구와 이를 Cloud 서비스에 이용하는 방법을 소개한다. 또, 추론과 학습에 있어서 TensorFlow XLA, TensorRT, TVM 등의 state-of-the-art 딥러닝 최적화 프레임워크를 능가하는 딥 러닝 최적화 기법의 연구에 대하여 소개한다. 딥 러닝 시스템의 성능 평가를 위하여 PyTorch나 TensorFlow와 같은 딥 러닝 프레임워크의 오버헤드 없이 순수하게 시스템의 성능을 측정할 수 있는 딥 러닝 벤치마크가 필요한데, 이를 위한 벤치마크 개발도 소개한다.

Abstract:

현재 딥러닝은 모델이 복잡해지고 (예: GPT-3), 학습 데이터는 증가하고 (예: OpenImage v6), 많은 학습과 추론을 처리해야 하는 추세이다.
본 강연에서는 큰 데이터, 큰 모델, 많은 요청을 처리하는 딥러닝 시스템 디자인시 어떤 문제들이 있는지 논의하고 최적의 딥러닝 시스템을 어떻게 설계할지 소개한다.
특히, 모델 변환없이 TensorRT 추론 시스템 보다 빠르게 추론을 수행하는 시스템과 큰 모델과 큰 데이터가 있는 경우 자동 분산 학습하는 시스템에 대해 소개한다.

Natural language has inherent structure. Words compose with one another to form hierarchical structures to convey meaning. While these compositional structures (such as parse trees) are crucial for mediating human language understanding, they are unobserved during human language acquisition. Yet, human learners have little trouble acquiring the syntax of their native language without explicit supervision. This has motivated the classical task of grammar induction, (i.e., data-driven discovery of syntactic structure from raw text), which has proven to be empirically difficult for artificial language learners. In this talk, I show how recent advances in model parameterization and inference can lead to improved computational tools for discovering syntactic structure from raw text.

Abstract:

Humans and animals learn to see the world mostly on their own, without supervision, yet today’s state-of-the-art visual recognition systems rely on millions of manually-annotated training images. This reliance on labeled data has become one of the key bottlenecks in creating systems that can attain a human-level understanding of the vast concepts and complexities of our visual world. Indeed, while computer vision research has made tremendous progress, most success stories are limited to specific domains in which lots of carefully-labeled data can be unambiguously and easily acquired.

In this talk, I will present my research in computer vision and deep learning on creating recognition systems that can learn disentangled visual representations with minimal human supervision. A key challenge is that, without strong supervision, deep networks can easily "cheat" and take undesirable shortcuts. However, given the right constraints, I’ll show that one can design learning algorithms that discover and generate meaningful structured disentangled representations from the data with little to no human supervision.


Bio:

Yong Jae Lee is an Associate Professor in the Computer Science Department at the University of California, Davis and an AI Visiting Faculty at Cruise. Before joining UC Davis in 2014, he was a Postdoctoral Fellow in the EECS Department at UC Berkeley (2013-2014) and Robotics Institute at Carnegie Mellon University (2012-2013), and obtained his PhD from the University of Texas at Austin in May 2012. Yong Jae is the recipient of several awards including the Army Research Office Young Investigator Program Award, NSF CAREER Award, UC Davis College of Engineering Outstanding Junior Faculty Award, numerous industry faculty awards, and the Most Innovative Award at the COCO Object Detection Challenge at ICCV 2019 for his team’s work on real-time instance segmentation. Yong Jae will join the Computer Sciences Department at the University of Wisconsin - Madison in Fall 2021.

본 발표에서는 언어 이해에 있어 지식을 활용하는 연구 사례에 대해 연세대학교 Data Intelligence 연구실의 최근 2년 연구를 바탕으로 소개한다. 관련 주제의 보다 자세한 정보는 http://dilab.yonsei.ac.kr/~swhwang에 있다.

Abstract:

Combinatorial structures arise naturally in modern machine learning problems including but not limited to representation learning with discrete generative factors, data mixup, and network pruning tasks. In this talk, I will discuss the latest research from SNUMLLAB (https://mllab.snu.ac.kr/) on efficient learning with combinatorial structures through the lens of principled combinatorial optimization.

Abstract:

본 강연에서는 자율주행을 실현하기 위해 필요한 컴퓨터 SW 기술들을 소개하고 그와 관련된 최신 연구들에 대해 논의한다. 구체적으로 자율주행 태스크들을 ECU에서 실시간 수행하기 위한 최적화 기술, 자율주행 기능을 설계단계에서부터 정확히 검증하기 위한 설계/검증 SW 기술, 전자적 쇼크, 방사능 등에 의한 Soft-Error에도 자율주행 안전성을 확보할 수 있는 신뢰성 기술 등에 대해 소개한다. 또한, 이러한 기술들을 적용해서 서울대학교에서 개발하고 있는 자율주행차량에 대해서도 소개한다.

Abstract:

In this talk, I am going to introduce upcoming challenges in building mobile deep learning systems for emerging extended reality applications. Also, I will present two specific systems recently developed in our group: (1) EagleEye to enable a person identification service in crowded urban spaces and (2) Heimdall to coordinate concurrent deep neural network tasks over resource-limited mobile GPUs.

Abstract

Reinforcement learning (RL) is a general-purpose machine learning framework, which considers an agent that makes sequential decisions in an environment to maximize its reward. Deep reinforcement learning approaches have made significant advances in the recent years by allowing the agent to learn a policy directly from raw observations. In this talk, I will first present how deep learning can be used to train RL agents to learn complex behaviors with several examples including AlphaStar which is the first AI to defeat top professional players in the game of Starcraft 2. Finally, I will present my latest work that attempts to meta-learn "how to learn" by interacting with environments, which can potentially open up entirely new approaches to RL by replacing hand-designed algorithms with meta-learned algorithms.


Bio

Junhyuk Oh is a senior research scientist at DeepMind. He received his Ph.D. from Computer Science and Engineering at the University of Michigan in 2018, co-advised by Prof. Honglak Lee and Prof. Satinder Singh, and B.S. from Computer Science and Engineering at Seoul National University in 2014. His research focuses on deep reinforcement learning problems, which include generalization, planning, multi-agent reinforcement learning, and meta-learning. Some of his work was featured at MIT Technology Review and Daily Mail. He served as a co-organizer of NIPS 2017 symposium on deep reinforcement learning, ICML 2018 workshop on exploration in reinforcement learning, and ICLR 2019 workshop on structures and priors in reinforcement learning.

Abstract:

Due to the advancement of generative adversarial networks, image generation and translation techniques are being actively developed, inspiring and accelerating humans' creative activities. In the light of this research direction, this talk will present my recent work on image-to-image translation and its applications to automatic image colorization.

Despite the rapid growth of object detection techniques, the current detection algorithms are still not robust enough to be applicable to the examples under certain conditions.
In this talk, I will introduce three well-known conditions that hinder the robust application of detection algorithms, and our proposed techniques to handle each problem.

This talk introduces two systems, Janus and Terra, that combine two different paradigms of deep learning frameworks. The systems achieve fast DL training by exploiting the techniques imposed by symbolic graph-based DL frameworks, while maintainingthe simple and flexible programmability of imperative DL frameworks at the same time.

Widely used deep learning frameworks heavily rely on the NVIDIAcuDNN library to improve performance. However, using cuDNN doesnot always give the best performance because of the following tworeasons: one is that it is hard to handle every case of versatileDNN models and GPU architectures with a library that has a fixedimplementation. We propose a deep learning optimization frameworkthat generates GPU kernel code considering both kernelimplementation parameters and GPU architectures. The proposedframework outperforms cuDNN/cuBLAS based implementation, and italso outperforms other state-of-the-art optimization frameworks.

Abstract: The attention mechanism is rapidly emerging as one of the most important key primitives in neural networks (NNs) for its ability to identify the relations within input entities. The attention-oriented NN models such as Google Transformer and its variants have established state-of-the-art on an extensive range of natural language processing tasks, and many other self-attention-oriented models are achieving competitive results in computer vision and recommender systems as well. Unfortunately, despite its great benefits, the attention mechanism is an expensive operation whose cost increases quadratically with the number of input entities that it processes, and thus accounts for a significant portion of the model runtime. This talk presents how hardware/software codesign enables the efficient acceleration of the attention mechanism.

Bio: Tae Jun is currently a postdoctoral researcher at Seoul National University. Tae Jun received his B.S.E degree from Duke University and received his masters and Ph.D. in Electrical Engineering from Princeton University under the supervision of Professor Margaret Martonosi. His main research area is hardware-software codesign for emerging applications and data access optimizations across systems and architectures. He is the recipient of Best Paper Award in MICRO-49, IEEE Micro Top Picks Honorable Mention in 2016, and Best Paper Award Nomination in ISPASS 2020. He is also the recipient of the Samsung Scholarship 2012-2017.

최근 그래프 신경망의 발전으로 그래프 구조를 갖는 데이터에 대한 신경망 학습이 주목받고 있다. 하지만 일반적인 그래프 신경망은 메세지 전달 방식의 연산을 이용하여 그래프 내 노드 표현 학습에 중점을 두기 때문에 그래프 데이터의 계층적 표현을 학습하는 것은 어렵다. 본 발표에서는 그래프의 구조적 정보와 노드 특징 정보를 활용하는 종단학습이 가능한 그래프 풀링 알고리즘을 통해 그래프의 계층적 표현을 학습하는 방법론에 대해 논의한다.

I introduce our recent works on deep decision networks which provide interpretable soft decisions on multivariate data

Pre-trained language models such as BERT and GPT have undoubtedly been at the center of recent advances in natural language processing,and there have also been many efforts to better utilize their power in various domains.

In this talk, I will briefly introduce a method that enables us to extract constituency parse trees from such strong models without furthertraining and illustrate how effective the induced trees are when compared against the trees induced from other unsupervised parsers.

Abstract:

Targeting global EDF scheduling, this we propose an optimal algorithm for parallelizing tasks with parallelization freedom. For this, we extend the interferencebased sufficient schedulability analysis and derive monotonic increasing properties of both tolerance and interference for the schedulability. Leveraging those properties, we propose a oneway search based conditionally optimal algorithm with polynomial time complexity.