Korean AI Association

  >   학술행사   >   국내학술대회


기조 & 초청강연
Plenary Talk 


Prof. Stephen G. Walker(University of Texas at Austin)
Title : Martingale Posterior Distributions
The talk will present a fundamental alternative to the thinking behind Bayesian analysis. Rather than update a prior to posterior via a likelihood function, the idea is to regard Bayesian uncertainty as being driven by what has not been seen. Usually this is the observations indexed from the sample size plus 1 to infinity. Using a density estimator for the next observation, this is sampled and updated, and iterated, the long term process generating a random parameter which can be regarded as a sample from a posterior. Indeed, under special cases, Bayesian posteriors can be sampled in this way. The framework allows for straightforward generalizations of the Bayesian approach which in particular can relax the strict assumption of exchangeability.
Stephen G. Walker is a full professor of Mathematics and SDS (Statistics & Data Science) at the University of Texas at Austin, having previously held positions at the University of Kent, Bath, and Imperial College, in the UK. He is currently executive editor for Journal of Statistical Planning and Inference, and associate editor for Journal of the American Statistical Association, having previously been associate editor for Scandinavian Journal of Statistics, Statistica Sinica, and Annals of Statistics. He has received an EPSRC Advanced Research Fellowship, 2001–2006, and has been funded by the NSF. Dr. Walker's research is predominately in Bayesian methods, including nonparametric, asymptotics, computational and methodological and foundational. Other areas of interest include hypothesis testing, inequalities, matrix, and linear algebra.

Special invited talk 1


Prof. Nhat Ho(University of Texas at Austin)
Title : Optimal Transport in Large-Scale Machine Learning Applications
From its origins in the seminal works by Monge and Kantorovich in the eighteenth and twentieth centuries, and through to the present day, the optimal transport (OT) problem has played a determinative role in the theory of mathematics and recently machine learning and data science. In the current era, the strong and increasing linkage between optimization and machine learning has brought new applications of OT to the fore. However, in these large-scale data-driven applications, there have still remained two major challenges: (1) Computational, namely, the OT is computationally expensive when the number of data is large; (2) Curse of dimensionality, namely, the required sample size for estimating the true underlying distribution of the data scales exponentially with the dimension. In this talk, we address these challenges via proposing several new notions of mini-batches and sliced optimal transport.
For the computational challenge, the mini-batch optimal transport (m-OT) has been successfully
used in practical applications that involve probability measures with intractable density, or probability measures with a very high number of supports. Despite its scalability advantage, m-OT does not consider the relationship between mini-batches which leads to undesirable estimation. Moreover, m-OT suffers from misspecified mappings, namely, mappings that are optimal on the mini-batch level but are partially wrong in the comparison with the optimal transportation plan between the original measures. To address these issues, we propose novel mini-batching scheme for optimal transport, named Mini-batch Partial Optimal Transport (m-POT). The m-POT utilizes partial optimal transport (POT) between mini-batch empirical measures. Finally, we carry out extensive experiments on various applications such as deep domain adaptation, partial domain adaptation, deep generative model, color transfer, and gradient flow to demonstrate the favorable performance of m-POT compared to current
mini-batch methods.

For the curse of dimensionality challenge, the sliced Wasserstein (SW) distance has been widely used to address this challenge. The value of sliced Wasserstein distance is the average of transportation cost between one-dimensional representations (projections) of original measures that are obtained by Radon Transform (RT). Despite its efficiency in the number of supports, estimating the sliced Wasserstein requires a relatively large number of projections in high-dimensional settings. Therefore, for applications where the number of supports is relatively small compared with the dimension, e.g., several deep learning applications where the mini-batch approaches are utilized, the complexities from matrix multiplication of Radon Transform become the main computational bottleneck. To address this issue, we propose to derive projections by linearly and randomly combining a smaller number of projections which are named bottleneck projections. We explain the usage of these projections by introducing
Hierarchical Radon Transform (HRT) which is constructed by applying Radon Transform variants recursively. We then formulate the approach into a new metric between measures, named Hierarchical Sliced Wasserstein (HSW) distance. Finally, we compare the computational cost and generative quality of HSW with the conventional SW on the task of deep generative modeling using various benchmark datasets including CIFAR10, CelebA, and Tiny ImageNet.
Nhat Ho is currently an Assistant Professor of Data Science and Statistics at the University of Texas at Austin. He is a core member of the University of Texas Austin Machine Learning Laboratory and senior personnel of the Institute for Foundations of Machine Learning. His current research focuses on the interplay of four principles of machine learning and data science: interpretability of models (deep generative models, convolutional neural networks, Transformer, model misspecification), stability, and scalability of optimization and sampling algorithms (computational optimal transport, (non)-convex optimization in statistical settings, sampling and variational inference, federated learning), and heterogeneity of data (Bayesian nonparametrics, mixture and hierarchical models).
Special invited talk 2


최진호 교수(Emory University)
Title : Controllable AI: Lessons from the Past & The Future of Conversational AI
Abs :
In this talk, I will share our winning experience in the Alexa Prize Socialbot Grand Challenge 3, an international university competition to build a chatbot capable of engaging in 20-minute conversations with anyone on any topic. Through this experience, I will highlight key lessons learned in the realm of human-computer interaction. I will then address the challenges and limitations faced by current deep learning-based chatbots and present a new comprehensive platform to evaluate multi-turn human-to-chatbot conversations in open-domain settings. Finally, I will introduce our integration of cutting-edge large language models with our previous technology to develop a Controllable Conversational AI platform. This platform aims to advance education by assisting college students in enhancing their academic success and career development. Moreover, we are developing a chatbot to provide support for individuals with mental health issues, such as trauma, by offering a service that helps them recognize traumatic events and diagnose related conditions. The talk will underscore the significance of Controllable AI and outline our future directions in this field.
Dr. Jinho Choi is an associate professor of Computer Science, Quantitative Theory and Methods, and Linguistics at Emory University. He received his BA in Computer Science and Mathematics (dual degree) from Coe College in 2002, MS in Computer and Information Science from the University of Pennsylvania in 2003 (advisor: Mitch Marcus), PhD in Computer Science and Cognitive Science (joint degree) from the University of Colorado Boulder in 2012 (advisor: Martha Palmer), and post-doctoral training at the University of Massachusetts Amherst in 2014 (advisor: Andrew McCallum). He was a full-time lecturer of Computer Science at the Korea Military Academy from 2004 to 2007 while he served his military duty in South Korea. He is the founder and the director of the Natural Language Processing Research Laboratory at Emory University.


Dr. Choi has been active in the field of Natural Language Processing (NLP). He has presented many state-of-the-art NLP models that automatically derive various linguistic structures from plain text. These models are publicly available in the NLP Toolkit called ELIT. He has also led the Character Mining project and introduced novel machine comprehension tasks for explicit and implicit understanding in multiparty dialogue. For the application side, Dr. Choi has developed innovative Biomedical NLP models by collaborating with several medical fields such as radiology, neurology, transplant, and nursing. His latest research focuses on building the conversational AI-based chatbot called Emora that aims to be a daily companion of everyone's life. With Emora, Dr. Choi's team won the 1st-place at the Alexa Prize Socialbot Grand Challenge 3 that came with $500,000 cash award.




성낙호 하이퍼스케일 AI 기술총괄(네이버클라우드)


Title :  Hyperscale AI, 우리의 경쟁력



전기의 발명에 비유될 만큼, 우리 삶을 크게 바꿔 놓으리라 기대되는 인공지능 기술에 대한 산업 현장에서의 이해와 향후 전망에 대해 이야기 합니다.



서울대학교 컴퓨터공학과 졸업
1999년 ~ 2017년 : 엔씨소프트(부장), 레드덕(디렉터), 헥스플렉스(CTO)
2017년 ~ 2022년 : 네이버 클로바 책임리더/이사
2023년 현재        :  네이버 클라우드 Hyperscale AI 기술총괄
                                네이버 클라우드 Hyperscale AI 이사
                                네이버 책임리더