Prof. Robert D. Nowak(University of Wisconsin-Madison)
Title: Deep Learning and Foundation Models: Theoretical Insights and Label-Efficient Learning
Abs :
This talk explores the intersection of deep learning theory with practical advancements in label-efficient learning. We begin by presenting a theory that characterizes the types of functions learned by ReLU networks, revealing that deep neural networks learn compositions of functions of bounded second-order variation in the Radon transform domain. This sheds new light on the role of weight decay regularization and provides principled approaches to more efficient training methods and network compression.
In the context of label efficiency, the shift toward large pre-trained foundation models brings new challenges. Traditional theories underpinning label-efficient methods, like semi-supervised and active learning, were designed for training classical models and rather than today’s fine-tuning strategies. Addressing this gap, the LabelBench project—a collaborative, open-source initiative—provides a testbed for optimizing label-efficient fine-tuning of vision models. Early results show that combining strategies can significantly boost label efficiency, revealing new approaches for training large models with limited labeled data.
Bio:
Robert Nowak is the Grace Wahba Professor of Data Science and Keith and Jane Nosbusch Professor in Electrical and Computer Engineering at the University of Wisconsin-Madison. His research focuses on machine learning, optimization, and signal processing. He serves on the editorial boards of the SIAM Journal on the Mathematics of Data Science and the IEEE Journal on Selected Areas in Information Theory.
Prof. Xiao Wang(Purdue University)
Title: Harnessing AI for Bayesian Inference: From Neural Conformal Inference to Neural Adaptive Empirical Bayes
Abs :
The fusion of artificial intelligence and Bayesian inference has unlocked new possibilities for tackling complex statistical challenges. This talk delves into two AI-driven methodologies that advance Bayesian inference with neural network capabilities. The first, Neural Conformal Inference, offers a likelihood-free approach that maps observed data to model parameters using deep neural networks. By circumventing traditional discretization errors and integrating conformal prediction, this new framework ensures rigorous uncertainty quantification and reliable posterior coverage. The second approach, Neural Adaptive Empirical Bayes, constructs flexible priors using implicit generative models and combines this with variational inference to optimize hyperparameters directly from data. This adaptive strategy enhances predictive accuracy and provides comprehensive uncertainty measures, addressing the challenges of high-dimensional, complex data structures.
Bio:
Education
2005 Ph.D. in Statistics, University of Michigan, Ann Arbor, MI
2000 M.S. in Mathematics, University of Science and Technology of China
1997 B.S. in Mathematics, University of Science and Technology of China
Research
AI
Machine Learning
Nonparametric Statistics
Functional Data Analysis
Reliability
Honors & Awards
Professional Achievement Award, Purdue University, 2022-2023
Elected Fellow of the American Statistical Association (ASA) 2021
Elected Fellow of the Institute of Mathematical Statistics (IMS) 2021
Regina and Norman F. Caroll Research Award, Purdue University 2017- 2018
Title: 지능형 업무 협업 솔루션(Brity Works + Copilot)
Abs :
본 강연에서는 생성형 AI 기반의 협업 솔루션이 기업의 업무를 어떻게 혁신할 수 있는지 살펴봅니다. 현재 삼성그룹 및 다양한 기업에서 사용 중인 협업 솔루션(브리티웍스)과 생성형 AI 서비스(브리티코파일럿)의 실제 적용 사례을 소개하고, 앞으로 Copilot 서비스가 Personal Agent 형태로 진화하게 되면, 어떠한 새로운 업무 경험을 가져오게 될지 함께 생각해봅니다.
Bio:
2023.12 ~ 현재 삼성SDS IW사업팀장 / 상무
2021.01 ~ 2023.11 삼성SDS C&C상품기획그룹 그룹장
2017.11 ~ 2020.12 삼성SDS Knox Portal사업그룹 상품 리더
2014.04 ~ 2017.10 삼성SDS 컨퍼런싱솔루션그룹 개발 리더
