UPcoming schedule
Generalization in Deep Learning through the Lens of Kernels
Speaker: Wei Hu, postdoc at UC Berkely (assistant professor at U of Michigan, starting Fall 2022)
Date: Feb. 25, Friday, 1PM
Abstract: The empirical success of deep learning has posed significant challenges to classical generalization theory, due to the excess capacity of overparameterized neural networks. This talk is about how studying kernels (or equivalently, overparameterized linear models) can contribute to the theoretical understanding of generalization in deep learning. First, I will review the theory of neural tangent kernels (NTKs), which is formally related to a class of wide neural networks. I will touch on some of the insights we can obtain from this theory as well as its limitations. Next, we will examine the generalization theory for kernels, using several real-world settings (such as the NTK corresponding to a pre-trained ResNet applied to CIFAR-10) as a testbed. We find that, even in these linear settings, most existing theoretical analyses fail to qualitatively capture the empirical generalization phenomena; on the other hand, a random matrix theory perspective gives rise to an estimator that accurately predicts the generalization.
Bio: Wei Hu is a postdoctoral scholar at the University of California, Berkeley and an incoming assistant professor at the University of Michigan (starting in Fall 2022). He obtained his Ph.D. degree from Princeton University in 2021, and his B.Eng. degree from Tsinghua University in 2016, both in Computer Science. He is broadly interested in the theoretical and scientific foundations of modern machine learning, in particular deep learning and related topics. He aims to obtain a solid, rigorous, and practically relevant theoretical understanding of machine learning methods, as well as to develop principles to make machine learning systems better in terms of reliability and efficiency.
Adaptive Sampling for Exploration in Bandits and RL
Lecturer: Alexandre Proutiere, professor at KTH, the Royal Institute of Technology
Date: Feb.14 - 17, Monday - Thursday, 3-6PM
Lecture notes: lec1, lec2, lec3, lec4, lec5, lec6, lec7
Video: link
Abstract: In this short course, we review recent theoretical advances in the design of adaptive sampling methods for exploration in various types of Reinforcement Learning problems. By adaptive, we mean that the data samples sequentially gathered by the learner are decided in an online manner depending on the observations made so far. These observations are leveraged to design future exploration steps, and overall, their use boosts the learning process. We investigate adaptive sampling methods for regret minimization and best policy identification in stochastic bandits, Markov Decision Processes (MDPs), and controlled linear systems. We first present generic and unified techniques to derive problem-specific information-theoretical limits for these problems. The fundamental limits specify the sampling strategy that leads to the most efficient learning process, and in turn, are instrumental in the design of adaptive exploration algorithms. The course outline is as follows:
Introduction to learning problems: bandits, learning in MDPs and linear systems
Learning algorithms and their performance: regret, sample complexity, minimax vs. instance-specific guarantees
Information-theoretical limits through the change-of-measure argument
Concentration inequalities
Regret minimization in unstructured and structured bandits
Sample complexity for best arm identification in bandits
Learning in MDPs
Open problems and future research directions
Bio: Alexandre Proutiere’s research evolves around theoretical aspects of learning and control of complex dynamical systems. He received the degree in mathematics from École Normale Supérieure, Paris, France, the degree in engineering from Télécom ParisTech, Paris, France, and the Ph.D. degree in applied mathematics from École Polytechnique, Palaiseau, France, in 2003. He is an Engineer from Corps of Mines. In 2000, he joined France Telecom Research and Development as a Research Engineer. From 2007 to 2011, he was a permanent researcher with the Microsoft Research, Cambridge, U.K. He is currently a Professor with the Department of Decision and Control Systems at KTH, Royal Institute of Technology, Stockholm, Sweden. He was a recipient in 2009 of the ACM Sigmetrics Rising Star Award and the Best Paper Awards at ACM Sigmetrics Conference in 2004 and 2010, and the ACM Mobihoc Conference in 2009. He was an Associate Editor of the IEEE/ACM Transactions on Networking and of the IEEE Transactions on Information Theory. He is currently an Editor of JMLR and an area chair at NeurIPS.
Full schedule
Label-Efficient Deep Learning
Kihyuk Sohn, Research Scientist at Google Cloud AI
Jan. 17, Monday, 10 AMCurrent Trends in Drug Discovery with Machine Learning and Computational Chemistry
Seongok Ryu, research scientist at Galux
Feb. 3, Thursday, 2 PMUnderstanding Infinite-Width Deep Neural Networks
Jaehoon Lee, Senior Research Scientist at Google Brain
Feb. 4, Friday, 10AMModeling information flow in Online Social Media using Hawkes Point Processes
Marian-Andrei Rizoiu, senior lecturer (associate professor) at University of Technology Sydney
Feb. 7, Monday, 2PMGraph Neural Networks for Subgraph Matching
Qing Wang, professor at Australian National University
Feb. 8, Tuesday, 2PMBenchmarks for Unsupervised Reinforcement Learning and Preference-based Reinforcement Learning
Kimin Lee, Researcher at Google Brain
Feb. 9, Wednesday, 1PMRobust and Efficient Federated Learning
Sai Praneeth Karimireddy, Postdoc at UC Berkely
Feb. 10, Thursday, 10AMShortcut learning in Machine Learning: Challenges, Examples, Solutions
Sanghyuk Chun, research scientist at Naver AI Lab
Feb. 10, Thursday, 2PMEfficient Deep Learning on Model, Data, Label and Beyond
Zhiqiang Shen, professor at HKUST
Feb. 11, Friday, 5PMAdaptive Sampling for Exploration in Bandits and RL (special lecture)
Alexandre Proutiere, professor at KTH
Feb. 14 - 17, Monday - Thursday, 3-6PMRecent Advances in Text-to-Image Generation Models
Saehoon Kim, researcher at Kakao Brain
Feb. 21, Monday, 4PMGeneralization in Deep Learning through the Lens of Kernels
Wei Hu, postdoc at UC Berkely (assistant professor at U of Michigan, starting Fall 2022)
Feb. 25, Friday, 1PM