Abstract: A central question in modern cryptography is how mutually distrusting parties can collaborate without compromising their privacy/security. This tutorial will introduce the problem and go over some basic definitions, constructions, and impossibility results. The focus will be on information theoretic security, i.e., security without relying on computational assumptions, and on stochastic resources shared by parties which enable protocols that achieve such security.

Biography: Vinod Prabhakaran received the M.E. degree from Indian Institute of Science in 2001 and the Ph.D. degree from the University of California at Berkeley in 2007. He was a post-doctoral researcher at the University of Illinois at Urbana–Champaign and the Ecole Polytechnique Fédérale de Lausanne, Switzerland. Since 2011, he has been with the School of Technology and Computer Science, Tata Institute of Fundamental Research, Mumbai, where he is now a Professor. He was a Distinguished Lecturer for the IEEE Information Theory Society during 2024 and 2025 and a TPC co-chair for ITW 2022. He serves as the area editor of security and privacy for the IEEE Transactions on Information Theory.

Abstract: Modern machine learning systems, particularly foundation models in vision, language, and multimodal domains, have achieved remarkable performance by leveraging massive datasets and large-scale computation. However, this paradigm entails substantial computational and environmental costs and is increasingly constrained by data scarcity, supervision quality, and limits in human expertise. These challenges have fueled growing interest in data- and supervision-efficient learning: how to train high-performing models using fewer, more informative data points and weaker or alternative forms of supervision.

This talk presents a unified theoretical perspective and practical algorithms for data- and supervision-efficient learning. We begin with the theoretical foundations of data efficiency in supervised and multimodal contrastive learning, emphasizing principled methods for identifying high-impact training samples that improve optimization and generalization. We then review recent advances in dataset distillation and synthetic data generation, showing how large datasets can be compressed into compact representations without sacrificing downstream performance.

We next turn to supervision efficiency in foundation models, discussing how machine-generated signals such as teacher predictions and synthetic data can serve as scalable alternatives to costly human supervision. In particular, we highlight knowledge distillation as a central mechanism for transferring knowledge from teacher to student models, enabling strong generalization across a wide range of teacher–student settings. We conclude by discussing emerging challenges in alignment and generalization as models approach or surpass human-level expertise, including weak-to-strong generalization as a promising paradigm for training stronger models using weaker supervision.

Biography: Hye Won Chung is an Associate Professor in the School of Electrical Engineering at KAIST, with joint appointments in the School of Computing and the Graduate School of AI. She received her Ph.D. and M.S. in Electrical Engineering and Computer Science from MIT and her B.S. from KAIST. From 2014 to 2017, she was a research fellow at the University of Michigan.

Her research lies at the intersection of information theory, data science, and machine learning, with a focus on the algorithmic and theoretical foundations of efficient, robust, and trustworthy AI systems. She is currently an IEEE Information Theory Society Distinguished Lecturer for 2025–2026. She delivered a tutorial at IEEE ISIT 2024 and has served as an Area Chair and Technical Program Committee member for major conferences including ICML, NeurIPS, and IEEE ISIT.

Abstract: Information theory, control theory, and network scheduling are traditionally three separate subjects in the general field of applied mathematical sciences for engineering. In this lecture, we will start by reviewing the basic concepts of each of these three pillars. Later, the discussion will shift towards combining information theory and control theory under the umbrella of the rate-cost function. In the third part of this lecture, we will formally establish the connections between rate-cost function and the new metric of Age-of-Infomration (AoI) in the network scheduling literature using the context of remote sensing and estimation, including the introduction of some latest results in the AoI literature.

Biography: Chih-Chun Wang is a Professor and the Associate Head of Facilities, Staff, and Planning of the Elmore Family School of ECE of Purdue University. He received the B.E. degree from National Taiwan University, and the M.S./Ph.D. degrees from Princeton University. He joined Purdue University in 2006 and became a Professor in 2017. He is an IEEE Fellow, and the author of 44 journal papers. He served as an associate editor of IEEE-IT in 2014 to 2017. He was the technical co-chair of the 2017 IEEE Information Theory Workshop, the general workshop co-chair of the Age and Semantics of Information Workshop of INFOCOM 2026, and Student Travel Grant Co-chair of ISIT 2023. His research interests focus on algorithmic design and performance analysis of wireless communication networks and network schedulers. His other research interests include stochastic control and optimization, iterative inference, channel and network coding, network information theory, feedback channels, distributed storage, coded caching, and security. Dr. Wang received the National Science Foundation Faculty Early Career Development (CAREER) Award in 2009.

Abstract: Information theory is fundamentally concerned with defining measures that quantify and distinguish information sources. In the classical setting, a rich family of such quantities exists, including relative entropy and Rényi divergences, each supported by clear operational interpretations—for instance, in hypothesis testing. 

In the quantum setting, however, the picture becomes significantly more intricate. Due to the non-commutative nature of quantum theory, there is no unique way to generalize classical divergences, leading instead to a variety of inequivalent candidates. Remarkably, even operational considerations single out different divergences as relevant in different contexts, for example in quantum hypothesis testing and channel coding. In this lecture, we will explore the consequences of this diversity and highlight structural properties that make certain divergences particularly well-suited for specific applications. Central to this discussion is the data processing inequality, which captures the fundamental principle that information cannot increase under noise. We will examine how different quantum divergences behave under data processing and show that not all are made equal in quantifying information loss—especially when considering strengthened forms of the data processing inequality. Participants will gain intuition for selecting appropriate divergences for different tasks, along with their respective advantages and limitations., we will start by reviewing the basic concepts of each of these three pillars. Later, the discussion will shift towards combining information theory and control theory under the umbrella of the rate-cost function. In the third part of this lecture, we will formally establish the connections between rate-cost function and the new metric of Age-of-Infomration (AoI) in the network scheduling literature using the context of remote sensing and estimation, including the introduction of some latest results in the AoI literature.

Biography: Christoph Hirche is a Junior Professor at Leibniz University Hannover, where he leads the research area “Quantum Learning and Algorithms.” His research focuses on quantum information theory, in particular on divergences, data processing inequalities, and the mathematical structure of quantum states and channels. He obtained his PhD from the Universitat Autònoma de Barcelona and held several postdoctoral positions before becoming a Marie Skłodowska-Curie Global Fellow at the National University of Singapore and the Technical University of Munich.

Abstract: Modern AI agents often need to interact continually with an unknown environment, creating a fundamental tension between exploration and exploitation. This tutorial discusses how information-theoretic tools can be used to quantify this trade-off and design algorithms that balance the two. A central theme is the interplay between learning a model of the environment and using that model to make decisions. We consider settings ranging from multi-armed and contextual bandits to Markov decision processes and beyond. The goal is to provide a rigorous and principled way to approach interactive decision-making problems.

Biography: Chen-Yu Wei is an Assistant Professor in the Department of Computer Science at the University of Virginia. Previously, he was a Postdoctoral Associate at MIT and a Research Fellow at the Simons Institute for the Theory of Computing. He received his Ph.D. in Computer Science from the University of Southern California. His research focuses on the theory of interactive machine learning, with emphasis on robust and adaptive learning in non-stationary environments, sample-efficient reinforcement learning with strategic exploration, and decentralized multi-agent learning. His work has been recognized with Best Paper Awards at the Conference on Learning Theory (COLT 2021) and the Conference on Algorithmic Learning Theory (ALT 2022).

Abstract: Sequential resource allocation lies at the heart of modern communications and networking, such as scheduling transmissions, managing queues, and minimizing information staleness. The Restless Multi-Armed Bandit (RMAB) provides a principled framework for these problems, but its exact solution is PSPACE-hard. The Whittle index policy resolves this through an elegant decomposition: assign a scalar priority index to each arm and activate the highest-ranked ones. This two-part tutorial develops the full story from foundations to the learning frontier.

Part 1 covers the mathematical formulation of restless bandits and the Whittle index, illustrated through networking applications in dynamic multichannel access, Age-of-Information scheduling, and multi-class queueing. We then examine extensions to heterogeneous multi-channel systems via partial indexability, and to contextual restless bandits with global side information.

Part 2 addresses the challenge of unknown transition dynamics via reinforcement learning. We survey two complementary approaches: treating the Whittle index as a break-even point (QWI, QWINN) and as the solution to an auxiliary optimal control problem (NeurWIN, DeepTOP, DIP). We compare these approaches in convergence speed, scalability, and multi-resource generality, and close with open research directions.

Biography: I-Hong Hou (Senior Member, IEEE) is a Professor in the ECE Department of the Texas A&M University. He received his Ph.D. from the Computer Science Department of the University of Illinois at Urbana-Champaign. His research interests include wireless networks, edge/cloud computing, and reinforcement learning. His work has received the 2025 IEEE Communications Society William R. Bennett Prize, the Best Paper Award from ACM MobiHoc 2017 and ACM MobiHoc 2020, and Best Student Paper Award from WiOpt 2017. He has also received the C.W. Gear Outstanding Graduate Student Award from the University of Illinois at Urbana-Champaign, and the Silver Prize in the Asian Pacific Mathematics Olympiad.