Talks

Speakers

Brief Introduction

Professor Anand D. Sarwate (Distinguished Lecture of ITSoc)

Bio: I am an Associate Professor in the Department of Electrical and Computer Engineering at Rutgers, The State University of New Jersey. I also am a member of the graduate faculty in the Department of Computer Science and the Department of Statistics. I like to work on problems that involve probability, mathematical statistics, and optimization, with applications in information theory, communication, signal processing, and machine learning. I’m particularly interested in how these things intersect in the context of distributed/decentralized systems with constraints like privacy, bandwidth, latency, power, and so on.

Area of Expertise: Distributed optimization and signal processing, machine learning and statistics, information theory, and privacy-preserving data analysis.

Title: Are modern ML models like scientific instruments?

Abstract: The landscape of machine learning evolves rapidly and the complexity of the networks and their architectures defies easy comprehension. AI is touted as the next “scientific revolution” by allowing the processing and pattern-finding in increasingly massive data sets. The fine-tuning paradigm assumes that the features extracted by large pre-trained models are useful for many downstream tasks, which is similar to data collected from “classical” measurement devices such as cameras or medical scanners. While functionally playing similar roles, our understanding for and requirements of scientific instruments is quite different. Digging into these differences turns up some interesting fundamental things we still need to understand about training processes, how to compare models, and how feature embeddings can be used. While the results in this talk are largely empirical, they point to interesting directions for (infomation?) theoretical investigation.

Professor Shancheng Zhao

Bio: Shancheng Zhao received the bachelor’s degree in software engineering and Ph.D. degree in communication and information systems from Sun Yat-sen University, Guangzhou, China, in 2009 and 2014, respectively. From 2013 to 2014, he was a Graduate Visiting Student with the University of California at Los Angeles, Los Angeles, CA, USA. He is currently a Professor and Vice Dean of the College of Information Science and Technology, Jinan University, Guangzhou. His current research interests include finite-length codes, spatially coupled codes, and their applications. He was a co-recipient of the Best Paper Award at the IEEE GlobeCom. He serves as an Associate Editor for IET Quantum Communication, Physical Communication (Elsevier), and Alexandria Engineering Journal.

Title: Multiply-Chained Product Code: Construction, Design, and Decoding

Abstract: We introduce the multiply-chained product-like codes in this talk, including its construction, design, and decoding. We first introduce and analyze the multiply-chained zipper codes. We then introduce and analyze the GII-zipper codes. We finally introduce an efficient decoder for product-like codes based on random flipping. 

Professor Suihua Cai

Bio: Cai Suihua is an associate professor at the School of Computer Science, Sun Yat-sen University. He obtained his Ph.D. in Information and Communication Engineering from Sun Yat-sen University in 2019. His research interests are in information theory, channel coding and their applications. In recent years, he has published more than 20 papers in journals such as Transaction on Information Theory and Transactions on Communications, as well as important information theory conferences like ISIT (International Symposium on Information Theory). He has also been authorized 11 invention patents.

Title: Block Markov Superposition Transmission of LDPC Codes with High-Throughput Cooperative Layered Decoders

Abstract: Block Markov Superposition Transmissions are a class of spatial coupling construction for block codes to improve performance. Unlike spatially coupled LDPC codes, BMST-LDPC code offer the advantage of directly reutilizing the encoder and decoder of the existing basic LDPC codes. However, the dependence between BMST layers introduces substantial decoding delays. This talk proposes to address this challenge by introducing the layered decoding of QC-LDPC codes in the BMST system. By exploiting the parallel processing capabilities of layered decoding across BMST layers, we design the cooperate layered decoding for BMST-LDPC codes. The proposed coding scheme may find applications in scenarios requiring high-throughput and reliable data transmission, such as future wireless networks and high-speed storage devices, where QC-LDPC codes are widely adopted.

Professor Pascal O. Vontobel

Bio: Pascal O. Vontobel received the Diploma degree in electrical engineering in 1997, the Post-Diploma degree in information techniques in 2002, and the Ph.D. degree in electrical engineering in 2003, all from ETH Zurich, Switzerland.

From 1997 to 2002 he was a research and teaching assistant at the Signal and Information Processing Laboratory at ETH Zurich, from 2006 to 2013 he was a research scientist with the Information Theory Research Group at Hewlett-Packard Laboratories in Palo Alto, CA, USA, and since 2014 he has been with the Department of Information Engineering at the Chinese University of Hong Kong, where, since 2023, he has been a (full) professor, department chairman, and graduate division head. Besides this, he was a postdoctoral research associate at the University of Illinois at Urbana-Champaign (2002-2004), a visiting assistant professor at the University of Wisconsin-Madison (2004-2005), a postdoctoral research associate at the Massachusetts Institute of Technology (2006), and a visiting scholar at Stanford University (2014). His research interests lie in coding and information theory, quantum information processing, data science, communications, and signal processing.

Dr. Vontobel was an Associate Editor for the IEEE Transactions on Information Theory (2009-2012), an Awards Committee Member of the IEEE Information Theory Society (2013-2014), a Distinguished Lecturer of the IEEE Information Theory Society (2014-2015), an Associate Editor for the IEEE Transactions on Communications (2014-2017), and currently is a Thomas Cover Dissertation Awards Committee Member of the IEEE Information Theory Society (2023-2025). Moreover, he was a TPC co-chair of the 2016 IEEE International Symposium on Information Theory, the 2018 IEICE International Symposium on Information Theory and its Applications, and the 2018 IEEE Information Theory Workshop, along with being the director of the 2021 and 2023 Croucher Summer Course in Information Theory, co-organizing several topical workshops, and being on the technical program committees of many international conferences. Furthermore, he was multiple times a plenary speaker at international information and coding theory conferences, he received an exemplary reviewer award from the IEEE Communications Society, and was awarded the ETH medal for his Ph.D. dissertation. He is an IEEE Fellow.

https://www.ie.cuhk.edu.hk/faculty/vontobel-pascal-olivier/

Title: On Counting Arrays with Certain Properties

Abstract: Consider the set of arrays of size n by n that contain only zeros and ones. How many of these arrays have exactly k ones per column and k ones per row? For k = 1, it is straightforward to find the answer (n factorial). However, for k > 1, solving this problem (even approximately) becomes more interesting and challenging.

In this presentation I will discuss how the above question is motivated by applications in constrained coding, and how it can be approached with tools from graphical models, matroid theory, and real stable polynomials.

(Relevant background about graphical models, matroid theory, and real stable polynomials will be introduced as needed.)

(Based on joint work with Yuwen Huang.)

Professor Chao-Kai Wen

Bio: Dr. Chao-Kai Wen, a professor at the National Sun Yat-sen University’s Institute of Communication Engineering, is an IEEE Fellow and a Highly Cited Researcher by Clarivate Analytics. He served as the Director of the Institute of Communication Engineering at National Sun Yat-sen University from 2018 to 2021 and as the Vice Dean of the College of Engineering from 2020 to 2023. Prior to joining the university, he worked as an engineer at the Industrial Technology Research Institute (ITRI) and MediaTek (MTK) from 2004 to 2009.

Dr. Wen’s primary research focus is on communication signal processing. He currently serves as an editor, associate editor, and guest editor for several IEEE journals, including the IEEE Transactions on Wireless Communications, IEEE Transactions on Communications, IEEE Wireless Communications Letters, IEEE Communications Letters, and IEEE Communication Magazine. He has received multiple awards, including the 2016 GLOBECOM and 2022 ICC Best Paper Awards, the 2023 Jack Neubauer Memorial Award, and the Best Editor Award for both IEEE Wireless Communications Letters and IEEE Communications Letters.

Title: End-User MIMO Evolution toward 6G

Abstract: In 6G, the trend of transitioning from massive antenna elements to even more massive ones is continued. However, installing additional antennas in the limited space of user equipment (UE) is challenging, resulting in limited capacity scaling gain for end users, despite network side support for increasing numbers of antennas. To address this issue, we propose an end-user-centric collaborative MIMO (UE-CoMIMO) framework that groups several fixed or portable devices to provide a virtual abundance of antennas. This talk outlines how advanced L1 relays and conventional relays enable device collaboration to offer communication, localization, and sensing enhancements.

Professor Farzan Farnia

Bio: Farzan Farnia is an assistant professor at the Department of Computer Science and Engineering at the Chinese University of Hong Kong. Prior to joining CUHK, he was a postdoctoral research associate at the Laboratory for Information and Decision Systems, Massachusetts Institute of Technology, from 2019-2021. He received his M.Sc. and Ph.D. degrees in electrical engineering from Stanford University and his B.Sc. degree in electrical engineering and mathematics from Sharif University of Technology. At Stanford, he was a graduate research assistant at the Information Systems Laboratory advised by David Tse. Farzan’s research interests include machine learning, deep learning theory, optimization, and information theory. He has been the recipient of a Stanford Graduate Fellowship from 2013-2016 and the Numerical Technology Founders Prize as the second top performer of Stanford Electrical Engineering Ph.D. Qualifying Exams in 2014.

Title: Scalable Evaluation of Generative Models

Abstract: Evaluating generative models remains a core challenge in machine learning, especially in the reference-free setting where no ground-truth samples are available. In this talk, I will present recent work on developing scalable and adaptive frameworks for evaluating generative models. We first introduce a method that leverages random Fourier features to efficiently assess the quality of generated samples without relying on reference datasets. Building on this, we explore an online evaluation approach based on multi-armed bandits, allowing for dynamic model selection under limited feedback. I will also discuss ongoing developments and open challenges in building practical and reliable evaluation tools for generative models.

Professor Chu-Hsiang Huang

Bio: Chu-Hsiang Huang received B.S. and M.S. degrees in Electrical Engineering from National Taiwan University, Taiwan in 2007 and 2009, respectively. and his Ph.D. degree in Electrical Engineering from University of California, Los Angeles in 2015. He is now with National Taiwan University (NTU) as an assistant professor. Before join NTU, he was working in Qualcomm Technologies, Inc. as a RAN working group delegate for 3GPP Standard Organization. Besides representing Qualcomm in 3GPP standard meetings, he is working on product development projects including multi-user interference mitigation, energy efficient receiver and demodulation algorithm for Qualcomm flagship modems as a senior staff engineer. He was a research assistant for NTU-INTEL research center in Taiwan in 2010. His research interest includes next generation wireless communication system design, communication system standardization, artificial intelligence and machine learning, statistical communication theory .

Title: Air-Interface Design for Life Cycle Management and Interoperability of AI/ML Models in Wireless Communication Systems

Abstract: The performance improvement by adapting AI models in the communication systems is widely recognized by the research communities, and it leads to the standardization of air-interface design enabling wider and more efficient application of AI models in the communication system. During the standardization discussion, performance guarantee for AI models operating in different communication conditions/configurations is intensively studied. Model size limitation due to the storage and processing capability constraints of the mobile devices brings additional concerns of the generalizability of the AI models against the changing environments. Therefore, life cycle management (LCM) becomes essential for AI-model-based transceiver algorithms to manage model adaptation and updating based on different monitoring and statistical analysis functionalities, and guarantee the transceiver algorithm performance across various environments and configurations. When considering two-sided model, in which the AI models in BS and UE collaborate to implement the desired functionalities, interoperability across BS and UE side models is brought into the model management discussion. Since the AI models on the BS and the UE are interacting with each other, matching the AI models to ensure the correct execution of the designated functionalities becomes a new challenge when the AI models on the two sides are independently developed by different entities. In this talk, we provide an overview of 3GPP standardization progress on the general framework for LCM functionality development, and the preliminary air-interface design of functionalities like monitoring for specific use cases. In addition, we cover the interoperability perspective of model management focusing on two-sided model by going through the evaluations of specification-aided collaborative development process and verification/testing procedures. We conclude the talk by reviewing potential future directions of the LCM development and interoperable design from the air-interface specification perspective.

Professor Hsin-Po Wang

Bio: Hsin-Po Wang is an Assistant Professor in EE and GICE at National Taiwan University. His research interests lie in information theory and coding theory, where he applies techniques in algebra, combinatorics, and probability theory to polar codes, group testing, distributed storage, and distributed computation. Hsin-Po earned his B.Sc. in Mathematics at National Taiwan University and completed his Ph.D. in Mathematics at the University of Illinois Urbana-Champaign. He has held research positions at UC San Diego, UC Berkeley, and the Simons Institute for the Theory of Computing.

Title: Sculpting Rational Distributions Accurately and Efficiently

Abstract: Simulating an arbitrary discrete distribution $D \in [0, 1]^n$ using fair coin tosses incurs trade-offs between entropy complexity and space and time complexity. Shannon’s theory suggests that $H(D)$ tosses are necessary and sufficient, but does not guarantee exact distribution. Knuth and Yao showed that a decision tree consumes fewer than $H(D) + 2$ tosses for one exact sample. Drapper and Saad’s recent work addresses the space and time aspect, showing that $H(D) + 2$ tosses, $O(n \log(n) \log(m))$ memory, and $O(H(D))$ operations are all it costs, where $m$ is the common denominator of the probability masses in $D$ and $n$ is the number of possible outcomes.

In this talk, we show how to recycle leftover entropy to break the ``$+2$’’ barrier. With $O((n + 1/\varepsilon) \log(m/\varepsilon))$ memory and $O(\log(m/\varepsilon)^2 / \varepsilon)$ operations, the entropy cost is reduced to $H(D) + \varepsilon$, proving Shannon right.

Professor Ziye Ma

Bio: The speaker Ziye Ma is currently a presidential assistant professor in the computer science department at the City University of Hong Kong. Prior to this, he completed his PhD in the EECS department at UC Berkeley under the guidance of Somayeh Sojoudi and Javad Lavaei. His research is mostly focused on non-convex problems in machine learning, low-rank matrix recovery, and robustness in ML. He serves as regular reviewers of top ML venues and several of his papers received oral designations in these conferences.

Title: Efficient and Fast Training with new Zero-th order Hybrid Optimizer.

Abstract: Optimizing large-scale nonconvex problems, common in machine learning, demands balancing rapid convergence with computational efficiency. First-order (FO) stochastic methods like SVRG provide fast convergence and good generalization but incur high costs due to full-batch gradients in large models. Conversely, zeroth-order (ZO) algorithms reduce this burden using estimated gradients, yet their slow convergence in high-dimensional settings limits practicality. We introduce VAMO (VAriance-reduced Mixed-gradient Optimizer), a stochastic variance-reduced method combining FO mini-batch gradients with lightweight ZO finite-difference probes under an SVRG-style framework. VAMO’s hybrid design uses a two-point ZO estimator to achieve a dimension-agnostic convergence rate of $\mathcal{O}(1/T + 1/b)$, where $T$ is the number of iterations and $b$ is the batch-size, surpassing the dimension-dependent slowdown of purely ZO methods and significantly improving over SGD’s $\mathcal{O}(1/\sqrt{T})$ rate. Additionally, we propose a multi-point ZO variant that mitigates the $O(1/b)$ error by adjusting number of estimation points to balance convergence and cost, making it ideal for a whole range of computationally constrained scenarios. Experiments including traditional neural network training and LLM finetuning show VAMO outperforms established FO and ZO methods, offering a faster, more flexible option for improved efficiency.

Professor Linqi Song

Bio: Prof. Linqi Song is an Associate Professor in the Computer Science Department at the City University of Hong Kong. He was the Chair of IEEE Information Theory Society Hong Kong Chapter. He received the Ph.D. degree in Electrical Engineering from UCLA, and the B.S. and M.S. degrees in Electronic Engineering from Tsinghua University. His research interests encompass information theory, coded computation, machine learning, and artificial intelligence. He has published more than 180 high-impact papers in top-tier journals and conferences, such as IEEE TIT, JSAC, TPAMI, NeurIPS, ICLR, ACL, CVPR, and EMNLP. He has filed several invention patents and won three silver medals at the Geneva International Exhibition of Inventions. Prof. Song has received several prestigious awards, including the Hong Kong Research Grants Council’s Early Career Scheme Award in 2019, the Best Paper Award at IEEE MIPR 2020, and the Best Paper Award at China Communications 2023. He serves as an Associate Editor for Intelligent and Converged Networks and a Guest Editor for Journal of Franklin Institute and Digital Signal Processing. Prof. Song has led projects funded by the National Natural Science Foundation of China, the Hong Kong Research Grants Council, and the Hong Kong Innovation and Technology Commission, as well as various provincial, municipal, and industry collaboration projects.

https://scholars.cityu.edu.hk/en/persons/linqi-song(a665d7a3-8847-404d-a56a-2b10b470327c).html

Title: Managing Heavy-tailed Gradients in Distributed Learning via Improved Quantization Strategies

Abstract: In distributed learning, the transmission of gradients with heavy-tailed distributions can significantly strain communication bandwidth, which is a key bottleneck in scaling AI models across multiple computing nodes. As model sizes grow and training data becomes increasingly decentralized, efficient communication of gradients is critical to ensuring the scalability and performance of distributed Stochastic Gradient Descent (SGD). Existing quantization methods often fail to adequately balance communication efficiency with convergence accuracy, especially in the presence of heavy-tail gradients. To address these challenges, we propose a novel two-stage quantization strategy that enhances communication efficiency while preserving model accuracy. Our method begins by truncating gradients to mitigate the detrimental effects of long-tail noise, followed by a non-uniform quantization step that compresses gradients based on their statistical distribution. This approach directly tackles the communication bottleneck while minimizing the loss in convergence performance. We rigorously analyze the convergence properties of our quantized distributed SGD and provide theoretical guarantees for its performance. Additionally, we derive optimal closed-form solutions for both the truncation threshold and the non-uniform quantization levels by minimizing the convergence error within given communication constraints. Our theoretical insights, supported by extensive experimental results, show that the proposed approach outperforms existing quantization methods, striking an optimal balance between communication efficiency and convergence accuracy, making it highly suitable for large-scale distributed learning.

Professor Ming-Hsun Yang

Bio: Ming-Hsun Yang received the Ph.D. degree in communications engineering from the National Yang Ming Chiao Tung University, Taiwan, in 2018. From 2019 to 2021, he was a Postdoctoral Researcher with the Institute of Communications Engineering, National Yang Ming Chiao Tung University. Since 2022, he has been with the Department of Communication Engineering at National Central University, where he is currently an Assistant Professor. His research interests include statistical signal processing, machine learning for wireless communications, and compressed sensing with applications in wireless sensor networks. In 2024, Dr. Yang received the Best Paper Award for Young Scholars from the IEEE IT/COM Society Taipei/Tainan Chapters.

Title: Sparse Affine Sampling for Ambiguity-Free and Efficient Sparse Phase Retrieval

Abstract: Conventional sparse phase retrieval schemes can recover sparse signals from the magnitude of linear measurements only up to a global phase ambiguity. In this talk, we will introduce a novel approach that instead utilizes the magnitude of affine measurements to achieve ambiguity-free signal reconstruction. The proposed method relies on two-stage approach that consists of support identification followed by the exact recovery of nonzero signal entries. In the noise-free case, perfect support identification using a simple counting rule is guaranteed subject to a mild condition on the signal sparsity, and subsequent exact recovery of the nonzero signal entries can be obtained in closed-form. The proposed approach is then extended to the bounded noise scenario. In this setting, perfect support recovery still holds under mild conditions on the noise power, and robust signal reconstruction up to a bounded error can be achieved using linear least-squares estimation. The obtained analytic performance guarantee also sheds light on the construction of the sensing matrix and bias vector. In particular, we show that a near optimal performance can be achieved with high probability using a random construction scheme.

Professor Pingzhi Fan

Bio: Pingzhi Fan received his MSc degree in Computer Science from the Southwest Jiaotong University (SWJTU), PRC, in 1987, and Ph.D. degree in Electronic Engineering from the Hull University, U.K., in 1994. He is currently a distinguished professor of SWJTU, director of the institute of mobile communications, and honorary dean of the Sino-UK SWJTU-Leeds Joint School. He is also a visiting/guest professor of Leeds University (UK, 1997-) and Shanghai Jiaotong University (China, 1999-); and an honorary professor of the University of Nottingham Ningbo China (UNNC, 2025.1-). He served as vice president for academic affairs of SWJTU between 2007-2014, dean of the graduate school (SWJTU) between 2007-2012, dean of the school of info sci & tech (SWJTU) between 1998-2008. He has published over 900 research works, including 8 books and edited books published by John Wiley & Sons Ltd, RSP, Springer, IEEE Press, Nova Science Publishers, etc., over 430 journal papers published in various internationally renowned English journals (IEEE Trans on Com, IEEE Trans on WC, IEEE Trans on VT, IEEE Trans on IT, IEEE JSAC, IEEE Trans on SP, IEEE Trans on PDS, IEEE Trans on Computer, IEEE CL/WCL, IEICE Trans on Fundamentals, IEE/IET Electronic Letters, IEE Proc on Communications etc.), as well as over 160 Chinese journal papers and over 330 referred international conference papers. He is also an inventor of 30+ granted Chinese and PCT patents. His works has been cited 30,280+ times, h-index=71, Google Scholar, 30 March 2025.

Title: Non-Uniform Pilot Pattern for MIMO-OFDM Systems Based on Difference Sets

Abstract: OFDM performs well in frequency-selective channels by resisting multipath interference. Due to the limited number of effective multipath components in wireless environments, the channel impulse response often exhibits sparsity in the time domain. This enables sparse channel estimation with high accuracy and low pilot overhead. While Cyclic Difference Sets (CDS)-based pilot schemes perform well in single-antenna systems, they may cause severe interference in Multiple-Input Multiple-Output (MIMO) scenarios. To address this problem, this talk introduces the concept of Identical Multi-set Intersection (IMI) for pilot design in MIMO-OFDM systems, effectively suppressing inter-antenna interference. In addition, two pilot placement strategies, Zero Setting (ZS) and Partial Orthogonal Covering Codes (POCC), are proposed. Our results confirm that the proposed methods significantly improve the estimation accuracy in MIMO-OFDM systems.

Professor Li Congduan

Bio: Dr. Li Congduan, Associate Professor and Doctoral Supervisor at Sun Yat sen University with the Hundred Talents Program, Young Rising Star in Information Theory at the China Electronics Society, Overseas High-level Talent in Shenzhen, IEEE Senior Member, and Member of the China Internet of Things Youth Technical Committee. His research focuses on communication networks and artificial intelligence technology. More than 70 academic papers have been published on related topics and over ten Chinese invention patents have been authorized. In the past 5 years, he has participated many projects as PI, such as the National Natural Science Foundation, the National Key R&D Program, key project of basic and applied basic research in Guangdong Province, key fund for scientific and technological innovation in Shenzhen, etc.

Title: Distributed Computation and Learning meets Network Coding

Abstract: Network coding is a promising technology to increase the network throughput, especially in the multi-source communication networks. It can also be used in distributed systems such as storage, computation, and machine learning. In this talk, we will share some new results in distributed computation and learning systems with emphasis on the applications of network coding technique to guarantee the fault-tolerance, increase the efficiency and privacy in them.