Speaker:
Dr. Quanshi Zhang

Abstract:
The interpretability of deep neural networks has received increasing attention in recent years, and diverse methods of explainable AI (XAI) have been developed. Currently, most XAI methods are designed in an experimental manner without solid theoretic foundations, or simply fit explanation results to people’s cognition instead of objectively reflecting the true knowledge in the DNN. The lack of theoretic supports has hampered the future development of XAI. Therefore, in this talk, Dr. Quanshi Zhang will review several studies of explainable AI theories of his research group in recent years, which use the system of game-theoretic interactions to explain the attribution, the adversarial robustness, model generalization, visual concepts learned by the DNN, and the aesthetic level of images.

Biography:
Dr. Quanshi Zhang is an associate professor at Shanghai Jiao Tong University, China. He received the Ph.D. degree from the University of Tokyo in 2014. From 2014 to 2018, he was a post-doctoral researcher at the University of California, Los Angeles. His research interests are mainly machine learning and computer vision. In particular, he has made influential research in explainable AI (XAI) and received the ACM China Rising Star Award. He was the co-chairs of the workshops towards XAI in ICML 2021, AAAI 2019, and CVPR 2019. We is the speaker of the tutorials on XAI at IJCAI 2020 and IJCAI 2021.

Join Zoom Meeting:
https://cuhk.zoom.us/j/98782922295

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Dr. Zichao Yang

Abstract:
In recent years, advances in deep learning for NLP research have been mainly propelled by massive computation and large amounts of data. Despite the progress, those giant models still rely on in-domain data to work well in down-stream tasks, which is hard and costly to obtain in practice. In this talk, I am going to talk about my research efforts towards overcoming the challenge of learning with limited supervision by designing efficient NLP models. My research spans three directions towards this goal: designing structural neural networks models according to NLP data structures to take full advantage of labeled data, effective unsupervised models to alleviate the dependency on labeled corpus and data augmentation strategies which creates large amounts of labeled data at almost no cost.

Biography:
Zichao Yang is currently a research scientist working at Bytedance. Before that he obtained his Ph.D from CMU working with Eric Xing, Alex Smola and Taylor Berg-Kirkpatrick. His research interests lie in machine learning and deep learning with applications in NLP. He has published dozens of papers in top AI/ML conferences. He obtained his MPhil degree from CUHK and bachelor degree from Shanghai Jiao Tong University. He worked at Citadel Securities as a quantitative researcher, specializing in ML research for financial data, before joining Bytedance. He also interned in Google DeepMind, Google Brain and Microsoft Research during his Phd.

Join Zoom Meeting:
https://cuhk.zoom.us/j/94185450343

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. HAN Dongsu

Abstract:
Internet video has experienced tremendous growth over the last few decades and is still growing at a rapid pace. Internet video now accounts for 73% of Internet traffic and is expected to quadruple in the next five years. Augmented reality and virtual reality streaming, projected to increase twentyfold in five years, will also accelerate this trend.

In this talk, I will argue that advances in deep neural networks present new opportunities that can fundamentally change Internet video delivery. In particular, deep neural networks allow the content delivery network to easily capture the content of the video and thus enable content-aware video delivery. To demonstrate this, I will present NAS, a new Internet video delivery framework that integrates deep neural network based quality enhancements with adaptive streaming.

NAS incorporates a super-resolution deep neural network (DNN) and a deep re-inforcement neural network to optimize the user quality of experience (QoE). It outperforms the current state of the art, dramatically improving visual quality. It improves the average QoE by 43.08% using the same bandwidth budget or saving 17.13% of bandwidth while providing the same user QoE.

Finally, I will talk about our recent research progress in supporting live video and mobile devices in AI-assisted video delivery that demonstrate the possibility of new designs that tightly integrate deep learning into Internet video streaming.

Biography:
Dongsu Han (Member, IEEE) is currently an Associate Professor with the School of Electrical Engineering at KAIST. He received the B.S. degree in computer science from KAIST in 2003 and the Ph.D. degree in computer science from Carnegie Mellon University in 2012. His research interests include networking, distributed systems, and network/system security. He has received Best Paper Award and Community Award from USENIX NSDI. More details about his research can be found at http://ina.kaist.ac.kr.

Join Zoom Meeting:
https://cuhk.zoom.us/j/93072774638

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Dr. Huaicheng Li

Abstract:
The increasing complexity in storage software and hardware brings new challenges to achieve predictable performance and efficiency. On the one hand, emerging hardware break long-held system design principles and are held back by aged and inflexible system interfaces and usage models, requiring radical rethinking on the software stack to leverage new hardware capabilities for optimal performance. On the other hand, the computing landscape is becoming increasingly heterogeneous and complex, demanding explicit systems-level support to manage hardware-associated complexity and idiosyncrasy, which is unfortunately still largely missing.

In this talk, I will discuss my efforts to build low-latency and cost-efficient datacenter storage systems. By revisiting existing storage interface/abstraction designs and software/hardware responsibility divisions, I will present holistic storage stack designs for cloud datacenters, which deliver orders of magnitude of latency improvement and significantly improved cost-efficiency.

Biography:
Huaicheng is a postdoc at CMU in the Parallel Data Lab (PDL). He received his Ph.D. from University of Chicago. His interests are mainly in Operating Systems and Storage Systems, with a focus on building high-performance and cost-efficient storage infrastructure for datacenters. His research has been recognized by two best paper nominations at FAST (2017 and 2018) and has also made real impact, with production deployment in datacenters, code integration to Linux, and a storage research platform widely used by the research community.

Join Zoom Meeting:
https://cuhk.zoom.us/j/95132173578

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Dr. Yaoqing Yang

Abstract:
Machine learning (ML) models are increasingly being deployed in safety-critical applications, making their generalization and reliability a problem of urgent societal importance. To date, our understanding of ML is still limited because (i) the narrow problem settings considered in studies and the (often) cherry-picked results lead to incomplete/conflicting conclusions on the failures of ML; (ii) focusing on low-dimensional intuitions results in a limited understanding of the global structure of ML problems. In this talk, I will present several recent results on “generalization metrics” to measure ML models. I will show that (i) generalization metrics such as the connectivity between local minima can quantify global structures of optimization loss landscapes, which can lead to more accurate predictions on test performance than existing metrics; (ii) carefully measuring and characterizing the different phases of loss landscape structures in ML can provide a more complete picture of generalization. Specifically, I show that different phases of learning require different ways to address failures in generalization. Furthermore, most conventional generalization metrics focus on the so-called generalization gap, which is indirect and of limited practical value. I will discuss novel metrics referred to as “shape metrics” that allow us to predict test accuracy directly instead of the generalization gap. I also show that one can use shape metrics to achieve improved compression and out-of-distribution robustness of ML models. I will discuss theoretical results and present large-scale empirical analyses for different quantity/quality of data, different model architectures, and different optimization hyperparameter settings to provide a comprehensive picture of generalization. I will also discuss practical applications of utilizing these generalization metrics to improve ML models’ training, efficiency, and robustness.

Biography:
Dr. Yaoqing Yang is a postdoctoral researcher at the RISE Lab at UC Berkeley. He received his PhD from Carnegie Mellon University and B.S. from Tsinghua University, China. He is currently focusing on machine learning, and his main contributions to machine learning are towards improving reliability and generalization in the face of uncertainty, both in the data and the compute platform. His PhD thesis laid the foundation for an exciting field of research—coded computing—where information-theoretic techniques are developed to address unreliability in computing platforms. His works have won the best paper finalist at ICDCS and have been published multiple times in NeurIPS, CVPR, and IEEE Transactions on Information Theory. He has worked as a research intern at Microsoft, MERL and Bell Labs, and two of his joint CVPR papers with MERL have both received more than 300 citations. He is also the recipient of the 2015 John and Claire Bertucci Fellowship.

Join Zoom Meeting:
https://cuhk.zoom.us/j/99128234597

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Mr. Guodong Zhang

Abstract:
Deep learning has achieved huge successes over the last few years, largely due to three important ideas: deep models with residual connections, parallelism, and gradient-based learning. However, it was shown that (1) deep ResNets behave like ensembles of shallow networks; (2) naively increasing the scale of data parallelism leads to diminishing return; (3) gradient-based learning could converge to spurious fixed points in the multiagent setting.

In this talk, I will present some of my works on understanding and addressing these issues. First, I will give a general recipe for training very deep neural networks without shortcuts. Second, I will present a noisy quadratic model for neural network optimization, which qualitatively predicts scaling properties of a variety of optimizers and in particular suggests that second-order algorithms would benefit more from data parallelism. Third, I will describe a novel algorithm that finds desired equilibria and saves us from converging to spurious fixed points in multi-agent games. In the end, I will conclude with future directions towards building intelligent machines that can learn from experience efficiently and reason about their own decisions.

Biography:
Guodong Zhang is a PhD candidate in the machine learning group at the University of Toronto, advised by Roger Grosse. His research lies at the intersection between machine learning, optimization, and Bayesian statistics. In particular, his research focuses on understanding and improving algorithms for optimization, Bayesian inference, and multi-agent games in the context of deep learning. He has been recognized through the Apple PhD fellowship, Borealis AI fellowship, and many other scholarships. In the past, he has also spent time at Institute for Advanced Study of Princeton and industry research labs (including DeepMind, Google Brain, and Microsoft Research).

Join Zoom Meeting:
https://cuhk.zoom.us/j/95830950658

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Ms. Jiasi Shen

Abstract:
Software now plays a central role in numerous aspects of human society. Current software development practices involve significant developer effort in all phases of the software life cycle, including the development of new software, detection and elimination of defects and security vulnerabilities in existing software, maintenance of legacy software, and integration of existing software into more contexts, with the quality of the resulting software still leaving much to be desired. The goal of my research is to improve software quality and reduce costs by automating tasks that currently require substantial manual engineering effort.

I present a novel approach for automatic software rejuvenation, which takes an existing program, learns its core functionality as a black box, builds a model that captures this functionality, and uses the model to generate a new program. The new program delivers the same core functionality but is potentially augmented or transformed to operate successfully in different environments. This research enables the rejuvenation and retargeting of existing software and provides a powerful way for developers to express program functionality that adapts flexibly to a variety of contexts. In this talk, I will show how we applied these techniques to two classes of software systems, specifically database-backed programs and stream-processing computations, and discuss the broader implications of these approaches.

Biography:
Jiasi Shen is a Ph.D. candidate at MIT EECS advised by Professor Martin Rinard. She received her bachelor’s degree from Peking University. Her main research interests are in programming languages and software engineering. She was named an EECS Rising Star in 2020.

Join Zoom Meeting:
https://cuhk.zoom.us/j/91743099396

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. Wen Wujie

Abstract:
Thanks to recent model innovation and hardware advancement, machine learning (ML) has now achieved extraordinary success in many fields ranging from daily image classification, object detection, to security- sensitive biometric authentication and autonomous vehicles. To facilitate fast and secure end-to-end machine learning services, extensive studies have been conducted on ML hardware acceleration and data or model-incurred adversarial attacks. Different from these existing efforts, in this talk, we will present a new understanding of the efficiency and security challenges in accelerated ML services. The talk starts with the development of the very first “machine vision” (NOT “human vision”) guided image compression framework tailored for fast cloud-based machine learning services with guaranteed accuracy, inspired by an insightful understanding about the difference between machine learning (or “machine vision”) and human vision on image perception. Then we will discuss “StegoNet”- a new breed stegomalware taking advantage of machine learning service as a stealthy channel to conceal malicious intent (malware). Unlike existing attacks focusing only on misleading ML outcomes, “StegoNet” for the first time achieves far more diversified adversarial goals without compromising ML service quality. Our research prospects will be also given at the end of this talk, offering the audiences an alternative thinking about developing efficient and secure machine learning services.

Biography:
Wujie Wen is an assistant professor in the Department of Electrical and Computer Engineering at Lehigh University. He received his Ph.D. from University of Pittsburgh in 2015. He earned his B.S. and M.S. degrees in electronic engineering from Beijing Jiaotong University and Tsinghua University, Beijing, China, in 2006 and 2010, respectively. He was an assistant professor in the ECE department of Florida International University, Miami, FL, during 2015-2019. Before joining academia, he also worked with AMD and Broadcom for various engineer and intern positions. His research interests include reliable and secure deep learning, energy-efficient computing, electronic design automation and emerging memory systems design. His works have been published widely across venues in design automation, security, machine learning/AI etc., including HPCA, DAC, ICCAD, DATE, ICPP, HOST, ACSAC, CVPR, ECCV, AAAI etc. He received best paper nominations from ASP-DAC2018, ICCAD2018, DATE2016 and DAC2014. Dr Wen served as the General Chair of ISVLSI 2019 (Miami), Technical Program Chair of ISVLSI 2018 (Hong Kong), as well as the program committee for many conferences such as DAC, ICCAD, DATE, etc. He is an associated editor of Neurocomputing and IEEE Circuit and Systems (CAS) Magazine. His research projects are currently sponsored by US National Science Foundation, Air Force Research Laboratory and Florida Center for Cybersecurity etc.

Join Zoom Meeting:
https://cuhk.zoom.us/j/98308617940

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. LI Yu

Abstract:
In this talk, I will give an overview of the research in our group. Essentially, we are developing new machine learning methods to resolve the problems in computational biology and health informatics, from sequence analysis, biomolecular structure prediction, and functional annotation to disease modeling, drug discovery, drug effect prediction, and combating antimicrobial resistance. We will show how to formulate problems in the biology and health field into machine learning problems, how to resolve them using cutting-edge machine learning techniques, and how the result could benefit the biology and healthcare field in return.

Biography:
Yu Li is an Assistant Professor in the Department of Computer Science and Engineering at CUHK. His main research interest is to develop novel and new machine learning methods, mainly deep learning methods, for solving the computational problems in healthcare and biology, understanding the principles behind the bio-world, and eventually improving people’s health and wellness. He obtained his PhD in computer science from KAUST in Saudi Arabia, in Oct 2020. He obtained MS degree in computer science from KAUST at 2016. Before that, he got the Bachelor degree in Biosciences from University of Science and Technology of China (USTC).

Join Zoom Meeting:
https://cuhk.zoom.us/j/98928672713

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Mr. Dennis Lee

Abstract:
With the ever increasing demand and aging population, it is envisioned that adoption of AI technology will support Hospital Authority to tackle various strategic service challenges and deliver improvements. HA has setup AI Strategy Framework two years ago and begun setup process & infrastructure to support AI development and delivery. The establishment of AI Lab and AI delivery center is aimed to flourish AI innovations by engaging internal and external collaboration for Proof of Concept development; and also to build data and integration pipeline to validate AI solution and integrate into the HA services at scale.

By leverage 3 platforms to (1) Improve awareness of HA staff (2) Match AI supply and Demand (3) data pipeline for timely prediction, we can gradually scale AI innovations and solution in Hospital Authority. Over the past year, many clinical and non-clinical Proof of Concept has been developed and validated. The AI Chest X-ray pilot project has been implemented for General Outpatient Clinics and Emergency Department with the aim to reduce report turnaround time and provide decision support for abnormal chest x-ray imaging.

Biography:
Mr. Dennis Lee currently serves as the Senior System Manager for Artificial Intelligence Systems of the Hong Kong Hospital Authority. He current work involve developing the Artificial Intelligence and Big Data Platform to streamline data acquisition for facilitating HA data analysis via Business Intelligence, to develop Hospital Command Center dashboards, and solution deployment for Artificial Intelligence. Mr. Lee has also been leading the Corporate Project management office and as program managers for several large scale system implementations.

Join Zoom Meeting:
https://cuhk.zoom.us/j/95162965909

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Mr. Wenbo Guo

Abstract:
Nowadays, security researchers are increasingly using AI to automate and facilitate security analysis. Although making some meaningful progress, AI has not maximized its capability in security yet due to two challenges. First, existing ML techniques have not reached security professionals’ requirements in critical properties, such as interpretability and adversary-resistancy. Second, Security data imposes many new technical challenges, which break the assumptions of existing ML Models and thus jeopardize their efficacy.

In this talk, I will describe my research efforts to address the above challenges, with a primary focus on strengthening the interpretability of blackbox deep learning models and deep reinforcement learning policies. Regarding deep neural networks, I will describe an explanation method for deep learning-based security applications and demonstrate how security analysts could benefit from this method to establish trust in blackbox models and conduct efficient finetuning. As for DRL policies, I will introduce a novel approach to draw critical states/actions of a DRL agent and show how to utilize the above explanations to scrutinize policy weaknesses, remediate policy errors, and even defend against adversarial attacks. Finally, I will conclude by highlighting my future plan towards strengthening the trustworthiness of advanced ML techniques and maximizing their capability in cyber defenses.

Biography:
Wenbo Guo is a Ph.D. Candidate at Penn State, advised by Professor Xinyu Xing. His research interests are machine learning and cybersecurity. His work includes strengthening the fundamental properties of machine learning models and designing customized machine learning models to handle security-unique challenges. He is a recipient of the IBM Ph.D. Fellowship (2020-2022), Facebook/Baidu Ph.D. Fellowship Finalist (2020), and ACM CCS Outstanding Paper Award (2018). His research has been featured by multiple mainstream media and has appeared in a diverse set of top-tier venues in security, machine learning, and data mining. Going beyond academic research, he also actively participates in many world-class cybersecurity competitions and has won the 2018 DEFCON/GeekPwn AI challenge finalist award.

Join Zoom Meeting:
https://cuhk.zoom.us/j/95859338221

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. Wayne Luk

Abstract:
This talk describes recent research on meta-programming techniques for mapping high-level descriptions to multiple hardware platforms. The purpose is to enhance design productivity and maintainability. Our approach is based on decoupling functional concerns from optimisation concerns, allowing separate descriptions to be independently maintained by two types of experts: application experts focus on algorithmic behaviour, while platform experts focus on the mapping process. Our approach supports customisable optimisations to rapidly capture a wide range of mapping strategies targeting multiple hardware platforms, and reusable strategies to allow optimisations to be described once and applied to multiple applications. Examples will be provided to illustrate how the proposed approach can map a single high-level program into multi-core processors and reconfigurable hardware platforms.

Biography:
Wayne Luk is Professor of Computer Engineering with Imperial College London and the Director of the EPSRC Centre for doctoral training in High Performance Embedded and Distributed Systems. His research focuses on theory and practice of customizing hardware and software for specific application domains, such as computational finance, climate modelling, and genomic data analysis. He is a fellow of the Royal Academy of Engineering, IEEE, and BCS.

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. XU Hong

Abstract:
As cloud computing becomes ubiquitous, the network stack in this virtualized environment is becoming a focal point of research with unique challenges and opportunities. In this talk, I will introduce our efforts in this space.

First, from an architectural perspective, the network stack remains a part of the guest OS inside a VM in the cloud. I will argue that this legacy architecture is becoming a barrier to innovation/evolution. The tight coupling between the network stack and the guest OS causes many deployment troubles to tenants and management and efficiency problems to the cloud provider. I will present our vision of providing the network stack as a service as a way to address these issues. The idea is to decouple the network stack from the guest OS, and offer it as an independent entity implemented by the cloud provider. I will discuss the design and evaluation of a concrete framework called NetKernel to enable this vision. Then in the second part, I will focus on container communication, which is a common scenario in the cloud. I will present a new system called PipeDevice that adopts a hardware-software co-design approach to enable low-overhead intra-host container communication using commodity FPGA.

Biography:
Hong Xu is an Associate Professor in Department of Computer Science and Engineering, The Chinese University of Hong Kong. His research area is computer networking and systems, particularly big data systems and data center networks. From 2013 to 2020 he was with City University of Hong Kong. He received his B.Eng. from The Chinese University of Hong Kong in 2007, and his M.A.Sc. and Ph.D. from University of Toronto in 2009 and 2013, respectively. He was the recipient of an Early Career Scheme Grant from the Hong Kong Research Grants Council in 2014. He received three best paper awards, including the IEEE ICNP 2015 best paper award. He is a senior member of both IEEE and ACM.

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. Kai Chen
Associate Professor
Department of Computer Science & Engineering, HKUST

Abstract:
Communication overhead poses a significant challenge to distributed DNN training. In this talk, I will overview existing efforts toward this challenge, study their advantages and shortcomings, and further present a novel solution exploiting the domain-specific characteristics of deep learning to optimize the communication overhead of distributed DNN training in a fine-grained manner. Our solution consists of several key innovations beyond prior work, including bounded-loss tolerant transmission, gradient-aware flow scheduling, and order-free per-packet load-balancing, etc., delivering up to 84.3% training acceleration over the best existing solutions. Our proposal by no means provides an ultimate answer to this research problem, instead, we hope it can inspire more critical thinkings on intersection between Networking and AI.

Biography:
Kai Chen is an Associate Professor at HKUST, the Director of Intelligent Networking Systems Lab (iSING Lab) and HKUST-WeChat joint Lab on Artificial Intelligence Technology (WHAT Lab), as well as the PC for a RGC Theme-based Project. He received his BS and MS from University of Science and Technology of China in 2004 and 2007, and PhD from Northwestern University in 2012, respectively. His research interests include Data Center Networking, Cloud Computing, Machine Learning Systems, and Privacy-preserving Computing. His work has been published in various top venues such as SIGCOMM, NSDI and TON, etc., including a SIGCOMM best paper candidate. He is the Steering Committee Chair of APNet, serves on the Program Committees of SIGCOMM, NSDI, INFOCOM, etc., and the Editorial Boards of IEEE/ACM Transactions on Networking, Big Data, and Cloud Computing.

Join Zoom Meeting:
https://cuhk.zoom.us/j/98448863119?pwd=QUJVdzgvU1dnakJkM29ON21Eem9ZZz09

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. Sinno Jialin Pan
Provost’s Chair Associate Professor
School of Computer Science and Engineering
Nanyang Technological University

Abstract:
How to develop a loop to integrate existing knowledge to facilitate deep learning inference and then refine knowledge from the learning process is a crucial research problem. As first-order logic has been proven to be a powerful tool for knowledge representation and reasoning, interest in integrating firstorder logic into deep learning models has grown rapidly in recent years. In this talk, I will introduce our attempts to develop a unified integration framework of first-order logic and deep learning with applications to various joint inference tasks in NLP.

Biography:
Dr. Sinno Jialin Pan is a Provost’s Chair Associate Professor with the School of Computer Science and Engineering at Nanyang Technological University (NTU) in Singapore. He received his Ph.D. degree in computer science from the Hong Kong University of Science and Technology (HKUST) in 2011. Prior to joining NTU, he was a scientist and Lab Head with the Data Analytics Department at Institute for Infocomm Research in Singapore. He joined NTU as a Nanyang Assistant Professor in 2014. He was named to the list of “AI 10 to Watch” by the IEEE Intelligent Systems magazine in 2018. He serves as an Associate Editor for IEEE TPAMI, AIJ, and ACM TIST. His research interests include transfer learning and its real-world applications.

Join Zoom Meeting:
https://cuhk.zoom.us/j/97292230556?pwd=MDVrREkrWnFEMlF6aFRDQzJxQVlFUT09

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Dr. Jinwei Gu
R&D Executive Director
SenseBrain (aka SenseTime USA)

Abstract:
Smart sensing and perception refer to intelligent and efficient ways of measuring, modeling, and understanding of the physical world, which act as the eyes and ears of any AI-based system. Smart sensing and perception sit across the intersection of three related areas – computational imaging, representation learning, and scene understanding. Computational imaging refers to sensing the real world with optimally designed, task-specific, multi-modality sensors and optics which actively probes key visual information. Representation learning refers to learning the transformation from sensors’ raw output to some manifold embedding or feature spaces for further processing. Scene understanding includes both the low-level capture of a 3D scene of its physical properties, as well as high-level semantic perception and understanding of the scene. Advances in this area will not only benefit computer vision tasks but also result in better hardware, such as AI image sensors, AI ISP (Image Signal Processing) chips, and AI camera systems. In this talk, I will present several latest research results including high quality image restoration and accurate depth estimation from time-of-flight sensors or monocular videos, as well as some latest computational photography products in smart phones including under-display cameras, AI image sensors and AI ISP chips. I will also layout several open challenges and future research directions in this area.

Biography:
Jinwei Gu is the R&D Executive Director of SenseBrain (aka SenseTime USA). His current research focuses on low-level computer vision, computational photography, computational imaging, smart visual sensing and perception, and appearance modeling. He obtained his Ph.D. degree in 2010 from Columbia University, and his B.S and M.S. from Tsinghua University, in 2002 and 2005 respectively. Before joining
SenseTime, he was a senior research scientist in NVIDIA Research from 2015 to 2018. Prior to that, he was an assistant professor in Rochester Institute of Technology from 2010 to 2013, and a senior researcher in the media lab of Futurewei Technologies from 2013 to 2015. He serves as an associate editor for IEEE Transactions on Computational Imaging (TCI) and IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), an area chair for ICCV2019, ECCV2020, and CVPR2021, and industry chair for ICCP2020. He is an IEEE senior member since 2018. His research work has been successfully transferred to many products such as NVIDIA CoPilot SDK, DriveIX SDK, as well as super resolution, super night, portrait restoration, RGBW solution which are widely used in many flagship mobile phones.

Join Zoom Meeting:
https://cuhk.zoom.us/j/97322964334?pwd=cGRJdUx1bkxFaENJKzVwcHdQQm5sZz09

Enquiries: Ms. Karen Chan at Tel. 3943 8439

Speaker:
Prof. DOU Qi

Abstract:
With advancements in information technologies and medicine, the operating room has undergone tremendous transformations evolving into a highly complicated environment. These achievements further innovate the surgery procedure and have great promise to enhance the patient’s safety. Within the new generation of operating theatre, the computer-assisted system plays an important role to provide surgeons with reliable contextual support. In this talk, I will present a series of deep learning methods towards interdisciplinary researches at artificial intelligence for surgical robotic perception, for automated surgical workflow analysis, instrument presence detection, surgical tool segmentation, surgical scene perception, etc. The proposed methods cover a wide range of deep learning topics including semi-supervised learning, relational graph learning, learning-based stereo depth estimation, reinforcement learning, etc. The challenges, up-to-date progresses and promising future directions of AI-powered context-aware operating theaters will also be discussed.

Biography:
Prof. DOU Qi is an Assistant Professor with the Department of Computer Science & Engineering, CUHK. Her research interests lie in innovating collaborative intelligent systems that support delivery of high-quality medical diagnosis, intervention and education for next-generation healthcare. Her team pioneers synergistic advancements across artificial intelligence, medical image analysis, surgical data science, and medical robotics, with an impact to support demanding clinical workflows such as robotic minimally invasive surgery.

Enquiries: Miss Karen Chan at Tel. 3943 8439

Speaker:
Prof. HO Tsung Yi

Abstract:
Advances in microfluidic technologies have led to the emergence of biochip devices for automating laboratory procedures in biochemistry and molecular biology. Corresponding systems are revolutionizing a diverse range of applications, e.g. point-of-care clinical diagnostics, drug discovery, and DNA sequencing–with an increasing market. However, continued growth (and larger revenues resulting from technology adoption by pharmaceutical and healthcare companies) depends on advances in chip integration and design-automation tools. Thus, there is a need to deliver the same level of design automation support to the biochip designer that the semiconductor industry now takes for granted. In particular, the design of efficient design automation algorithms for implementing biochemistry protocols to ensure that biochips are as versatile as the macro-labs that they are intended to replace. This talk will first describe technology platforms for accomplishing “biochemistry on a chip”, and introduce the audience to both the droplet-based “digital” microfluidics based on electrowetting actuation and flow-based “continuous” microfluidics based on microvalve technology. Next, the presenter will describe system-level synthesis includes operation scheduling and resource binding algorithms, and physical-level synthesis includes placement and routing optimizations. Moreover, control synthesis and sensor feedback-based cyberphysical adaptation will be presented. In this way, the audience will see how a “biochip compiler” can translate protocol descriptions provided by an end user (e.g., a chemist or a nurse at a doctor’s clinic) to a set of optimized and executable fluidic instructions that will run on the underlying microfluidic platform. Finally, present status and future challenges of open-source microfluidic ecosystem will be covered.

Biography:
Tsung-Yi Ho received his Ph.D. in Electrical Engineering from National Taiwan University in 2005. His research interests include several areas of computing and emerging technologies, especially in design automation of microfluidic biochips. He has been the recipient of the Invitational Fellowship of the Japan Society for the Promotion of Science (JSPS), the Humboldt Research Fellowship by the Alexander von Humboldt Foundation, the Hans Fischer Fellowship by the Institute of Advanced Study of the Technische Universität München, and the International Visiting Research Scholarship by the Peter Wall Institute of Advanced Study of the University of British Columbia. He was a recipient of the Best Paper Awards at the VLSI Test Symposium (VTS) in 2013 and IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems in 2015. He served as a Distinguished Visitor of the IEEE Computer Society for 2013-2015, a Distinguished Lecturer of the IEEE Circuits and Systems Society for 2016-2017, the Chair of the IEEE Computer Society Tainan Chapter for 2013-2015, and the Chair of the ACM SIGDA Taiwan Chapter for 2014-2015. Currently, he serves as the Program Director of both EDA and AI Research Programs of Ministry of Science and Technology in Taiwan, VP Technical Activities of IEEE CEDA, an ACM Distinguished Speaker, and Associate Editor of the ACM Journal on Emerging Technologies in Computing Systems, ACM Transactions on Design Automation of Electronic Systems, ACM Transactions on Embedded Computing Systems, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, and IEEE Transactions on Very Large Scale Integration Systems, Guest Editor of IEEE Design & Test of Computers, and the Technical Program Committees of major conferences, including DAC, ICCAD, DATE, ASP-DAC, ISPD, ICCD, etc. He is a Distinguished Member of ACM.

Enquiries: Miss Karen Chan at Tel. 3943 8439

Speaker:
Prof. FARNIA Farzan

Abstract:
Generative Adversarial Networks (GANs) represent a game between two machine players designed to learn the distribution of observed data.

Since their introduction in 2014, GANs have achieved state-of-the-art performance on a wide array of machine learning tasks. However, their success has been observed to heavily depend on the minimax optimization algorithm used for their training. This dependence is commonly attributed to the convergence speed of the underlying optimization algorithm. In this seminar, we focus on the generalization properties of GANs and present theoretical and numerical evidence that the minimax optimization algorithm also plays a key role in the successful generalization of the learned GAN model from training samples to unseen data. To this end, we analyze the generalization behavior of standard gradient-based minimax optimization algorithms through the lens of algorithmic stability. We leverage the algorithmic stability framework to compare the generalization performance of standard simultaneous-update and non-simultaneous-update gradient-based algorithms. Our theoretical analysis suggests the superiority of simultaneous-update algorithms in achieving a smaller generalization error for the trained GAN model.

Finally, we present numerical results demonstrating the role of simultaneous-update minimax optimization algorithms in the proper generalization of trained GAN models.

Biography:
Farzan Farnia is an Assistant Professor 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 master’s and PhD degrees in electrical engineering from Stanford University and his bachelor’s degrees 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 Professor David Tse. Farzan’s research interests span statistical learning theory, information theory, and convex optimization. He has been the recipient of the Stanford Graduate Fellowship (Sequoia CapitalFellowship) between 2013-2016 and the Numerical Technology Founders Prize as the second top performer of Stanford’s electrical engineering PhD qualifying exams in 2014.

Enquiries: Miss Karen Chan at Tel. 3943 8439

Speaker:
Prof. CHAN Siu On
Assistant Professor
Department of Computer Science and Engineering, CUHK

Abstract:
This talk will summarize my works in two unrelated areas in complexity theory: distributional learning and extended formulation.

(1) Distributional Learning: Much of the work on distributional learning assumes the input samples are identically and independently distributed. A few recent works relax this assumption and instead assume the samples to be drawn as a trajectory from a Markov chain. Previous works by Wolfer and Kontorovich suggested that learning and identity test problems on ergodic chains can be reduced to the corresponding problems with i.i.d. samples. We show how to further reduce essentially every learning and identity testing problem on the (arguably most general) class of irreducible chans, by introducing the concept of k-cover time. The concept of k-cover time is a natural generalization of the usual notion of cover time.

The tight analysis of the sample complexity for reversible chains relies on a previous work by Ding-Lee-Peres. Their analysis relies on the so-called generalized second Ray-Knight isomorphism theorem, that connects the local time of a continuous-time reversible Markov chain to the Gaussian free field. It is natural to ask whether similar analysis can be generalized to general chains. We will discuss our ongoing work towards this goal.

(2) Extended formulation: Extended formulation lower bounds aim to show that linear programs (or other convex programs) need to be large in solving certain problems, such as constraint satisfaction. A natural open problem is whether refuting unsatisfiable 3-SAT instances requires linear programs of exponential size, and whether such a lower bound holds for every “downstream” NP-hard problem. I will discuss our ongoing work towards relating extended formulation lower bounds, using techniques from resolution lower bounds.

Biography:
Siu On CHAN graduated from the Chinese University of Hong Kong. He got his MSc at the University of Toronto and PhD at UC Berkeley. He was a postdoc at Microsoft Research New England. He is now an Assistant Professor at the Chinese University of Hong Kong. He is interested in the complexity of constraint satisfaction and learning algorithms. He won a Best Paper Award and a Best Student Paper Award at STOC 2013.

Enquiries: Miss Karen Chan at Tel. 3943 8439

Speaker:
Prof. YU Bei

Abstract:
Deep neural networks (DNNs) are currently widely used for many artificial intelligence AI applications with state-of-the-art accuracy, but they come at the cost of high computational complexity. Therefore, techniques that enable efficient computing of deep neural networks to improve key metrics—such as energy efficiency, throughput, and latency—without sacrificing accuracy are critical. This talk provides a structured treatment of the key principles and techniques for enabling efficient computing of DNNs, including implementation level, model level, and compilation level techniques.

Biography:
Bei Yu is currently an Associate Professor at the Department of Computer Science and Engineering, The Chinese University of Hong Kong. He received PhD degree from Electrical and Computer Engineering, the University of Texas at Austin in 2014. His current research interests include machine learning with applications in VLSI CAD and computer vision. He has served as TPC Chair of 1st ACM/IEEE Workshop on Machine Learning for CAD (MLCAD), served in the program committees of DAC, ICCAD, DATE, ASPDAC, ISPD, the editorial boards of ACM Transactions on Design Automation of Electronic Systems (TODAES), Integration, the VLSI Journal. He is Editor of the IEEE TCCPS Newsletter.

Prof. Yu received seven Best Paper Awards from ASPDAC 2021 & 2012, ICTAI 2019, Integration, the VLSI Journal in 2018, ISPD 2017, SPIE Advanced Lithography Conference 2016, ICCAD 2013, six other Best Paper Award Nominations (DATE 2021, ICCAD 2020, ASPDAC 2019, DAC 2014, ASPDAC 2013, and ICCAD 2011), six ICCAD/ISPD contest awards.

Enquiries: Miss Karen Chan at Tel. 3943 8439

Speaker:
Prof. TAO Yufei

Abstract:
This talk will present some results obtained by Yufei Tao and his students in recent years. These results span several active fields in database research nowadays – machine learning, crowdsourcing, massively parallel computation, and graph processing – and provide definitive answers to a number of important problems by establishing matching upper and lower bounds. The talk will be theoretical in nature but will assume only undergraduate-level knowledge of computer science, and is therefore suitable for a general audience.

Biography:
Yufei Tao is a Professor at the Department of Computer Science and Engineering, the Chinese University of Hong Kong. He received two SIGMOD Best Paper Awards (2013 and 2015) and a PODS Best Paper Award (2018). He served as a PC co-chair of ICDE 2014 and the PC chair of PODS 2020, and gave an invited keynote speech at ICDT 2016. He was elected an ACM Fellow in 2020 for his contributions to algorithms on large-scale data. Yufei’s research aims to develop “small-and-sweet”

algorithms: (i) small: easy to implement for deployment in practice, and (ii) sweet: having non-trivial theoretical guarantees. He particularly enjoys working on problems that arise at the cross-intersection of databases, machine learning, and theoretical computer science.

Enquiries: Miss Karen Chan at Tel. 3943 8439

Speaker:
Prof. WANG Liwei

Abstract:
Natural Dialogue Systems, including recent eye-catching multimodal (vision + language) dialog systems, need a better understanding of utterances to generate reliable and meaningful language. In this talk, I will introduce several research works that my LaVi Lab (multimodal Language and Vision Lab) has done together with our collaborators in this area. In Particular, I will discuss those essential components in natural dialog systems, including controllable language generation, language reasoning, and utterance rewriting, published in recent top NLP and AI conferences.

Biography:
Prof. WANG Liwei received his Ph.D. from the Computer Science Department at University of Illinois at Urbana Champaign (UIUC) in 2018. After that, he joined Tencent AI Lab, NLP group at Bellevue, US as a senior researcher, leading multiple projects in multimodal (language and vision) learning and NLP. In Dec 2020, Dr. Wang joined the Computer Science and Engineering Department at CUHK as an assistant professor. In the meanwhile, he is also serving as the Editorial Board of IJCV and program committee in top NLP conferences. Recently, his team won 2020 BAAI-JD Multimodal Dialogue Challenge and also the Referit3D CVPR 2021 challenge. The research goal of Prof. Wang’s LaVi Lab is to build multi-modal interactive AI systems that can not only understand and recreate the visual world but also communicate like human beings using natural language.

Enquiries: Miss Karen Chan at Tel. 3943 8439