Keynote Speakers
Academician, Chinese Academy of Engineering
Professor
Professor
IEEE, SPIE, OPTICA Fellow
Namgo Chair Professor
Member of the National Academy of Engineering of Korea
Namgo Chair Professor
Member of the National Academy of Engineering of Korea
CAE Fellow
NSERC Steacie Memorial Fellow、IEEE Fellow、EIC Fellow、AIUM Fellow、AAIA Fellow
Professor
NSERC Steacie Memorial Fellow、IEEE Fellow、EIC Fellow、AIUM Fellow、AAIA Fellow
Professor
Invited Speakers
Ian, Yi-Jen Chan
Cyntec Cooperate Inc.
CTO
Zhixuan Fang
Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua University
Hompage
Assistant Professor
Yuan Ni
平安科技
副总工程师
IEEE Fellow
Fellow of Academy of Engineering Singapore
Professor
Fellow of Academy of Engineering Singapore
Professor
Qifeng Tang
Shanghai Data Exchange
General Manager
Tammy Tang
Guangdong Grandawit Technology Co., Ltd.
General Manager
Lianfeng Wu
IDC China
VP & Chief Analyst
AAAS Fellow
IEEE Fellow
Chair Professor
IEEE Fellow
Chair Professor
Assistant Professor
Li Zhang
Hohai University
Professor
Speech Details
David Baker
University of Washington
Title:
Design of new protein functions using deep learning
Abstract:
Proteins mediate the critical processes of life and beautifully solve the challenges faced during the evolution of modern organisms. Our goal is to design a new generation of proteins that address current-day problems not faced during evolution. In contrast to traditional protein engineering efforts, which have focused on modifying naturally occurring proteins, we design new proteins from scratch to optimally solve the problem at hand. Increasingly, we develop and use deep learning methods to design amino acid sequences that are predicted to fold to desired structures and functions. We also produce synthetic genes encoding these sequences and characterize them experimentally. In this talk, I will describe several recent advances in protein design.
Yang Chai
The Hong Kong Polytechnic University
Title:
Bioinspired in-sensor computing for artificial vision
Abstract:
The visual scene in the physical world integrates multidimensional information (spatial, temporal, polarization, spectrum, etc) and typically displays unstructured characteristics. Conventional image sensors cannot process this multidimensional vision data, creating a need for vision sensors that can efficiently extract features from substantial multidimensional vision data. Vision sensors are able to transform the unstructured visual scene into featured information without relying on sophisticated algorithms and complex hardware. In this talk, I will describe our team’s efforts towards bioinspired in-sensor computing for artificial vision. I will talk about the framework of the in-sensor computing and demonstrate a few vision sensors for different scenarios, including visual adaptation, motion perception, as well as event-driven vision sensors for spiking neural network.
Ching-chuen CHAN
The Hong Kong Polytechnic University
Title:
Infinitive Possibilities in Science World - New Journey in Energy and Automotive Revolutions
Abstract:
The keynote speech begins with cultivating science spirit and understanding engineering philosophy, the contribution of science discovery, technology revolution and industry revolution to human civilization. Then discuss the key issues and challenges of energy revolution and automotive revolution. The automotive revolution is progressing from electrification to intelligent and internet connected. Electric vehicles will interlink mobility, energy and information. The friendly interaction between the vehicles and power grids, as well as the integration of people, vehicles, road and cloud are essential. The key technologies of electric vehicle power train, batteries and internet of vehicles will be discussed. The integration of energy, information, transportation and humanity will be highlighted. Our ultimate goal is the harmonious coexistence between human and nature.
Ian, Yi-Jen Chan
Cyntec Cooperate Inc.
Title:
Advanced Power Module and System for the Coming of AI Era
Abstract:
The integration of advanced power modules is becoming increasingly crucial in preparation for the coming of AI era. Key areas of investigation include the development of novel power modules capable of meeting the stringent efficiency and reliability requirements imposed by AI hardware. This presentation will highlight the solutions of 48V DC/DC power conversion for GPUs. A comprehensive overview of state-of-the-art technologies in power electronics and modules, to achieve a high efficiency and a high power density will be demonstrated and discussed.
Wing Cheung Eddie Chong
Raysolve Technology
Title:
Innovative Full-Color Micro-LED Micro-Display: A Revolutionary Technology for AR/XR industry
Abstract:
The miniaturization of LEDs has made them suitable for displays of different sizes. As a new display technology, Micro-LEDs have unique advantages in the field of micro-displays. They have high brightness, low power consumption, and high reliability, making them considered the only solution for AR/XR terminal devices. Building upon the foundation of traditional LED manufacturing processes, the development of Micro-LED micro-displays began with sapphire and flip-chip solutions, which were suitable for low-resolution displays in the initial stage. However, they faced numerous bottlenecks that couldn't meet the requirements of AR/XR displays. Additionally, due to limitations in GaN material growth and integration processes, achieving single-chip color displays has been a challenge. Now, Raysolve will redefine single-chip integration technology to achieve wafer-level full-color Micro-LED micro-display chips, making it the best micro-display technology for the future AR/XR industry.
Shu Hung Henry CHUNG
City University of Hong Kong
Title:
Wideband Series Harmonic Voltage Compensation Technology Using Look-Ahead State Trajectory Prediction for Network Stability Enhancement and Condition Monitoring
Abstract:
Today’s electric power systems have an increasing number of distributed power generation systems (DPGSs) connected to the utility grid via grid-connected inverters (GCIs). The entire system can be subject to different resonances, including local resonance caused by the GCI output filter, resonance between GCIs, and resonance with the utility grid. Extensive literature is devoted to studying stability, which is dominantly assessed by applying impedance-based stability analysis or Nyquist stability criterion. A wideband series harmonic voltage compensation technology with the operating bandwidth and impedance measurement up to 10kHz will be presented. It creates a virtually zero high-frequency impedance at the output of a plurality of GCIs, so that the impedance stability criterion can always be satisfied. Apart from enhancing network stability, a particle-swarm-optimization (PSO)-optimized multi-sine power perturbation is injected into the network by the compensation technology to determine the point-of-common-coupling impedance and equivalent impedance of the GCIs.
Yuchao Dai
Northwestern Polytechnical University
Title:
3D Reconstruction of Dynamic Scenes: Optimization, Learning and Generation
Abstract:
Dynamics are ubiquitous in the real world. Perceiving and reconstructing dynamic scenes in 3D from 2D images are of great significance, with wide-ranging applications in areas such as autonomous driving, consumer-level applications, augmented reality/metaverse, etc. 3D reconstruction of dynamic scenes aims to capture the dynamic 3D structure of observed scenes over time from continuous videos. It has evolved from explicit optimization methods to implicit representations using neural radiance fields and the latest methods based on deep diffusion models. This talk focuses on 3D reconstruction of dynamic scenes, discussing sparse reconstruction of single objects, dense reconstruction of multiple objects, and dense reconstruction of complex scenes under explicit optimization methods. Under implicit learning methods, it concentrates on novel view synthesis and 3D reconstruction of dynamic scenes. Finally, it provides an outlook on generation methods such as deep diffusion models and open problems and development trends in this field.
Zhixuan Fang
Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua University
Title:
Incentive Design and Optimization in Decentralized Network Systems
Abstract:
Decentralized network systems, such as decentralized ledger, has demonstrated powerful performance and security advantages in many applications. Researchers continue to focus on fault tolerance mechanisms in these distributed and decentralized systems, while the issues related to incentives and mechanism design still require further study. This report will discuss the game theory and mechanism design perspective in decentralized network systems, with motivating scenarios from blockchain mining and consensus design, aiming at improving the performance and robustness of these systems.
Fei Gao
Zhejiang University
Title:
生物启发的飞行智能
Abstract:
对于感知范围、计算能力、机动性能、资源功耗等受限的空中机器人系统（无人机），其在无先验信息、卫星导航拒止、障碍物稠密环境中的自主化、鲁棒化，协同化、智能化运动仍面临理论与技术上的重大挑战。观察自然界的飞行生物，我们发现鸟群可以组成庞大的集群，仅依靠它们的眼睛和大脑跨越数千公里完成自然迁徙，而鹰隼可以在长远的距离下锁定目标，穿梭于各种障碍物之间。本次报告，我将类比生物导航现象的作用机理，从敏锐感知、自主决策、飞行交互、集群智能等方面介绍团队在无人机单机与集群自主导航方法上所做出的最新创新贡献，展示在不依赖外部定位和计算设施，仅靠机载摄像头、计算芯片和传感器的无人机动态环境鲁棒感知、快速灵巧机动飞行、飞行吊载与抓握、集群避障编队等方面的最新研究成果。
Chulhong Kim
Pohang University of Science and Technology
Title:
Multi-modal Imaging: Photoacoustic Imaging Plus More
Abstract:
Trans-energy imaging modalities have been significantly explored to overcome existing problems in conventional imaging modalities with respect to spatial/temporal resolutions, penetration depth, signal-to-noise ratio, contrast, and so on. Among them, photoacoustic imaging, an emerging hybrid modality that can provide strong endogenous and exogenous optical absorption contrasts with high ultrasonic spatial resolution, has overcome the fundamental depth limitation while keeping the spatial resolution. The image resolution, as well as the maximum imaging depth, is scalable with ultrasonic frequency within the reach of diffuse photons. In this presentation, the following topics will be discussed; (1) multiscale and multiparametric trans-energy imaging systems, (2) novel deep-learning powered image processing, (3) recent clinical study results in pathology, endocrinology, oncology, cardiology, dermatology, and radiology, (4) label-free ultrafast ultrasound Doppler imaging, and (5) efforts to commercialization.
Erping Li
Zhejiang University
Title:
An Electromagnetic-Information-Theory for Internet-of-Things Scenarios
Abstract:
Electromagnetic information theory (EIT) is a promising framework applied in the next-generation Internet-of-Things (IoT) to overcome the shortcomings of either electromagnetic-only or information-only framework. However, efficient algorithms as well as clear EIT analysis methods for multi-channel characterization in complex electromagnetic space are still lacking. This presentation will touch on an EIT-based model is developed for the multi-channel mode analysis. The group-T-matrix-based multiple scattering fast algorithm, the mode-decomposition-based characterization method, and their joint theoretical framework in scattering environments are discussed, a case of image transmission with limited power is presented to illustrate how to use this EIT-based model to guide the electromagnetic design for real IoT applications.
Jinfei Liu
Zhejiang University
Title:
数据要素定价研究进展
Abstract:
数据是21世纪的“新石油”。数据要素市场连接起数据拥有者、数据中间商和数据购买者，为数据要素交易流通提供平台。数据要素定价是数据要素市场最关键的环节之一。近年来，数据要素定价在学术研究、政策法规和工业企业取得了蓬勃发展。本报告关注数据要素定价中的两个核心问题，数据产品定价与数据贡献评估，具体介绍如何根据不完全信息博弈理论实现数据产品定价，以及如何基于合作博弈Shapley值对数据产品各贡献方实现公平的数据贡献评估。
Shixia Liu
Tsinghua University
Title:
Data-Centric Explainable Artificial Intelligence: A Visual Analytics Perspective
Abstract:
The quality of training data is crucial to the success of supervised and semi-supervised learning. Errors in data have long been known to limit the performance of machine learning models. This talk presents the motivation and major challenges of data quality diagnosis and improvement. With that perspective, I will then discuss some of our explainable machine learning efforts from the data perspective: 1) analyzing and correcting poor annotation quality, 2) resolving the poor coverage of the training data caused by dataset bias, 3) enriching training data by fusing multimodal information.
Cewu Lu
Shanghai Jiao Tong University
Title:
Embodied Intelligence - Perception, Imagination, Execution (PIE) Framework and Embodied Large-Scale Model Exploration
Abstract:
TBD
Haojian Lu
Zhejiang University
Title:
面向医疗应用的微型软体机器人
Abstract:
微型机器人由于其尺寸小、运动灵活、环境适应能力强的特性，被广泛应用于生物医疗领域的研究，如生物组织培养、微创手术、靶向送药等。在微型医疗机器人通往体内实际应用的道路上，除去机器人本体的驱动与运动控制之外，另一个核心部分就是功能化，包括了对体内环境或病变区域的感知和测量，对多种类型药物的搭载和可控释放。本报告以建立可控、多功能、高性能的微型医疗机器人系统为目标，针对面临的操作可控性、医疗功能性、设计加工可靠性等问题，主要介绍面向体内复杂环境的微型机器人设计与运动控制、微型机器人无线传感与通讯集成设计、基于生物信息学的微型机器人搭载药物挖掘及筛选，旨在解决微型医疗机器人多功能集成、高效可控运动、智能化医疗问题，推进微型医疗机器人在临床上应用。
Yipeng Lu
Peking University
Title:
压电MEMS超声波传感器
Abstract:
超声换能器被广泛应用于医疗成像、工业无损检测、汽车超声波雷达等。基于体压电材料的传统超声换能器具有体积大、不易加工、低带宽等缺点，限制了其在很多高端智能系统以及消费电子的应用。针对以上迫切需求，MEMS技术为超声换能器的发展与应用注入了新的动力，在降低大批量生产成本的同时，实现了低功耗、小型化、一体化集成的高性能微型超声换能器（MUT）阵列。MUT把超声技术的应用推向了新的高度，实现了其在智能手机、汽车电子、智能家居、自动驾驶、机器人以及医疗器械等新兴领域的应用包括超声指纹识别传感器、XR/元宇宙的人机交互、即使诊断的超声成像设备、以及超声可穿戴等。本报告会侧重于介绍MUT的前沿进展包括其机理、工艺、系统集成、应用研究等。
Youyou Lu
Tsinghua University
Title:
SuperFS: A File System for High-Performance Hardware
Abstract:
数据是当前信息技术发展的主要驱动力。云与大数据、超算以及智算应用处理数据规模越来越大，存储系统的性能瓶颈问题日益显现。本报告将介绍本课题组研制的新一代分布式文件系统SuperFS。SuperFS在2022年SC公布的国际超算500强（IO500）10节点元数据榜单上获得了第一名；在2023年ISC公布的国际超算500强（IO500）榜单上，SuperFS部署于“鹏城云脑II”上，获得了总榜单第一名。本报告将介绍SuperFS在面向高速存储设备和网络设备上的软硬件协同设计，包括数据存储中的数据拷贝、异步访问等技术，以及元数据存储中扁平目录结构、并行目录查找以及软硬件协同的元数据组织与结构设计等技术。
Yuan Ni
平安科技
Title:
人工智能技术赋能医疗转型
Abstract:
随着科技的飞速发展，人工智能技术在医疗领域的应用日益广泛，为医疗行业带来了前所未有的转型机遇。本次演讲，我将分享平安科技在医疗知识图谱和大模型领域的探索，以及将这类技术应用在在线问诊、慢病管理、体检解读等多个领域的实践经验。 平安科技作为我国领先的人工智能企业，致力于利用先进技术为医疗行业提供解决方案。在医疗知识图谱方面，我们通过构建大规模的医疗知识图谱，实现了医疗信息的结构化、智能化，为医生和患者提供更加精准、全面的医疗信息服务。同时，大模型技术在医疗领域的应用也取得了显著成果，通过深度学习等算法，实现对医疗数据的智能分析，提高疾病预测、诊断和治疗的准确率。在实际应用场景中，我们将这些先进技术应用于在线问诊、慢病管理、体检服务等多个领域。为客户提供“省时、省心、省钱”的健康管理服务。人工智能技术在医疗领域的应用将为行业带来深刻的变革，提高医疗服务质量，降低医疗成本，为患者带来更好的就医体验。
Tony Quek
Singapore University of Technology and Design
Title:
A Pathway towards Future Network Intelligence: RAN Intelligent Controller meets Digital Twin Networks
Abstract:
The RAN intelligent controller (RIC) is cloud native and a central component of an AI and virtualized RAN network. The RIC enables to deployment of machine learning and AI techniques to optimize resources and services in the RAN. Thus, it is an important component that brings intelligence, agility, and programmability to the radio access network. On the other hand, digital twin networks (DTW) are real-time replicas of physical networks. They are emerging as a powerful technology for design, diagnosis, simulation, what-if-analysis, and AI-driven real-time optimization and control of 6G wireless networks. In this talk, we will share our journey in building up AI RAN networks that allow us to bring together RIC and DTW to better understand how to design future wireless networks. Furthermore, we will also share Singapore’s first national Future Communications Research and Development Programme (FCP), which funds our work in this area.
Haiding Sun
University of Science and Technology of China
Title:
Monolithic Integration of III-Nitrides-based Optoelectronic Devices
Abstract:
The III-nitride family (AlN, GaN, InN, and their alloys) has been regarded as one of the most significant semiconductors, and has therefore been heavily investigated in the past two decades for various optoelectronics (light emitting diodes, lasers, photodetectors, etc.). Excitingly, the InGaN-based light emitters have revolutionized the lighting industry, promoting energy-efficient and eco-friendly modern solid-state lighting and lasing technologies. Similarly, by alloying GaN with AlN, we can synthesize AlGaN ternary alloys with tunable direct bandgaps from 3.4 eV (GaN) to 6.1 eV (AlN), corresponding to a wide UV spectral range from 360 nm to 210 nm, for a broad range of applications including air/water purification, UV curing, data storage, optical communication, etc. Hence, the AlGaN alloys have attracted enormous attention in the nitride community for use in efficient AlGaN-based UV LED, lasers and photodetectors. In this presentation, we will focus on the development of miniaturized AlGaN-based highly efficient UV LEDs and detectors with monolithic integration of these optoelectronics on silicon platform for on-chip and free-space optical communications.
Dean Ta
Fudan University
Title:
多模超声-光声骨成像方法及仪器
Abstract:
报告介绍了跨尺度高分辨率的多模态超声、光声骨成像方法及仪器研制。主要内容有：骨定量超声/光声新型成像原理、方法和技术；骨超声和光声多模态成像算法、图像融合算法；研制的跨尺度高分辨率的多模态超声骨成像仪器。该仪器实现了信号获取、参数成像及分析、特征提取、生化成分解析及全波反演超声成像、复合平面波仿真成像等功能。最后指出了存在的问题及努力的方向。
Qifeng Tang
Shanghai Data Exchange
Title:
Corporate data asset management from the perspective of data elements
Abstract:
1.The multiplier effect of data elements and new quality productive forces 2.The process of data assetization and the path to value realization 3.Corporate data utilization and data asset management methods 4.Practice and innovation of Shanghai Data Exchange in corporate data asset management
Tammy Tang
Guangdong Grandawit Technology Co., Ltd.
Title:
Industrial Al: The Digital Innovation Engine for Manufacturing
Abstract:
In this presentation, I will share how Artificial Intelligence (AI) drives the digital transformation of the manufacturing industry and leads the innovative development of enterprises. Firstly, I will introduce the development trend of generative AI in the global manufacturing industry. Through a series of actual scenario cases, I will elaborate on how AI technology helps experts in the enterprise field to enhance their ability to analyze complex problems and predict the future, and how to realize experience retention and data assetization. Finally, I will present the development path of generative AI in the manufacturing industry and the future outlook of the industrial metaverse.
Meixia Tao
Shanghai Jiao Tong University
Title:
AI-Empowered Coding and Transmission for Semantic Communications
Abstract:
As a new communication paradigm beyond Shannon, semantic communication (SemCom) can significantly reduce the required communication bandwidth and enhance downstream task performance by extracting and transmitting information directly relevant to the receiver's tasks. SemCom is envisioned to have transformative potential in 6G, facilitating a wide range of intelligent services, such as metaverse, smart surveillance, intelligent transportation, and robotic collaboration. In this talk, I will first provide an overview of the deep-learning-enabled SemCom frameworks as well as the main design challenges. Then I will introduce our latest research progress on novel designs of coding and transmission for SemCom towards practical implementation. These include: transmit data preprocessing via domain adaptation, receiver channel denoising via diffusion models, joint coding and modulation for digital semantic transmission, and codebook-assisted semantic coding.
Yijie Wang
Harbin Institute of Technology
Title:
Simultaneous wireless power and data transfer technology
Abstract:
This report presents a simultaneous wireless power and data transmission system with full-duplex multiple-input multiple-output (MIMO) communication channels. The system utilizes power coils to transfer data without the need for additional antennas and constructs two independent communication channels to form 2×2 MIMO communication channels. Compared to systems with only a single communication channel, high-priority information, such as control signals and feedback signals, can be transmitted by a dedicated channel to improve communication reliability. In addition, the use of two channels also enhances the communication speed, and data can be transmitted in parallel over these two channels. To address the challenge of designing high-order circuits in data channels, its reduced-order equivalent model is proposed, and a 1.1 kW prototype prototype with two 1 Mbps full-duplex channels and a total transmission rate of 2 Mbps is experimentally constructed to validate the feasibility of the proposed method.
Chentao Wu
Shanghai Jiao Tong University
Title:
大规模存储设备故障预测的研究
Abstract:
当前云计算、大数据系统需要用到海量的存储设备，但存储设备的故障会导致系统宕机和服务的不可用，极大影响上层应用的平稳运行。本报告针对大规模存储设备故障预测算法开展研究，重点介绍SSD和内存故障预测的方法和机制，以提高大规模云计算、大数据系统的可靠性。
Dan Wu
Zhejiang University
Title:
Microstructural Imaging of the brain and tumor
Abstract:
Diffusion MRI provides an noninvasive tool to probe tissue microstructural information. This talk will introduce our recent effort in the development of diffusion MRI pulse sequence, microstructural models, and high-performance graident system to enhanced the ability for microstructural mapping. The diffusion MRI based microstructural mapping has demonstrated its clinical value as a "virtual biopsy" for the diagnosis of cancer and brain disordrs.
Lianfeng Wu
IDC China
Title:
Can Large Language Models and Generative AI Create the Future?
Abstract:
Countries, industries, enterprises, and individuals are all entering a major AI transformation period. Those who develop and utilize generative AI better will have greater competitiveness in the future. By leveraging generative AI, countries can develop new productive capacities, enterprises can innovate business models, schools can transform teaching methods, and individuals can enhance labor productivity. However, today's generative AI is still in its early stages; using it well can create the 'future,' while poor utilization may lead to elimination. Mr. Wu's speech first outlined the current state and future trends of generative AI, reviewed the value it brings, highlighted the challenges faced in applying generative AI, and proposed main strategies for accelerating the deployment of generative AI to create value.
Zhengxing Wu
中国科学院自动化研究所
Title:
水下智能仿生机器人系统
Abstract:
鱼类经过漫长的自然进化，形成了卓越非凡的游动能力。仿生机器鱼以自然界鱼类为研究对象，旨在揭示鱼类高速、高机动及高效率游动的奥秘，为新型水下智能移动机器人系统的开发及应用，提供理论基础和技术支撑。本报告将重点介绍团队近年来在仿生机器鱼方面取得的研究进展，包括以鱼类、海豚、蝠鲼等生物为仿生原型，所开展的运动机理分析、仿生系统设计、智能运动控制等内容，最后分析和展望未来的研究方向和发展前景。
Wen Xia
Harbin Institute of Technology, Shenzhen
Title:
一种面向非易失内存的数据去重文件系统研究
Abstract:
NVM因其优良内存与存储特性有望成为下一代存储介质。但是目前NVM价格昂贵，面向NVM的数据去重技术的研究变得格外重要。现有的NVM去重文件系统无法充分发挥NVM的IO特性从而低估了NVM的价值，因此本工作提出Light-Dedup以充分发挥NVM的IO特性。Light-Dedup主要具备两个数据关键技术点：（1） 通过结合非加密哈希与预测预取字节级比对实现高性能冗余数据检测；（2） 通过精心设计粗粒度元数据布局大大降低去重元数据读写导致的IO放大问题。实验表明Light-Dedup能够提升至多7.98倍去重吞吐，并能够将元数据IO放大维持在较低水平。
Ying Xu
Southern University of Science and Technology
Title:
Studies of Disease Biology Need New Theoretical Frameworks
Abstract:
Chronical diseases, such as cancer, Alzheimer’s disease, and diabetes, generally have considerable changes in their cellular chemical conditions such as the pH and the O2 level, which result in changes in the cellular physical conditions, including the membrane potential and the intracellular polarity. These changes will profoundly alter the kinetics and thermodynamics of cellular chemical reactions, leading to the so-called metabolic reprogramming (MR) at a systems level, created by stress-induced genetic mutations and epigenetic alterations for survival. The affected cells may have to make drastic and fundamental changes, such as the significant simplification of their polarity system (which defines what a cell can and cannot do) via mutations as in cancer, to generate sustained metabolic exits for the newly created MRs, which give rise to specific phenotypes of a disease. To elucidate the operating logic of each disease type, or specifically how the biological pathways encoded in our genome behave in a fundamentally distinct microenvironment from the normal physiological conditions, we need to have a new research framework, which requires, at least, consideration of biology at the basic chemistry level.
Junchi Yan
Shanghai Jiao Tong University
Title:
Machine learning for Constrained Problems: Examples on Combinatorics and Autonmous Driving
Abstract:
In this talk, the speaker will introduce some recent works on machine learning of solving constrained problems, with two typical examples: 1) combinatorial optimization and 2) autonomous driving. For the former, some neural architectures are devised for end-to-end problem learning and solving; for the latter, model-based reinforcement learning is adopted with a latent world model. The hope is that the talk could attract more attention to the constrained problem solving via machine learning.
Alfred Yu
University of Waterloo
Title:
Deep Learning for Smart Design of Ultrasound Systems
Abstract:
Ultrasound is undoubtedly a popular medical imaging modality and is becoming known for its high-frame-rate imaging capabilities. However, high-frame-rate ultrasound has yet to flourish in point-of-care applications due to the lack of suitable portable hardware, and its ability to offer time-resolved flow visualization is hampered by Doppler aliasing artifacts. Can we leverage advances in deep learning and artificial intelligence to overcome bottlenecks in ultrasound system design? Can we train neural networks to detect Doppler aliasing artifacts and correct them in real time? This talk will introduce our laboratory’s quest to learn deep and learn smart about ultrasound imaging systems to make high-frame-rate ultrasound viable for portable use and flow estimation. We will demonstrate how deep learning solutions can be devised to resolve data transfer bottlenecks in ultrasound systems and, in turn, enable robust generation of high-frame-rate ultrasound images with data acquired from few array channels. We will also show how deep learning has enabled the design of advanced Doppler flow imaging platforms that remain resilient against higher-order Doppler aliasing artifacts with multi-cycle wraparounds. Related algorithms, real-time engineering efforts, and clinical applications will be presented throughout the presentation.
Huishuai Zhang
Wangxuan Institute of Computer Technology, Peking University
Title:
How do Momentums Accelerate Optimizers? An Empirical Study
Abstract:
Momentum has become a crucial component in deep learning optimizers, necessitating a comprehensive understanding of how it accelerates stochastic gradient descent (SGD). For the question of “when”, we establish a meaningful comparison framework that examines the performance of SGD with Momentum (SGDM) under the effective learning rates. For the question of “why”, we find that the momentum acceleration is closely related to a sudden jump of the directional Hessian along the update direction. Momentum improves the performance of SGDM by preventing or deferring the above sudden jump. Additionally, in our study of the newly developed optimizer "Lion," we also show the significant impact of evolved momentum design on the performance of adaptive optimizers in the context of optimizing transformers.
Jie Zhang
Peking University
Title:
近数据计算的研究进展
Abstract:
随着大数据时代的到来，人工智能、图计算、大数据等新型应用对服务器集群的算力和存储能力提出了更高的要求。然而，传统的冯诺依曼体系结构及配套的系统软件存在数据迁移开销大的天然劣势，无法满足新型应用的实际需求。当今的内存和存储系统正经历了重大的技术转变。基于这种技术的提升，研究人员需要重新思考和设计现有的系统组织和硬件架构。本次报告主要分享我们在近数据计算领域的研究进展，我们提出的解决方案能够有效减少大量软件栈的开销并且优化计算机体系结构消除传统硬件的物理限制。
Li Zhang
Hohai University
Title:
Modulation and Control of Grid-connected Inverters Using SiC Devices
Abstract:
This presentation primarily introduces the modulation and control of three-level grid-connected inverters using a combination of Si and SiC devices. It covers the commutation strategy of Si and SiC hybrid bridge arms, three-level discontinuous pulse width modulation (DPWM) technology, and grid-connected control methods for non-ideal grid conditions. The aim is to improve the efficiency and power density of grid-connected inverters at a low cost while enhancing their adaptability to grid conditions. Finally, the practicality of the proposed technology is validated through an engineering prototype.
Wenjun Zhang
Shanghai Jiaotong University
Title:
Generative Video Communication
Abstract:
Video communication is the technology that enables efficient transmission of video information. With video traffic now accounting for about 80% of all internet traffic, the significance of video communication has become increasingly prominent. However, the current progressive evolution of video communication technology faces a series of challenges such as saturation in compression efficiency, increased transmission energy consumption, and limited service formats. It is imperative to break through the existing video communication architecture and develop new momentum to support the continuous rapid development in the field of video communication. Therefore, this talk introduces the potential impact of generative artificial intelligence technology on video communication, focusing on the concept and important connotations of generative video communication, especially in the application scenarios of fully referenced, semi-referenced, and non-referenced generative video communication. Additionally, this talk will present our related work on generative video communication in areas such as elastic coding, collaborative transmission, and utility evaluation. Preliminary results have validated the significant role of generative video communication in improving coding efficiency, enhancing transmission capabilities, and enriching evaluation dimensions.
Yiming Zhang
厦门市智能存储与计算重点实验室
Title:
Efficient virtualized storage
Abstract:
KVM（基于内核的虚拟机）是Linux上主要的VM管理程序，它遵循标准virtio框架来对客户VM的I/O设备进行半虚拟化。传统上，KVM依赖QEMU来实现virtio设备族的后端，比如virtio-blk/-net，其中KVM（内核空间）和QEMU（用户空间）之间的协作是实现安全灵活存储管理（如迁移）的关键。上述方法的缺点是，基于KVM/QEMU的半虚拟化使得客户I/O路径中有多个上下文切换，延长了I/O操作时间。随着快速NVMe存储设备的广泛使用，对于基于KVM/QEMU的virtio-blk存储半虚拟化，软件开销显著。为了缩短I/O路径，virtio-blk的变种vhost-kernel-blk和vhost-user-blk分别在内核和用户空间执行所有客户I/O处理。然而，这种方式放弃了KVM和QEMU之间的协作，从而影响了管理灵活性、系统安全性和兼容性。针对上述问题，我们对虚拟化存储技术进行研究，显著提高了半虚拟化的效能。
YunYan Zhang
Zhejiang University
Title:
III-V nanowire technologies on Si for photonic and electronic applications
Abstract:
Recently, nanowires (NWs) with a one-dimensional (1D) columnar shape have gained great attention and have been envisioned as nanoscale materials for next-generation technology with good functionality, superior performance, high integration ability and a potential for low cost. Due to those advantages, the limitations, caused by the requirement of lattice and thermal expansion coefficient in the traditional thin film epitaxial technique, can be greatly alleviated, giving more freedom in material combinations, offering more flexibility in the band structure engineering and providing new theories for the device structure design. For example, III−V NW structures can be grown on a Si platform, thus enabling fabrication of high-quality photonic light sources for low-cost, ultra-high-density integration, solving one of the major challenges currently limiting the exploitation of Si photonics. We have demonstrated the self-catalyzed III-V (GaAsP, InAsP, InAsSb, AlGaInN et al) NW growth on Si substrates by using molecular beam epitaxy and achieved for the first time stacking-fault-free zinc blend crystal structure (Figure 1). - Many advanced structures have been developed, such as the defect-free deep quantum dots, deep quantum wells, horizontal NWs, branched NWs et al. After the surface passivation with an ultrathin InP layer of 2−3 nm, these NWs have their surface states reduced by at least 70% and hence show good room temperature performance with a very long-term stability (at least 3 years in ambient atmosphere). Using these NWs, several high-performance optoelectronic devices have been developed, such as NW lasers with the highest efficiency/lowest threshold at room temperature,7 high-temperature single photon sources,9 detectors with record-high detectivity and sensitivity, photovoltaics with a record-high fill factor, solar water splitting cells with a long lifetime . These works set a good foundation for developing high-performance optoelectronics on the cost-effective Si platform in a CMOS compatible way, which is promising to make significant breakthrough to a wide range of fields, such as Si photonics, and quantum computing.
Haitao Zhao
Northwestern Polytechnical University
Title:
A Novel Progressive Wave Gyroscope Based on Acousto-optic Effects
Abstract:
The gyroscope is an indispensable part of inertial navigation system, which plays an irreplaceable role in extensive potential applications, including industrial engineering and consumer electronics. However, in some applications involving extreme environments (e.g., overloads exceeding 20,000g), the performance of popular MEMS gyroscopes significantly degrades, affecting their accuracy, sensitivity, and stability. As an all-solid-state structure, a surface acoustic wave (SAW) gyroscope completely eliminates the movable mass and suspending support structure found in MEMS gyroscopes, providing excellent impact resistance. However, the extremely weak Coriolis force significantly reduces the sensitivity of SAW gyroscopes, presenting substantial challenges for practical applications in angular velocity measurement. In this presentation, we introduce a brand-new, high-sensitivity surface acoustic wave (SAW) gyroscope based on acousto-optic effects. Unlike traditional SAW gyroscopes, this study utilizes acousto-optic effects to detect changes in the refractive index caused by mechanical strain, measuring angular velocity through the output optical power intensity of the optical waveguide. This innovative structure combines the advantages of both conventional microscale vibrating gyroscopes and optical gyroscopes, offering a feasible and insightful solution for enhancing the performance of SAW gyroscopes.
Guoyan Zheng
Shanghai Jiao Tong University
Title:
Deep Learning for Surgical Video Analysis
Abstract:
Accurate analysis of surgical videos is of great significance to ensure the safety and precision of treatment of patients, especially in minimally invasive surgery. Typical tasks include surgical workflow analysis, instrument recognition, surgical scene reconstruction, and future event prediction. Although data-driven deep learning techniques have played an important role in the field of medical image analysis, surgical video analysis is still facing many challenges. In this talk, we will discuss recent progress that we have achieved on intelligent surgical video analysis using deep learning-based methods.
Weishi Zheng
Sun Yat-sen University
Title:
Visual Perception and Learning for Robot Behavior
Abstract:
Visual behavior perception and learning for physical robots is an important part of embodied intelligence research. This talk will report the related research carried out by our research group on general robot free grasping modeling, including robot 6-DOF grasping modeling, dexterous hand grasping modeling, dynamic grasping and multi-robot cooperation, and robot behavior quality assessment.
Miao Zhu
Shanghai Jiao Tong University
Title:
Novel Power Electronics Equipment and Operation Principles towards Future Offshore DC Network
Abstract:
In recent years, with the rapid development of renewable energy and power electronics, DC has garnered significant attention in R&D of power systems. This lecture will revolve around the future offshore DC network, focusing on basic theories and key techniques of the power electronics equipment embedded within it. In this lecture, two types of advanced power electronics equipment will be elaborated: multi-port DC substations and DC power flow controllers. Their respective operation principles and application mechanisms in offshore DC networks will also be discussed. Power electronics equipment serves as the core and foundation of future offshore DC networks, as they can be tailored to meet specific demands of various electrical nodes. This lecture will provide insights into principles and characteristics of DC networks, and is suitable for researchers and developers in relevant fields of electrical engineering.
