The emergence of large language models, such as ChatGPT, has ushered in a new era of generative AI. With their ability to generate human-like responses, these models have unlocked vast opportunities for applications such as customer service, virtual assistants, and content creation, and fundamentally transformed the landscape of intelligent search. However, these opportunities are accompanied by unique challenges that need to be addressed to fully harness the potential of ChatGPT. While ChatGPT is a powerful tool, it lacks real-time awareness and may sometimes generate incorrect or nonsensical information. Another challenge is the potential misuse of AI-generated content. In this talk, we will delve into the challenges and opportunities presented by this new era of AI.

In recent years, knowledge graphs (KGs) have attracted tremendous interest and attention from both industry and academia, as evidenced by the many types of KGs developed including encyclopedia KGs, commonsense KGs, and KGs for medical science, covering a wide range of applications domains like search engines, question-answering and recommendations. For different application domains, however, the ways of constructing, reasoning, and manipulating KGs are quite different. In this talk, I shall introduce a collaborative project of building a university curriculum platform (called K-Cube) based on educational KGs. Among various functions and components, K-Cube supports a novel course KG construction framework guided by a standard ontology. To reduce the redundancy, we learn a backbone based on related Wiki data items and hierarchy, thereby avoiding to use named-entity recognition. As part of the reasoning, we design a machine reading comprehension task with pre-defined questions to extract relations, thereby improving the accuracy. Furthermore, KG Views are devised to support more advanced applications such as deriving instruction plans, for which two-way synchronization is supported to accommodate editing changes on the source KG and/or the derived views. In addition, KG manipulation operations including visualization (in both 2D and 3D spaces), navigation, and utilization have been developed and are to be introduced through an experimental prototype of KCube we have implemented. The ample facilities of K-Cube greatly accommodate learning path/material recommendations, effective content exploration, and efficient course management, among other advantages.

Large language models can improve productivity in many areas, including database research. In this talk, we will use examples to illustrate the potentials and limitations in the following areas: (1) natural language querying to databases; (2) data augmentation to generate more data following adhering to the schema and statistical properties of the existing databases; (3) data cleansing to identify data quality issuers; and (4) knowledge graph construction with entity extraction and resolution.

“图在大数据领域是一种重要的数据组织方式，被用于建模具有复杂连接性的数据，并在各种应用场景中有广泛的应用。在过去的十年里，学术界和工业界在图数据分析和管理方面投入了非常多精力。在今天的讲座中，我将介绍图计算中的一个子话题，图结构稳定性计算。回顾其发展。”

Artificial Intelligence is rapidly transforming our world, and ChatGPT is at the forefront of this revolution. In this talk, we will introduce the nature of ChatGPT, explore its applications and capabilities, and investigate the employment prospects in various industries brought by ChatGPT, as well as its impact on future education and the development of humanity. We will also discuss the risks and opportunities that AI brings to humanity. While AI has the potential to significantly enhance our lives, from improving healthcare and education to advancing scientific research and boosting economic growth, there are serious concerns about the impact of AI on employment, privacy, security, and ethics. As AI continues to advance at an unprecedented pace, we must carefully consider these risks and opportunities and work to ensure that AI benefits all of humanity, not just a select few. Through this talk, we hope to provide insights into the complex challenges and exciting possibilities presented by AI, and spark meaningful discussions around how we can build a better future with AI for humanity.

数据系统旨在为信息社会提供数据服务，但这些系统的使用需通过专门的语言，技术壁垒较高。近年自然语言处理能力的发展，使得用户以自然语言与数据系统进行交互成为人们的期望。近期研究人员使用深度学习技术产生交互模型，取得了一定的成功，大模型的引入也进一步提升了交互能力，但结果仍不足以达到实用的水平，究其原因在于对系统中的数据语义缺乏充分的理解。本报告对数据系统的自然语言交互任务现状与挑战进行一个简要分析，并介绍一个数据系统语义增强的自然语言交互方法，籍此提升数据系统的自然语言交互能力。

In the platform-less Web3 world, users' data will no longer be held by centralized platforms, so the values of the data will be returned to the users. However, being platform-less poses unique challenges, especially in revolutionizing the underlying data infrastructure. For example, how data can be stored without a centralized database; how users collaborate to support necessary operations of creation, reading, and update of data; how the decentralized data can be retrieved with privacy, integrity, and high efficiency; how the emerging web3 requirements be met, including validation of data ownership and tracing of usage records? In this talk, I will present the latest research and development of decentralized web3 data infrastructure and answer the above questions. More specifically, I will summarize the requirements of web3 data infrastructure, categorize the unique challenges, and propose a web3 system architecture to address the challenges. I will also introduce some of our recent research towards building the infrastructure of Web3, including collaborative edge computing and blockchain-based data sharing.

近年来知识图谱和基于深度学习的大规模图计算技术蓬勃发展，其效果在搜索、电商、社交等场景中已经得到充分验证。在网络安全垂直领域，威胁情报及终端恶意软件数据所呈现的关系天然适合用图来组织和表示，研究利用大规模图计算和图神经网络技术在百亿级安全知识图谱中进行分析和挖掘，有助于从海量安全数据中获取高价值情报信息、提前预知安全风险。本次报告议题聚焦于深度图神经网络技术在威胁情报挖掘和恶意软件行为分析两个场景的研究与实践，同时综合NLP、向量搜索等技术来发现未知安全威胁、提高人工分析效率。最后还将简单展望大语言模型与知识图谱的协同联动。

搜索与推理作为跨媒体智能的两种重要表现形式，构成了网络空间信息处理和复杂跨媒体内容理解的能力。搜索为推理提供了广泛的信息基础，而推理则进一步整合和归纳搜索得到的多源信息，为用户提供更个性化和精准化的内容。近期，大规模预训练模型在跨媒体推理、自然语言处理等领域取得了重要的进展。大模型作为一种特殊的数据结构，通过模型参数学习的方式，从海量预训练数据中隐式学习与存储了丰富的世界知识，与信息搜索和推理存在密切的联系。本次报告将探讨大规模预训练模型在跨媒体搜索与推理中的前沿技术和实践，包括如何结合大模型实现细粒度语义对齐（如LOUPE）、个性化内容可定制的跨媒体综合推理系统（如HuggingGPT）。

The outstanding performance of deep neural networks has promoted deep learning applications in a broad set of domains, including in the areas of intelligent search for ubiquitous cyberspace (Big search). However, the potential risks caused by adversarial samples have hindered the large-scale deployment of deep learning. In these scenarios, adversarial perturbations, imperceptible to human eyes, significantly decrease the model’s final performance. Many prior works have been published on adversarial attacks and their countermeasures in the realm of deep learning. It is difficult to evaluate the real threat of adversarial attacks or the robustness of a deep learning model, as there are no standard evaluation methods. Hence, with this talk, we attempt to offer the first analysis framework for a systematic understanding of adversarial attacks. The framework is built from the perspective of cybersecurity so as to provide a lifecycle for adversarial attacks and defences. In addition, we provided a case study to show the defense on a deep learning attack in this framework.

从图数据中搜索紧密连接的社区子图是图数据分析领域的一个基础性问题，其在社交网络分析、推荐系统、欺诈团伙挖掘，金融风险分析等领域具有重要应用。如何对现实应用中大图数据的社区结构进行建模，以及如何高效搜索大图数据的社区结构一直是学术界和工业界的研究热点。本次报告将主要介绍我们课题组在社区建模与搜索方法方面取得的最新研究进展，包括继承性稠密子图建模与搜索方法、时序社区建模与搜索方法、以及基于继承性稠密子图计数的高阶社区建模与搜索方法等工作。

ChatGPT为代表的大语言模型在近年来引起了极大的关注，代表了人工智能在语言理解、知识表示、逻辑推理等能力上的重大突破。受此影响，和很多其它领域一样，信息检索也正在跨入“大模型时代”，以New Bing为代表的新一代生成式检索范式正在形成，以缓和生成结果的幻觉和实时性等问题。然而这种“检索+生成”的松耦合结合方式存在很多问题，如何构建原生的、无缝结合检索与生成的大模型是一个重要的开放问题。

近年来以BERT、GPT为代表的预训练模型，使人工智能技术进入“预训练-微调”的全新范式，特别是最近ChatGPT引爆了全社会对大模型技术的关注。本报告重点介绍大模型与过去深度学习模型相比，在模型框架、微调适配以及推理计算等方面的重要特性，探讨大模型未来的研发应用范式。

在云计算时代，云原生分布式数据库因其弹性扩展、高可用、分布式等特性而获得了大量应用。为满足弹性扩展、弹性计算以及按需按量使用等企业级应用需求，云原生数据库需要探索新的体系架构，如通过分布式共享存储shared-storage来实现shared-everything的架构和存储计算分离、基于shared-nothing分布式架构来提供分布式查询和分布式事务处理。同时，金融级高可用、异地多活等技术挑战也是云原生数据库必须提供的关键能力。数据库系统在快速的向云原生化、平台化、一体化、智能化的四化方向演进。基于对以上技术挑战的不断探索和实践，我们自研了云原生数据库PolarDB，提供企业级云原生分布式数据库能力，提供一站式一体化的数据库能力。同时我们也自研了企业级云原生数据仓库AnalyticDB(ADB)。PolarDB和ADB经受了阿里巴巴双十一世界级的交易峰值挑战并在阿里云上取得了巨大的商业化成功。PolarDB同时也深度结合机器学习和安全加密等最新技术，提供面对未来下一代企业级应用从OLTP到OLAP再到HTAP的智能化、安全的云原生分布式数据库系统。