Title: A Paradigm Shift in Machine Learning: From Data to Data- Knowledge Environment
题目：机器学习范式的转变：从数据到数据-知识环境
Abstract: Machine Learning (and AI) are inherently data-driven. Data are a lifeblood of design methodologies and drive current commonly encountered development practices. The usage of data is behind spectacular successes of AI but also leads in some far-reaching failures. From the usage perspective in the ML learning environment, data and knowledge are evidently different. Data are numeric and precise. Knowledge is general and usually expressed at the higher level of abstraction bringing an aspect of information granularity. We carefully revisit the key and promising trends that just have recently emerged under the banner of physics-informed Machine Learning and neuro-symbolic paradigm of AI. Having this in mind, we introduce a general concept of knowledge landmarks (knowledge anchors). The landmarks provide a vehicle that along with data navigate the design of ML models. The role of knowledge is two-fold. First, it fills data gaps existing in the input space. Second, it serves as a regularization term that is essential in avoiding constructing physically infeasible models. In the design process, we elaborate on two fundamental issues, namely (i) we discuss an origin and construction of knowledge landmarks, and (ii) we present a realization of learning mechanisms in the presence of data and knowledge. Main ways of elicitation of knowledge landmarks are identified and discussed, in particular, we elaborate on techniques based on large and diversified amounts of data collected in the past and encapsulated in the form of prototypes (formed through clustering techniques). Along this line of knowledge acquisition, the relevance of the landmarks is quantified through their granularity giving rise to granular knowledge landmarks. Another alternative of building knowledge landmarks is based on their acquisition through LLMs; in this case they are qualitative and described as symbols. The constraints imposed on the data-based ML model are quantified through a collection of magnitude and change of magnitude landmarks. Once knowledge landmarks have been acquired, a unified data-knowledge environment is constructed. In virtue of the low number of knowledge landmarks, the knowledge regularization is formalized in the form of the Gaussian Process regression where the probabilistic information granules are included in the minimization of the augmented loss function. Detailed illustrative studies are showcased using rule-based architectures.
摘要：机器学习（及人工智能）本质上是数据驱动的。数据是设计方法论的生命线，并推动着当前普遍采用的开发实践。数据的使用既是人工智能取得惊人成功的关键，也导致了一些影响深远的失败。从机器学习环境的应用视角来看，数据与知识存在显著差异。数据是数值化的、精确的；而知识是概括性的，通常以更高层次的抽象形式呈现，体现了信息粒度的特性。我们系统梳理了近期在"物理信息机器学习"和"神经符号人工智能"两大范式下涌现的重要趋势。基于此，我们提出了"知识地标"（知识锚点）的核心概念。这些地标与数据协同作用，共同指导机器学习模型的设计。知识的价值体现在双重维度：其一，填补输入空间存在的数据空白；其二，作为正则化项防止构建物理不可行的模型。在设计过程中，我们重点解决两个根本问题：（1）探讨知识地标的来源与构建方法；（2）提出数据与知识双驱动下的学习机制实现方案。研究系统梳理了知识地标的主要获取途径：一方面基于历史海量异构数据，通过聚类技术形成原型进行知识封装，这种基于粒度量化的方法催生了"粒度化知识地标"；另一方面利用大语言模型获取定性表达的符号化地标。对于数据驱动的机器学习模型，我们通过"量级地标"和"量级变化地标"集合来实现约束量化。构建知识地标后，即可形成统一的数据-知识协同环境。鉴于知识地标数量有限，其正则化作用通过高斯过程回归实现——在增强损失函数的最小化过程中融入概率信息粒。最后，我们通过基于规则的架构开展了详细的案例验证研究。
Bio: Witold Pedrycz (IEEE Life Fellow) is Professor in the Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Canada. He is also with the Systems Research Institute of the Polish Academy of Sciences, Warsaw, Poland. Dr. Pedrycz is a foreign member of the Polish Academy of Sciences and a Fellow of the Royal Society of Canada. He is a recipient of several awards including Norbert Wiener award from the IEEE Systems, Man, and Cybernetics Society, IEEE Canada Computer Engineering Medal, a Cajastur Prize for Soft Computing from the European Centre for Soft Computing, a Killam Prize, a Fuzzy Pioneer Award from the IEEE Computational Intelligence Society, and 2019 Meritorious Service Award from the IEEE Systems Man and Cybernetics Society. His main research directions involve Computational Intelligence, Granular Computing, and Machine Learning. Professor Pedrycz serves as an Editor-in-Chief of WIREs Data Mining and Knowledge Discovery (Wiley), and Co-editor-in-Chief of J. of Data Information and Management (Springer).
简介：维托尔德·佩德里茨（Witold Pedrycz，IEEE终身会士）现任加拿大埃德蒙顿阿尔伯塔大学电气与计算机工程系教授，同时兼任波兰华沙系统研究所研究员。佩德里茨博士是波兰科学院外籍院士、加拿大皇家学会会士，曾获多项国际学术殊荣，包括：IEEE系统、人与控制论学会颁发的诺伯特·维纳奖、IEEE加拿大计算机工程奖章、欧洲软计算中心卡哈斯特软计算奖、基拉姆奖、IEEE计算智能学会模糊系统先驱奖，以及2019年IEEE系统人与控制论学会杰出服务奖。其核心研究方向涵盖计算智能、粒计算与机器学习三大领域。佩德里茨教授现任《WIREs数据挖掘与知识发现》（Wiley）主编，并担任《数据信息与管理学报》（Springer）联合主编。