The International Standardization Committee of Networked DIKWP for Artificial Intelligence Evaluation Committee Standard Meeting
Meeting Minutes
Date: May 24, 2024
Conference organization: The International Standardization Committee of Networked DIKWP for Artificial Intelligence Evaluation
In this meeting of the DIKWP International Standard Planning Committee, the experts conducted in-depth discussions around the definitions of key terms within the DIKWP framework, the principles of standard formulation, the demarcation of cognitive and semantic spaces, and the explanation of the knowledge generation process. The meeting emphasized the breakthroughs and limitations of large models, highlighting the importance of new standards for risk management. A proposal was made to establish differentiated evaluation standards for conceptual, semantic, and cognitive spaces to accommodate the transition from conceptual to semantic space in evaluations. Experts pointed out that standards need to be defined for three spaces: the conceptual space, which involves natural language expression; the semantic space, which transcends the constraints of conceptual symbols; and the cognitive space, suggesting that all three spaces might need to be interconnected. Different standards would be needed for each space since many existing standards are aimed at the conceptual space.
Regarding the definition of knowledge, experts discussed the comprehensiveness of the definition, emphasizing the need to clarify the subject of knowledge acquisition, its operational and interpretative aspects, and ensure the mathematical logic of the definition. Reflecting on the definitions of data and information, one expert proposed changing "rough understanding" to "understanding" and revising overly colloquial parts of the expression. Another expert suggested deleting the "academic definition" title, which was agreed upon by others. Some experts suggested discussing how knowledge is generally produced or expressing such interactions in other parts. A comprehensive explanation of the DIKWP network relationships was deemed beneficial. One expert mentioned that while this content involves significant work, it could be updated in future versions. Other areas besides data also need revisions. Another expert remarked that knowledge is essentially a product, and its production process should resemble a Douglas function, involving inputs, the model, and the resulting knowledge.
In the compilation notes, one expert suggested changing “assessing AI's understanding and processing ability in different cognitive networks” to "cognitive network levels," emphasizing that networks have levels and are not equal. The phrase "focus on the fairness, justice, and equality of models" should be revised to "focus on the sharing, fairness, justice, and privacy protection of models." It was proposed to establish a voting rule to ensure fairness since decisions are made freely. Drawing on experience from other related meetings, a two-thirds majority vote of all attendees was considered a straightforward and effective decision-making mechanism. However, due to the special nature of our meetings, with varying attendees and possible gaps in understanding previous discussions and modifications, a more flexible and fair decision-making approach was suggested. It was proposed that if five or more attendees agree, the decision can be considered valid. While this method may not guarantee 100% democracy and fairness, it reflects the majority's will. After discussion, it was agreed to revise the terms to "cognitive network levels" and "focus on the sharing, fairness, justice, and privacy protection of models," with each modification receiving unanimous expert agreement.
One expert suggested changing "its ability to relate cognitive, conscious, semantic, and conceptual spaces" to "through bionics, relating conceptual, cognitive, semantic, and conscious spaces," emphasizing the imitation of human perception and cognition processes. Since bionics involves perception through senses like eyes, ears, nose, tongue, body, and brain, another expert noted that incorporating bionics would add constraints, potentially excluding non-bionic methods. Hence, the term "bionics" was suggested to be removed, which was agreed upon. Experts emphasized the order of the four spaces, and through a vote, it was finalized to be "conceptual space, semantic space, cognitive space, and conscious space."
For improving knowledge structuring representation, experts recommended integrating the concept of weight in semantic networks, using graph theory to clarify the strength of relationships between nodes. It was noted that weights reflect not just data but also information relationships, suggesting exploring the comparison of the strength of relationships between nodes rather than a single weight value. One expert highlighted the traditional approach to representing weights between nodes. Considering the DIKWP framework's integrity, if weights fall under the data category, it's challenging to ensure standard integrity in our definitions. Weights may involve data and information relationships, converting relationships between knowledge nodes into data measurable by weight. Instead of weights, the strength comparison of information relationships might be more meaningful.
When layering, it enters the conceptual space because only in the conceptual space can layering occur. Layering is an essential engineering step, as it allows conscious processing and symbolization, making it more expressible. Although many brain activities are unconscious or subconscious, layering enables handling and conceptualizing. The aim is to process concepts, considering their transformation relationships, to achieve comprehensive discourse. Bridging these in the semantic space through conversion reflects knowledge expression in the conceptual space. While the semantic space is indefinite, the conceptual space is defined and serves as a tool.
An expert noted that AD represents traditional weights. AI hasn't obtained the transformation relationship between two nodes or dataization, but it can assess the strength difference. The goal is to make the AD system more automated and aligned with AI systems. Although AD systems handle all issues, subjectivity and threshold settings pose challenges. Dataization of cognitive or semantic content can lead to semantic loss, an inevitable loss in the DIKWP semantic space. The subjective objectification process, retaining subjectivity but not always needing numerical values, can utilize partial order relations without dataization, requiring an expansion in this area.
In the meeting summary, experts emphasized that the standard should outline general principles, forming universally applicable guidelines rather than specific algorithms, to enhance international applicability and accommodate different scenarios. Conceptual space mathematical modeling involves categorizing data concepts and relationships, distinguishing between DIKWP data concepts and information concepts.
Suggestions for follow-up actions:
l Continue refining the definitions of knowledge and subsequent stages (e.g., from knowledge to wisdom).
l Implement the agreed modifications and update the standard document.
l Explore balancing automation and subjectivity in standardization to reduce semantic loss.
Conclusion of the meeting:
The meeting successfully advanced the optimization of knowledge definitions within the DIKWP framework, established the direction for multi-space evaluation standards, and deeply discussed key issues in standardization, laying a solid foundation for subsequent international evaluation standards.
The editorial board: (in no particular order, in alphabetical order)
AIII International Research Institute of Artificial Intelligence, AGI-AIGC-GPT Evaluation DIKWP (Global) Laboratory, Blue Edu, Peking University, University of Science and Technology Beijing, Beijing Academy of Social Sciences, Beijing Institute of Standardization, Chengdu University of Information Science and Technology, Chongqing Police College, Dongguan Advantech Precision Manufacturing Co., Ltd., Guangxi Normal University, State Grid, Hainan University, Hainan Nuclear Power Co., Ltd., Hainan Universal Intelligent Technology Co., Ltd., Hainan Provincial Market Supervision Administration, the Second Affiliated Hospital of Hainan Medical College, Huazhong Agricultural University, Jiangsu Li Zhuo Information Technology Co., Ltd., Kenside (Zhuhai) Co., Ltd., People's Procuratorate of Liaoyang City, Liaoning Province, Nanjing Police Academy, Inner Mongolia University, Ningbo University, Tsinghua University Research Institute, Shandong University, Shanxi Provincial Bureau of Data, Shanghai Aerospace Information Institute of Science and Technology, Shangrao Normal University, Sangfor Corporation, World Artificial Consciousness Association, World Artificial Consciousness Conference, Tai Chi Computer Co., Ltd., Xi'an University of Technology, Southwest University of Political Science and Law, Guangdong-Hong Kong-Macao Greater Bay Area Standardization Research Center, China Standardization Institute Standardization Theory Institute of Strategic Studies, China Mechatronics Technology Application Association, China Academy of Information and Communications Technology, etc.
转载本文请联系原作者获取授权，同时请注明本文来自段玉聪科学网博客。
链接地址：https://wap.sciencenet.cn/blog-3429562-1435962.html?mobile=1
收藏
