Witold Pedrycz is a Professor and Canada Research Chair (CRC) in Computational Intelligence in the Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Canada. Dr. Pedrycz is a foreign member of the Polish Academy of Sciences and a Fellow of the Royal Society of Canada. He is also an author of 15 research monographs and more than 500 research publications. Dr. Pedrycz is the Editor-in-Chief of Information Sciences and Editor-in-Chief of WIREs Data Mining and Knowledge Discovery (Wiley).
Explainable AI: From Data to Symbols and Information Granules
With the progress and omnipresence of Artificial Intelligence (AI), two aspects of this discipline become more and more apparent. When tackling with some important societal underpinnings, especially those encountered in strategic areas, AI constructs call for higher explainability capabilities. Some of the recent advancements in AI fall under the umbrella of industrial developments (which are predominantly driven by numeric data). With the vast amounts of data, one needs to resort herself to engaging abstract entities in order to cope with complexity of the real-world problems and delivers transparency of the required solutions. All of those factors give rise to a recently pursued discipline of explainable AI (XAI). From the dawn of AI, symbols and ensuing symbolic process have assumed a central position and ways of symbol grounding become of interest. We advocate that in the realization of the two timely pursuits of XAI, information granules and Granular Computing (embracing fuzzy sets, rough sets, intervals, among others) play a significant role. The two profound features that facilitate explanation and interpretation are about an accommodation of the logic fabric of constructs and a selection of a suitable level of abstraction. They go hand-in-hand with the information granules. First, it is shown that information granularity is of paramount relevance in building linkages between real-world data and symbols encountered in AI processing. Second, we stress that a suitable level of abstraction (specificity of information granularity) becomes essential to support user-oriented framework of design and functioning AI artifacts. In both cases, central to all pursuits is a process of formation of information granules and their prudent characterization. We discuss a comprehensive approach to the development of information granules by means of the principle of justifiable granularity. Here various construction scenarios are discussed including those engaging conditioning and collaborative mechanisms incorporated in the design of information granules. The mechanisms of assessing the quality of granules are presented. In the sequel, we look at the generative and discriminative aspects of information granules supporting their further usage in the AI constructs. A symbolic manifestation of information granules is put forward and analyzed from the perspective of semantically sound descriptors of data and relationships among data. With this regard, selected aspects of stability and summarization of symbol- oriented information are discussed.