SOPhiA 2021

Salzburgiense Concilium Omnibus Philosophis Analyticis

SOPhiA ToolsDE-pageEN-page

Programm - Vortrag

Urteilsenthaltung und kuenstliche Intelligenz
(Epistemology, )

The rising applicability of artificial intelligence is accompanied by the problem of explainability. Especially current learning-based artificial systems prove to be highly efficient, but often do not allow for insights into the decision-making process. This contributes to serious problems such as the possibility of biased decisions, and the lacking of understanding and trust. On a different node, there is the fairly recent development in epistemology of considering the third doxastic stance, which is a neutral position between acceptance and rejection, to be a proper object of investigation in its own,. In this paper, we bring the two fields together. We argue that the discussion about suspension (of judgement), as the third neutral stance in epistemology, can fruitfully be applied to the area of artificial intelligence. To understand the demand for explaining a system_s decisions, it is reasonable to look first at those cases where decisions are (or should be) absent or postponed and the different reasons for this. It is argued one precondition for possibly generating meaningful and accurate explanations of an AI system_s choices, is for the system to recognise its own knowledge limits. Here, the recent debate about suspension can help to understand the different situations of a system reaching its knowledge limits and can serve as a tool to communicate this. We want to illustrate this possibility to apply suspension in AI within non-monotonic reasoning, which often serves as the formal framework of symbolic, logic-based AI systems. In a second step, this can also contribute to the discussion of explainability for non-symbolic, data-based AI systems, when neural-symbolic approaches, that try to combine symbolic and non-symbolic approaches, come into play.

Chair: Stefan Sleeuw
Zeit: 15:20-15:50, 11. September 2021 (Samstag)
Ort: HS E.002

Daniela Schuster
(University Konstanz, )

Testability and Meaning deco