Towards a Reconciliation Between Reasoning and Learning - A Position Paper

Abstract

The paper first examines the contours of artificial intelligence (AI) at its beginnings, more than sixty years ago, and points out the important place that machine learning already had at that time. The ambition of AI of making machines capable of performing any information processing task that the human mind can do, means that AI should cover the two modes of human thinking: the instinctive (reactive) one and the deliberative one. This also corresponds to the difference between mastering a skill without being able to articulate it and holding some pieces of knowledge that one can use to explain and teach. In case a function-based representation applies to a considered AI problem, the respective merits of learning a universal approximation of the function vs. a rule-based representation are discussed, with a view to better draw the contours of AI. Moreover, the paper reviews the relative positions of knowledge and data in reasoning and learning, and advocates the need for bridging the two tasks. The paper is also a plea for a unified view of the various facets of AI as a science.