An Exploration Into Incremental Learning: the INFLUENCE System

Antoine Cornuéjols

Abstract: ABSTRACT We advocate that the one operational difference between direct and incremental learning is that the later is sensitive to the circumstances under which the data are presented (order and rate). It then becomes of central importance to analyze the effects of its “education”on a system performances, and subsequently to study strategies to optimize this education process. These are the questions we have begun to investigate using a text understander system called INFLUENCE as a test bed. Original features of the system (based on a many particles dynamical system analogy) are described. Results obtained so far are then reported, and the conclusion underlines perspectives for this research.