Type: Distinguished Lecture Who: Professor Tom Mitchell Topic: What Drives Learning -- Current Data or Prior Knowledge? Dates: 3-Feb-94 Time: 4:00 Place: 7500 Wean Duration: 1.5 hours Host: Merrick Furst ABSTRACT Consider the problem of getting computers to improve automatically through experience. Computer learning methods are currently being explored from a variety of disciplines, including statistics, cognitive science, complexity theory, connectionism, control theory, symbolic reasoning, and philosophy. Regardless of perspective, at the heart of the learning problem lies one fundamental question: by what rational process can general rules be acquired from specific training examples? Across the disciplines, approaches to answering this question fall into two broad categories: inductive and analytical. Inductive learning methods depend on large amounts of data and as few prior assumptions as possible, whereas analytical learning methods use significant prior knowledge to reduce the need for training data. This talk will explore both these approaches to the generalization problem, and the question of how and why to unify them. We will illustrate approaches with problems from robot learning, and present results of a new algorithm (Explanation-Based Neural Network learning) that combines inductive and analytical methods for learning robot control and perception for the Xavier mobile robot.